Meta-Ensemble Learning for IMDb Ratings: A Stacked Hybrid Model Integrating Gradient Boosting and Deep Neural Networks

Meta-Ensemble Learning for IMDb Ratings: A Stacked Hybrid Model Integrating Gradient Boosting and Deep Neural Networks | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Meta-Ensemble Learning for IMDb Ratings: A Stacked Hybrid Model Integrating Gradient Boosting and Deep Neural Networks Md. Faishal Ahmed Rudro, Md. Shahriar Rahman Bhuiyan, Abu Dojana, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6153928/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Precise IMDb movie ratings predictions are vital to the stakeholders of the film industry since it will dictate investment, marketing, and content recommendation. Standard machine learning techniques are not capable of handling high-cardinality categorical columns, bad interactions between features, and non-linear relations between reviews and metadata. This paper proposes the Meta-Ensemble Predictor (MEP), a new state-of-the-art hybrid framework that integrates various gradient boosters (CatBoost, LightGBM, XGBoost) among themselves and with a deep neural network and tuned using a meta-learning algorithm with a Random Forest as the final predictor. With a database of 33,600 movies from 1960 to 2024 having metadata details like genre, director, cast, runtime, box office, and voter ratings, MEP model uses TF-IDF text feature vectorization, polynomial interaction of features, and dimensionality reduction methods for improved feature representation. The model presented here has achieved RMSE of 0.4389 and accuracy of 96.96% at a difference of 1-point from actual IMDb ratings when compared to other state-of-the-art models. The study showcases the strength of ensemble learning to learn the sophisticated patterns of the ratings of the movies in becoming a good film industry predictive analytics tool. IMDb Rating Prediction Meta-Ensemble Learning Gradient Boosting Deep Neural Networks Feature Engineering Hybrid Models Machine Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 1. Introduction Artificial intelligence (AI) and machine learning (ML) have been applied to decision-making in movie making, release, and viewing in the film industry. The single best predictor of whether or not a film succeeds is its IMDb rating, and it is a strong predictor of the level of acceptability among audiences and an influencer for critics in overtly informing marketing, investment, and recommendations to audiences on video sites [ 1 ]. Having the ability to precisely forecast IMDb ratings is crucial to stakeholders such as investors, producers, streaming services, and filmmakers to determine the possible market performance of a film. Yet, forecasting movie rating is difficult due to user preference subjectivity, intricate interactions among various movie features, and non-linear interactions among metadata features such as box office, production budget, director, cast, genre, user rating, and production budget. Past performances of actors and actresses are also discussed earlier that it is a very crucial factor in rating the movie, and a more appropriate model selection and modeling method is needed for rating estimation enhancement [ 2 ]. Current machine learning methods towards IMDb rating estimation are faced with several challenges. Classic methods such as decision trees and linear regression were ineffective with high-cardinality categorical features and also with the non-linear interactions between the movie features. Although gradient boosting models (i.e., LightGBM, XGBoost, CatBoost) and deep neural networks (DNNs) have proven to be useful in improving prediction accuracy, they are prone to overfitting, bias, and generalizability to out-of-sample values. Furthermore, several data mining methods have been used to evaluate the success of the movie business, yet the majority of the research studies present today are not able to work with textual information, for instance, actor and genre metadata, or even simply do not apply advanced feature engineering techniques that enable models to optimize [ 3 ]. These limitations necessitate the development of a more efficient machine learning paradigm that can more efficiently distinguish intricate patterns between IMDb ratings. This paper investigates a machine learning approach that is an ensemble approach and aims to improve IMDb rating prediction by ensembling different models, advanced feature extraction methods, and meta-learning methods. This framework, which is built, is learned over a dataset of 33,600 movies of the years 1960–2024 with compact heterogeneous features like director, crew, running time, budget, fan rating, gross revenue, and won awards. The information go through strenuous preprocessing in terms of TF-IDF vectorization of textual data, polynomial feature interaction for non-linear interaction, and dimensionality reduction techniques for achieving improved computational efficiency. The major contributions of this paper include: •Have suggested a feature engineering pipeline as an end-to-end framework, employing TF-IDF text metadata vectorization, polynomial expansion of numerical features, and dimensionality reduction for the sake of attaining higher computational efficiency. • Constructed a hierarchical ensemble learning system that combines a collection of machine learning models, including gradient boosting models (CatBoost, LightGBM, XGBoost) and deep neural networks (DNNs), along with a meta-learning strategy to drive optimal prediction. • Constructed and explored a high-dimensional 33,600-movie dataset with a wide collection of movie-focused features in order to achieve best predictive performance. • Was a better predictor than prior models with an RMSE of 0.4389, an R² of 0.8511, and a 96.96% accuracy with a 1-point range. The structure of the paper is as follows: section 2 presents related work for the task of IMDb rating prediction, including prior methodologies and their inadequacies. Section 3 presents the dataset, preprocessing methods, and feature engineering techniques in this study. Section 4 introduces the ensemble-based prediction system with the explanations of machine learning model combination and meta-learning strategies. Section 5 reports the experimental results, such as model evaluation and comparison with the current state-of-the-art techniques. Finally, Section 6 summarizes the research with major findings and suggestions for further research. 2. LITERATURE REVIEW The prediction of movie ratings has been an extensively researched subject over the last ten years with proposed techniques to maximize accuracy along with interpretability. Popular techniques include traditional approaches that have been widely used in CF and CBF. These suffer from data sparsity, cold-start problems, and weak generalization to new items and users. To overcome these limitations, machine learning (ML) and deep learning (DL) models have been incorporated into movie rating prediction systems, and this has shown remarkable performance gains. One of the researches investigated supervised machine learning algorithms in IMDb metadata-based movie rating prediction and identified Random Forest as the best algorithm with 92% accuracy. Although the research offered satisfactory results, it did not incorporate deep feature extraction techniques and meta-learning strategies, which constrained its generalization performance across different datasets [ 4 ]. Another method used a hybrid recommendation system, combining movie metadata and user behavior using weighted score aggregation and machine learning-based feature extraction. The model was better than simple recommendation methods but needed extensive hyperparameter tuning to achieve optimal prediction accuracy [ 5 ]. Gradient boosting algorithms like XGBoost, LightGBM, and CatBoost have been used as useful substitutes for working with structured data for rating prediction problems. One such application used a boost-based prediction model with 92.7% accuracy, which performed better than traditional machine learning algorithms. Still, problems continued to arise in rapid processing of categorical variables and computational complexity optimization [ 6 ]. Another study used XGBoost on IMDb ratings with 82% accuracy as stated, limited by the size of the dataset and the use of manually chosen features [ 7 ]. Work also on the use of matrix factorization (MF) and recurrent neural networks (RNNs) did better but at higher computational expense for real-time use [ 8 ]. Current research has been focused on deep learning-based techniques, i.e., artificial neural networks (ANNs) and convolutional neural networks (CNNs) for latent feature representation extraction of visual and textual information. One such strategy employed a CNN-LSTM model for rating prediction on IMDb and its competence in extracting sequence patterns in review comments. Despite achieving the best results compared to conventional ML models, this approach was linked to high computational expense and large-scale hyperparameter search [ 9 ]. Also, one of the experiments conducted for testing feature engineering on hybrid ML models showed that inclusion of TF-IDF vectorization and polynomial feature transformations for improving prediction accuracy was constrained by availability of good quality metadata [ 10 ]. Several recent works have written about ensemble learning techniques and feature engineering in domain-specific areas to further improve prediction. One work used an ensemble of decision trees and neural networks for predicting IMDb ratings with higher generalizability, reducing overfitting using regularization methods [ 11 ]. Another work used exploratory data analysis (EDA) of IMDb ratings to study features that affect rating distributions significantly. The results identified that genre, budget, and demographics have significant influence on movie ratings, underscoring the relevance of systematic feature selection for predictive models [ 12 ]. Alongside algorithm development, quality in the dataset also contributes to predictive accuracy. The IMDb dataset (1960–2024) comprises an enormous collection of 33,600 movies, which allows researchers to train the models over a big representative dataset covering metadata, revenue data, and user ratings [ 13 ]. This data has enabled the creation of stronger models that can effectively manage long-term trends in film ratings. It remains, however, challenging to make predictive models generalizable across times and changing audience tastes [ 14 ]. Lastly, fundamental machine learning concepts continue to influence movie rating prediction as a research topic. Experiments revealed how statistical learning methods like Lasso regression, decision trees, and Bayesian models remain essential to estimating feature importance and preventing model bias [ 15 ]. These classic approaches, when combined with deep learning-inspired architecture, can provide a good balance of a hybrid model in movie prediction. Another study utilized Heterogeneous Information Network-based recommendation systems (HI2Rec) on MovieLens datasets and reported notable improvements in rating prediction accuracy. Scalability problems were, however, encountered, especially when the model was used on large datasets like IMDb [ 16 ]. In the meantime, aesthetic-aware matrix factorization models have been introduced to improve content-based filtering by incorporating visual features of movie posters and trailers. These models have been more precise, but their dependency on multimodal data sources is challenging in computational efficiency [ 17 ]. Meta-learning methods have been increasingly popular in movie rating prediction within the domain. A research using Markov Factorization Process (MFMP) for recommendation indicated a notable performance improvement over conventional methods. However, the research mentioned that multimodal data fusion is an open problem [ 18 ]. Another research went as far as exploring long- and short-term information utilization in models from adversarial training, and they were able to make more content-aware recommendations with less overfitting problems [ 19 ]. The research landscape presents that ensemble learning techniques provide a robust methodology by combining strengths of individual models. Recent papers have attained state-of-the-art accuracy in the prediction of ratings for movies through the integration of boosting-based models and deep neural networks. Computational cost, model interpretability, and scalability are still issues of concern. The next section provides an overview of how the Meta-Ensemble Predictor follows up on such advancements to attain improved predictive results while showing a deficiency in current models. 3. Materials and Methods The Meta-Ensemble Predictor (MEP) framework is designed to leverage ensemble learning through fusing an ensemble of machine learning models and deep learning techniques towards improved IMDb Movie rating prediction. The approach that is being followed in the methodology begins with dataset gathering, preprocessing, feature engineering, model training, and ensemble stacking and extends to the best predictions. In Fig. 1 ,It illustrates the workflow of the suggested framework, which highlights the sequential process of transforming raw movie metadata into predictive information. 3.1 Dataset Description The dataset utilized here is used from the Kaggle IMDb Movies from 1960 to 2023 publicly available dataset [ 13 ]. The dataset comprises 33,600 movies distributed over a span of more than six decades that have an inclusive list of metadata details on every single film. The dataset incorporates crucial variables on which IMDb scores depend, i.e., production-oriented variables, audience measures of engagement, and business-oriented performance measures. Each entry in the movies dataset has an official title, release date, and runtime, which are sparse context features for rating prediction. The addition of the Motion Picture Association (MPA) rating provides more information about a movie's target audience and possible restriction on viewing. User-generated attributes like the rating and vote on IMDb are associated with the prediction task because they capture the reception and engagement of the user. Economic attributes like budget, worldwide gross box office, U.S. and Canadian gross, and opening weekend gross provide useful information about the profitability of a film, a trait that is typically linked with critical success. Movie crews such as directors, writers, and main actors are significant features of the dataset since they influence the expectations of audiences and the quality of the film. These attributes are preprocessed through encoding and vectorization to make them more convenient for model integration. Figure 2 illustrates average movie rating over time while Fig. 3 illustrates the distribution of Movie ratings where it is a normal distribution with its mode at mid-range scores. Categorical attributes such as genres, production country, shooting location, production companies, and spoken languages are crucial attributes to achieve the regional and cultural impact on the film. Award and award won, Oscars, wins, nominations, provide more richness to the dataset because award-winning films possess a higher IMDb rating. Release date is converted into numerical format so that time patterns that may affect rating prediction might be trapped. Figure 4 is language and genre-based rating. The dataset heterogeneity allows a worldwide approach to predictive modeling, encompassing both objective monetary data and unmeasurable artistic attributes. 3.2 Data Preprocessing The raw data were preprocessed in detail to be consistent, minimize feature representation loss, and optimize model performance. Due to the heterogeneity of movie features—ranging from numerical financial quantities to text-based categorical metadata—a formal multi-step preprocessing pipeline was employed. Steps for missing value imputation, categorical encoding, feature transformation, and normalization were incorporated to pre-process the data for predictive modeling. Figure 3 illustrates the end-to-end preprocessing pipeline that converts raw movie metadata into machine-learning-friendly inputs. One of the first preprocessing problems was missing values on all attributes. Missing values in the IMDb rating column, the target column, were imputed with the global mean rating so that there was numerical consistency without adding too much bias. Money attributes like budget, worldwide gross, U.S./Canada gross, and opening weekend had many missing values because figures might be unreleased, and so it was imputed so that there was numerical consistency. Rather than dropping these films, missing values were imputed by zero so that the integrity of the dataset wasn't lost and bias from hypothetical imputations was prevented. Categorical attributes like directors, writers, and actors also had missing values as metadata was missing. Missing values in those were imputed with the placeholder "Unknown" so that the model was provided structured input instead of excluding informative points. The data was also needed to be thoroughly cleaned and normalized, especially for categorical data. Film names contained inconsistencies such as leading digits, extra spaces, and special characters, which were stripped out to make it consistent. The Votes column, which was initially stored in abbreviated format (e.g., "687K" for 687,000 votes), was converted to its full numerical value in order to facilitate appropriate numerical analysis. The genre column, being a set of genres for one movie, was transformed to multi-hot encoded vector representation in such a manner that every genre would contribute equally towards the learning process of the model. The model was thus capable of learning how the various sets of genres affected IMDb ratings by being transformed in this fashion. Since machine learning algorithms are programmed to take numerical inputs, the categorical variables were converted into numbers through some encoding schemes. For example, high-cardinality categorical features like MPA ratings, directors, writers, production companies, and locations employed label encoding where each category was represented as a numerical value. More sophisticated transformations were required for text features like actors, genres, and languages, though. Term Frequency-Inverse Document Frequency (TF-IDF) vectorization was applied to these features to convert text data into numerical data without losing the relative importance of words and terms. Since TF-IDF matrices are of large dimension, Truncated Singular Value Decomposition (SVD) was applied for reducing the dimension without losing valuable feature information while maintaining computational efficiency. Feature transformation techniques were required to pre-process the data before it could be fed into machine learning algorithms. Polynomial feature interactions were formed in an attempt to reveal sophisticated interactions of powerful numerical features. Budget by gross revenue interactions, for instance, helped to reveal high-budget films that became box office hits, and vote by IMDb rating interactions helped to reveal user interaction. In addition, the vote column was also very imbalanced, with blockbusters registering unusually high vote counts, as is evident from Fig. 5 . The last but most important preprocess step was feature standardization. In an improved preprocessing of the data, outlier removal techniques like interquartile range (IQR) filtering were used to eliminate outlier values from budget and revenue columns to avoid models learning noise but typical data. With the use of these preprocess steps, the data not only stabilized but enabled the model to learn more meaningful patterns. Figure 6 displays word clouds for highly-rated movies, highlighting the most frequently occurring words in successful films. 3.3 Proposed Model Architecture The Meta-Ensemble Predictor (MEP) is introduced for ensemble-improving IMDb rating prediction. MEP applies ensemble methods of a lot of machine and deep learning models in trying to develop a universally inclusive generalized and stable predict feature. MEP utilizes four chief models—CatBoost, XGBoost, LightGBM, and Neural Network—and winds down with the finishing touch that forms a meta-learning layer used in generalizing and summarizing prediction through Random Forest Regressor. Each model within the ensemble has something unique to contribute, so MEP can catch subtle feature interactions that a single model may miss. CatBoost is a gradient boosting algorithm specifically designed for categorical feature handling. It is used here to imbed high-cardinality categorical columns such as film directors, stars, and categories that are highly significant in their prediction of viewer acceptability. XGBoost is used as another gradient boosting algorithm for the capture of non-linear interaction effects between numerical columns. It proves useful while performing economic calculations like budget, box office, and voting public and thus becomes the central part of MEP. LightGBM is specially designed for use with massive data and has computing speed with no compromise in accuracy. It proves very handy while performing TF-IDF vectorized text features and high-dimensional categorical features so that more expressiveness in the feature representation is achieved. Additionally, the Neural Network module is capable of introducing deep learning to the system such that it can learn abstract high-level feature representations which may be neglected by gradient boosting models. It uncovers underlying relationships between different features of the movie effectively and allows in-depth analysis of rating behavior. When all base models have predicted, they are passed on to the meta-model, a Random Forest Regressor, which determines the best ensemble of base model predictions to make use of to make the end prediction. The meta-model takes the average across predictions, resists overfitting, and generalizes better by making the end prediction an optimal ensemble of the strengths of numerous models. Hierarchical learning maintains the model variance and bias at their minimum value, and MEP is highly insensitive to variation in movie data. Figure 7. Proposed Model Diagram The Neural Network architecture in MEP architecture of the Neural Network is a multilayer feedforward fully connected neural network with many layers providing feature extraction as well as prediction enhancement. Input Layer The numerical attributes and one-hot encoded attributes of categorical attributes are input to the neural network, where both are input as such. They are normalized by Z-score normalization so that they remain uniform in the instance of differing attribute scales. Hidden Layers The network has three hidden layers, and all of them are thoroughly optimized to increase learning capacity and generalization. The first hidden layer with 1024 neurons uses the ReLU (Rectified Linear Unit) activation function that brings non-linearity into the model. The activation function prevents the vanishing gradient problem and enables the model to learn complex relationships in the data efficiently. The second hidden layer consisting of 512 neurons uses Batch Normalization, speeding up training by normalizing weight distributions. The method stabilizes learning to enable the network to converge quickly and optimize optimally on diverse data distributions. The third hidden layer consisting of 256 neurons uses Dropout Regularization (30%), a method where randomly during training some of the neurons are cut. This method avoids overfitting as it forces the model to learn more regularized feature representations instead of depending on individual neurons. The employment of the three hidden layers enables the deep learning model to learn abstract high-level feature representations, enhancing prediction accuracy and enabling the model to generalize from new data. Output Layer It has a neuron with a linear activation function that provides the predicted final IMDb rating as a real actual value. Training optimization of the neural network is done with the Adam algorithm and the adaptive learning rate having dynamic ranges of values for free convergence. Furthermore, early stopping is applied, wherein training is terminated when no improvement is seen in validation loss. L2 Regularization is employed to combat overfitting even further, in deep layers, managing model complexity and generalization. The meta-learning phase of MEP improves the prediction performance with the help of knowledge from the individual base models. The process starts on the premise that the base models are individually trained on the same pre-processed data and produce unique data feature predictions. The predictions are stacked in a second dataset, which is employed to train the Random Forest meta-model. The meta-model is monitoring the interaction among base model outputs with one another and determining the best distribution of weights for fine-tuning the final estimate. The second level of training removes inconsistencies and improves the predictive accuracy by making the most accurate predictions of base models to announce the final decision. Unlike the traditional single-model approaches, MEP reaches the best prediction performance by stacking heterogeneous learning paradigms together. The gradient boosting models perform best on structured tabular data, the Neural Network learns high-level abstracted feature representations, and the meta-model stacks the best of every model in the best possible way. With the state-of-the-art stacking-based ensemble learning paradigm, MEP tends to make state-of-the-art IMDb rating predictions with improved robustness, decreased bias, and improved generalization on different movie datasets. 4. Results and Evaluation 4.1. Performance Analysis The Meta-Ensemble Predictor (MEP) was comprehensively validated with the leading machine learning algorithms of the market, such as LightGBM, XGBoost, Gradient Boosting Regressor, and Neural Networks, to check MEP's prediction precision and overallizability. The performances of the models were assessed with four essential metrics, i.e., Root Mean Squared Error (RMSE), R² Score, Mean Absolute Error (MAE), and Accuracy Within ± 1-Point Deviation. RMSE was employed to estimate the standard deviation of the prediction error, with lower values meaning greater precision in prediction. R² value was employed to calculate the variance proportion in IMDb ratings explained by the model, with the higher values reflecting good performance. MAE offered a clear measure of mean absolute error, which simplified the comparison of the accuracy of the models. In addition, Accuracy Within ± 1-Point Deviation discussed how often the prediction was in 1-point range from the actual IMDb scores, thus ensuring the model output remained in effectively useful terms. These approximations as a whole provided a correct foundation to judge the efficiency of MEP in contrast to other methods. Table 1 shows the comparative accuracy of MEP compared to baseline models. From the results, it can be seen that MEP has the least RMSE value of 0.4389 and largest R² of 0.8511 and thus is the top-performing model in terms of predicting IMDb rating. Its 96.96% accuracy under a 1-point error implies MEP's very good consistency of prediction, significantly better than others. LightGBM, being highly computationally efficient, performed an RMSE of 0.7342 and an R² score of 0.5832, which is significantly lower than MEP.XGBoost, previously bragged about being able to model structured data, only managed a score of 0.3748 for R², which indicates low predictability. Gradient Boosting Regressor, while performing better than XGBoost, was unable to outperform MEP's accuracy and error minimization. The Neural Network model, although apt at feature extraction, was prone to overfitting and returned an R² of only 0.4361, which is lower than MEP's ensemble model. Table 1 MODEL PERFORMANCE COMPARISON Model RMSE R² Score Accuracy (Within ± 1-Point) MEP (Proposed Model) 0.4389 0.8511 96.96% LightGBM 0.7342 0.5832 85.55% XGBoost 0.8992 0.3748 76.88% Gradient Boosting Regressor 0.7417 0.5747 85.68% Neural Network 0.8540 0.4361 90.56% The Comparative outcomes mirror the strength of MEP's ability to reduce prediction errors at retaining superior generalization on heterogeneous IMDb ratings. 40.2% of the RMSE reduction over LightGBM confirms MEP's ensemble learning approach in grasping intricate patterns in the data in an overall masterful manner. In contrast to XGBoost and LightGBM, which are more variant and less interpretable on categorical-dominant datasets, MEP's hybrid approach balances tree-based learning with deep neural feature extraction to produce a more robust prediction model. The considerably enhanced R² score of 0.8511 proves that MEP captures more variation in IMDb ratings compared to other approaches. The precision in ± 1-point difference to 96.96% also indicates MEP's practical usability in real life, in the sense that nearly all the predictions are very close to actual IMDb ratings. Figure 9 illustrates the training loss vs valid loss graph and Fig. 10 illustrates Training R² score vs valid R² score. A deeper Residual analysis was carried out more rigorously to investigate the distribution of prediction errors for various levels of IMDb ratings. A well-calibrated predictive model would have residuals symmetrically spread around zero, reflecting unbiased predictions. Figure 11 shows the residual distribution of MEP, and it can be seen that most residuals are tightly clustered around zero, which reflects that MEP is not adding systematic overestimation or underestimation biases. This contrasts with models like XGBoost and LightGBM, which exhibit higher residual variance, especially for well-rated movies. The residual spread remains low for mainstream ratings (4–8), demonstrating the robustness of MEP in doing well with mainstream movie predictions. Residuals do open up a bit for extreme IMDb ratings (1–2 and 8–9), as would be expected from the subjective nature of audience preferences for extremely polarizing films. Figure 12, is the predicted rating vs true ratings plot, The points are tightly clustered about the red dashed line, meaning that the predictions are very close to the actual ratings. The spread of the prediction is not highly skewed or biased, i.e., the model generalizes very well for different rating values. 4.2 Discussion The Meta-Ensemble Predictor (MEP) shows a remarkable IMDb rating prediction using ensemble learning and feature engineering methods. In contrast to the conventional models like Random Forest, XGBoost, and LightGBM, which are not efficient when dealing with high-cardinality categorical features, feature interactions, and non-linear relationships, MEP combines several base models and a meta-model to offer more accurate predictions. Adding CatBoost, XGBoost, LightGBM, and a Neural Network enables MEP to identify complex patterns between the movie metadata, which has enhanced predictive accuracy. One of the strongest arguments MEP wins is that it boasts a complete engineering pipeline in the form of TF-IDF text feature vectorization, polynomial feature expansion, and Bayesian hyperparameter optimization. The prior work utilized rudimentary feature encoding methods and thus recorded accuracy levels of between 70% and 92.7%. Furthermore, much of the previous work was marred by small datasets and was thus non-generalizable. MEP, with 33,600 movies on training over the span of six decades, easily obviates this drawback with thorough generalization of a plenty of film characteristics. Table 2 presents a comparison of MEP's performance with existing research, demonstrating its higher accuracy and generalization capability. Table 2 PERFORMANCE COMPARISON OF MODELS Reference Model Accuracy (%) This Work Meta-Ensemble Predictor 96.96% [ 4 ] Random Forest 92.00% [ 7 ] XGBoost 82.00% [ 6 ] Random Forest 92.70% [ 11 ] Naïve Bayes 70.00% The results clearly demonstrate that MEP excels other standard machine learning models by achieving the highest accuracy (96.96%) while regulating overfitting and model robustness. Ensemble learning enables MEP to leverage strengths of different models, leading to stability across diverse rating distributions. In comparison to XGBoost-based studies (accuracy ~ 82%), susceptible to categorical encoding, MEP is found to be robust with actor, director, and genre-dependent attributes with enhanced categorical support using CatBoost. Another most significant reason for the supremacy of MEP is that MEP can process giant-sized datasets. Most of the previous models are trained on comparatively much smaller datasets in terms of movies, thereby poor general representation of rating patterns. But MEP is trained on 33,600 movies, thereby a better general understanding of user rating patterns over a period of several decades. Second, the old models lack the ability to convincingly mimic top-level interactions between attributes, whereas MEP uses a Random Forest meta-model to average and smooth the predictions to remove variance and include generalization. In addition, earlier research experiments have demonstrated high residual variance, especially when handling outlier IMDb ratings. MEP's residual analysis prefers uniform error distribution in Fig. 13 , as evidenced by the fact that the model accurately captures the pattern of the ratings without being overly affected by outliers. The ensemble architecture also prefers robustness, protecting against any of the base models individualizing predictions and thus creating lower margins for error and stabilizing ratings. Through the use of sophisticated ensemble learning, feature engineering, and hyperparameter tuning, MEP sets a new standard for IMDb rating prediction. By integrating conventional machine learning with deep learning techniques, MEP is more precise, generalizes better, and has more stable predictions, making it a very efficient predictive analytics framework in the entertainment industry. 5. Limitations and Future Work While the Meta-Ensemble Predictor (MEP) is heavily supplemented to forecast IMDb ratings, it is not without its flaws. One of its major flaws is computational complexity. The ensemble learning method, particularly with many base models and a meta-model, is very computationally intensive and training is very time-consuming as opposed to single-model strategies. It is thus less scalable to real-time where inference speed is crucial. Again, the limitation is to be vulnerable to overfitting in the treatment of deep feature transformation and stacking-based learning. Bayesian hyperparameter optimization and dropout regularization mitigate the limitation but must be further optimized such that it generalizes much more to fresh data. Apart from this are certain extrinsic factors like movie critics' scepticism, audience sentiment, and social opinion not modelled in an explicit format and that drive the performance of prediction. Future work will involve maximizing computational cost with synergistically guiding effective model distillation and pruning methods towards optimizing maximum MEP for real-time application. Real-time powered sentiment analysis of social media and review websites can similarly be applied to improve rating prediction accuracy by tracking dynamic audience mood. Application of transformer models such as BERT or GPT models to text feature extraction can similarly include contextual movie metadata awareness. Finally, semi-supervised learning techniques from unlabelled data can be applied theoretically on MEP to integrate legacy recommendation systems. Conclusion This study introduced the Meta-Ensemble Predictor (MEP), a new hybrid model that integrates gradient boosting models (CatBoost, XGBoost, LightGBM) with a deep learning-based Neural Network, meta-optimized using a Random Forest Regressor. The integration of heterogeneous learning strategies enabled MEP to learn intricate feature interactions, reduce overfitting, and improve generalization for IMDb rating prediction. By using sophisticated feature engineering methods, such as TF-IDF vectorization of text features, polynomial expansion of features, and Bayesian optimization of hyperparameters, MEP scored higher in its prediction accuracy with 96.96%—far superior to that of traditional machine learning methods. Comparative study also established MEP's superiority over current methodologies by being able to model high-dimensional data precisely and predict more uniformly with a high diversity of movie features. The results emphasize the significance of ensemble learning for predictive modeling, especially where user ratings based on subjective opinions are driven by multiple influencing factors. With its success in combining ordered numerical data with text and category features, MEP provides a robust and scalable solution for further breakthroughs in IMDb rating prediction and other wider uses in entertainment analytics. Declarations Author Contribution M.F.A.R. (Md. Faishal Ahmed Rudro) conceptualized and designed the study, developed the Meta-Ensemble Predictor (MEP) framework, and performed the primary machine learning model implementation. He also wrote the main manuscript text and conducted data preprocessing, feature engineering, and model evaluation.M.S.R.B. (Md. Shahriar Rahman Bhuiyan) contributed to dataset acquisition, preprocessing, and exploratory data analysis. He also assisted in hyperparameter tuning and the implementation of gradient boosting models (CatBoost, LightGBM, XGBoost).A.D. (Abu Dojana) worked on deep neural network model integration and contributed to performance comparison with state-of-the-art methods. He was also involved in fine-tuning the Random Forest meta-learner.M.R.H.S. (Md. Reduanul Haque Shakib) assisted in statistical analysis, residual evaluation, and model validation. He prepared figures and tables summarizing model performance metrics and contributed to result interpretation.A.A.O. (Arisha Ashraf Oishi) contributed to literature review compilation, manuscript formatting, and references organization. She also assisted in ensuring compliance with journal submission requirements.All authors reviewed the manuscript and approved the final version for submission. Data Availability The dataset used in this study, titled "IMDb Movies from 1960 to 2023", is publicly available on Kaggle at the following link:🔗 https://www.kaggle.com/datasets/raedaddala/imdb-movies-from-1960-to-2023This dataset contains metadata on 33,600 movies spanning from 1960 to 2024, including attributes such as genre, director, cast, runtime, box office revenue, and IMDb user ratings. The data was preprocessed and analyzed for model training and validation in this study.Researchers interested in replicating the results can access the dataset through the provided Kaggle link. Any additional preprocessing scripts and model configurations used in this study can be shared upon reasonable request to the corresponding author. References Internet Movie Data Base URL: https://www.imdb.com/ Persson K (2015) Predicting movie ratings: A comparative study on random forests and support vector machines. Bachelor Degree Project in Informatics, University of Skövde Tashman M (2015) The Association Between Film Industry Success and Prior Career History: A Machine Learning Approach, Master's thesis, Harvard Extension School Siddique A, Ahmed M, Rahman S (2023) Movies Rating Prediction Using Supervised Machine Learning Techniques. Int J Inform Syst Comput Technol (IJISCT) 6(2):45–51 Kumar S, De K, Roy PP (2020) IEEE Trans Comput Soc Syst 7(4):915–923. 10.1109/TCSS.2020.2993585 . Movie Recommendation System Using Sentiment Analysis from Microblogging Data Abidin D, Bostanci C, Site A (2018) Movie Rating Prediction with Machine Learning Algorithms on IMDB Data Set, in Proceedings of the International Conference on Advanced Technologies, Computer Engineering and Science (ICATCES'18) Dixit P, Verma R, Tiwari A (2021) Predicting the IMDB Rating by Using EDA and Machine Learning Algorithms, International Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT) , vol. 7, no. 1, pp. 55–63 Zhang J, Wang Y, Yuan Z, Jin Q (2020) Personalized Real-Time Movie Recommendation System: Practical Prototype and Evaluation, pp. 180–191 Brownlee J (2019) A Gentle Introduction to Gradient Boosting Algorithms for Machine Learning, Machine Learning Mastery , [Online]. Available: https://machinelearningmastery.com/gradient-boosting-algorithms/ Chen K, Wu Z (2022) Feature Engineering for Movie Rating Prediction Using Hybrid Machine Learning Models. IEEE Trans Knowl Data Eng 34(7):2501–2515 Sivakumar P, Kumar R, Sharma SK (2020) Movie Success and Rating Prediction Using Data Mining Algorithms, International Conference on Computational Intelligence and Data Science (ICCIDS) Johnson T, Patel M (2022) Machine Learning-Based Analysis of Movie Ratings Using IMDb Data, International Conference on Artificial Intelligence and Data Analytics (ICAIDA) IMDb D (1960–2024) Available: https://www.kaggle.com/datasets/raedaddala/imdb-movies-from-1960-to- 2023. Hastie T, Tibshirani R, Friedman J (2009) The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edn. Springer Bishop C (2006) Pattern Recognition and Machine Learning, 1st edn. Springer He M, Wang BO, Du X (2019) HI2Rec: Exploring Knowledge in Heterogeneous Information for Movie Recommendation. IEEE Access 7:30276–30284. 10.1109/ACCESS.2019.2902398 Chen X et al (2019) Exploiting Aesthetic Features in Visual Contents for Movie Recommendation, IEEE Access, vol. PP, no. c, p. 1. 10.1109/ACCESS.2019.2910722 Chen X, Liu D, Member S, Xiong Z (2020) Learning and Fusing Multiple User Interest Representations for Micro-Video and Movie Recommendations, vol. 9210, no. c, pp. 1–13. 10.1109/TMM.2020.2978618 Zhao W et al (2019) Leveraging Long and Short-Term Information in Content-Aware Movie Recommendation via Adversarial Training. IEEE Trans Cybern PP:1–14. 10.1109/TCYB.2019.2896766 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6153928","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":425001034,"identity":"e494be9e-412d-4397-a375-e2ddc297e138","order_by":0,"name":"Md. Faishal Ahmed Rudro","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCElEQVRIiWNgGAWjYFCCBIYPDAZAkh3I5mFgkAOJHXiAXwvjDLAWZogWY7CWBIJaGBBaEhsgVuMGuu0JjA0/Cury+Jt5n0m83WOTPj/s8EOgLXZyug3YtZidecDY2GNwuFjiMLuZ5Jxnabkbb6cZALUkG5sdwKHlRgL7Ax6DA4kNh9nYpHkOHM7dODsBpOVA4jbcWhgb/xjUJc6Hakk3nJ3+gaCWZh4D5sQNUC0J8tI5BGw587CxWcbgcOLGw2zMlnMOpBlukM4pOJBggMcvx5MPNr75U5c473gb4403B2zk5Wenb/7wocJODpcWBgbGBlS+AVilAS7l2IB8A0Elo2AUjIJRMMIAAIdUZaAP6LLtAAAAAElFTkSuQmCC","orcid":"","institution":"East West University","correspondingAuthor":true,"prefix":"","firstName":"Md.","middleName":"Faishal Ahmed","lastName":"Rudro","suffix":""},{"id":425001035,"identity":"9b485027-2747-4df5-83ff-0bcb658c7e4f","order_by":1,"name":"Md. Shahriar Rahman Bhuiyan","email":"","orcid":"","institution":"Ahsanullah University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Md.","middleName":"Shahriar Rahman","lastName":"Bhuiyan","suffix":""},{"id":425001036,"identity":"7a1477ea-12fe-4329-977b-09187a1f8a15","order_by":2,"name":"Abu Dojana","email":"","orcid":"","institution":"Ahsanullah University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Abu","middleName":"","lastName":"Dojana","suffix":""},{"id":425001037,"identity":"ba4ba4dc-df91-4c75-ac2e-f665fcf54d33","order_by":3,"name":"Md. Reduanul Haque Shakib","email":"","orcid":"","institution":"American International University- Bangladesh Dhaka","correspondingAuthor":false,"prefix":"","firstName":"Md.","middleName":"Reduanul Haque","lastName":"Shakib","suffix":""},{"id":425001039,"identity":"6534af49-4d74-46c1-998b-7177083bb632","order_by":4,"name":"Arisha Ashraf Oishi","email":"","orcid":"","institution":"East West University","correspondingAuthor":false,"prefix":"","firstName":"Arisha","middleName":"Ashraf","lastName":"Oishi","suffix":""}],"badges":[],"createdAt":"2025-03-04 11:23:10","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6153928/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6153928/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":77974119,"identity":"bec29888-1e1e-4ae2-b8cf-82e94db309d3","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":60837,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow Diagram\u003c/p\u003e","description":"","filename":"Picture1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/637407d426b71408e550d88f.jpg"},{"id":77974555,"identity":"72a9c061-d006-425f-b1b3-aca8efb67dfb","added_by":"auto","created_at":"2025-03-07 11:32:40","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":10138,"visible":true,"origin":"","legend":"\u003cp\u003eMovie Ratings Over Time\u003c/p\u003e","description":"","filename":"Picture2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/2d376250dccd155179f245bf.jpg"},{"id":77974553,"identity":"da822fc2-d400-4786-afd8-b18d400e9c0d","added_by":"auto","created_at":"2025-03-07 11:32:40","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10388,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of Movie Ratings\u003c/p\u003e","description":"","filename":"Picture3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/9c73a9338e9d091ef6924d65.jpg"},{"id":77974124,"identity":"96e12905-2e77-4ab2-be07-cb765f7ea869","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":32673,"visible":true,"origin":"","legend":"\u003cp\u003eGenre and Language wise Rating\u003c/p\u003e","description":"","filename":"Picture4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/88c9ebfd42c0ec56009fd1de.jpg"},{"id":77974155,"identity":"f75503f8-168c-425c-b3ef-9ea833401dae","added_by":"auto","created_at":"2025-03-07 11:24:41","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":16391,"visible":true,"origin":"","legend":"\u003cp\u003eVotes Distribution (Original VS Scaled)\u003c/p\u003e","description":"","filename":"Picture5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/ee71aa41706ef1581d272e8d.jpg"},{"id":77975604,"identity":"a7058add-2d13-43ed-a033-9b21e810de1c","added_by":"auto","created_at":"2025-03-07 11:40:40","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":85795,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWord Clouds for Highly-Rated Movies\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/ba3614fb2054b979c2da0092.jpg"},{"id":77974565,"identity":"9dd2e0fc-b0e4-4195-b138-1f5ba1e499a5","added_by":"auto","created_at":"2025-03-07 11:32:40","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":72037,"visible":true,"origin":"","legend":"\u003cp\u003eProposed Model Diagram\u003c/p\u003e","description":"","filename":"Picture7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/76c35ca1f51f0b77bfc38848.jpg"},{"id":77974122,"identity":"a43f61e8-4ff9-4301-9394-0600eb9875f8","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":16990,"visible":true,"origin":"","legend":"\u003cp\u003eBase Model Contribution to Meta-Model\u003c/p\u003e","description":"","filename":"Picture8.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/b3d676d96dc61ecdd1900d33.jpg"},{"id":77975955,"identity":"798c6733-9935-4538-9f72-f88cadd3b9ef","added_by":"auto","created_at":"2025-03-07 11:48:40","extension":"jpg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":8535,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining Loss VS Valid Loss\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture9.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/3b6c1f2d007b696879eb1479.jpg"},{"id":77974560,"identity":"73a41fa7-a91a-4f85-868c-2e82c8c98c6b","added_by":"auto","created_at":"2025-03-07 11:32:40","extension":"jpg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":11320,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTraining R² VS Valid R²\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture10.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/2c1583521bd352ac85f44fe2.jpg"},{"id":77974131,"identity":"94bc4829-f183-4578-8447-0bedf9581b48","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":7978,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResidual Distribution\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture11.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/d7a3b188a2f57927fc4a1a36.jpg"},{"id":77974132,"identity":"1b561120-d57a-4226-bfb9-48db2df60b3c","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":10712,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrediction VS True Ratings\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture12.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/8ef7a0388ed5bf6183f1bb84.jpg"},{"id":77974136,"identity":"89a69e73-191e-4cfb-a711-d2c680467c06","added_by":"auto","created_at":"2025-03-07 11:24:40","extension":"jpg","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":13352,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eError Distribution\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"Picture13.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/d50c13044fd4962c13a9bfc6.jpg"},{"id":78361810,"identity":"9b2464f0-af00-45de-b996-73b436286e42","added_by":"auto","created_at":"2025-03-12 12:24:00","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1051846,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6153928/v1/db1f7868-dc8e-4f7b-a5ad-4fed0af84a21.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Meta-Ensemble Learning for IMDb Ratings: A Stacked Hybrid Model Integrating Gradient Boosting and Deep Neural Networks","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eArtificial intelligence (AI) and machine learning (ML) have been applied to decision-making in movie making, release, and viewing in the film industry. The single best predictor of whether or not a film succeeds is its IMDb rating, and it is a strong predictor of the level of acceptability among audiences and an influencer for critics in overtly informing marketing, investment, and recommendations to audiences on video sites [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Having the ability to precisely forecast IMDb ratings is crucial to stakeholders such as investors, producers, streaming services, and filmmakers to determine the possible market performance of a film. Yet, forecasting movie rating is difficult due to user preference subjectivity, intricate interactions among various movie features, and non-linear interactions among metadata features such as box office, production budget, director, cast, genre, user rating, and production budget. Past performances of actors and actresses are also discussed earlier that it is a very crucial factor in rating the movie, and a more appropriate model selection and modeling method is needed for rating estimation enhancement [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCurrent machine learning methods towards IMDb rating estimation are faced with several challenges. Classic methods such as decision trees and linear regression were ineffective with high-cardinality categorical features and also with the non-linear interactions between the movie features. Although gradient boosting models (i.e., LightGBM, XGBoost, CatBoost) and deep neural networks (DNNs) have proven to be useful in improving prediction accuracy, they are prone to overfitting, bias, and generalizability to out-of-sample values. Furthermore, several data mining methods have been used to evaluate the success of the movie business, yet the majority of the research studies present today are not able to work with textual information, for instance, actor and genre metadata, or even simply do not apply advanced feature engineering techniques that enable models to optimize [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. These limitations necessitate the development of a more efficient machine learning paradigm that can more efficiently distinguish intricate patterns between IMDb ratings.\u003c/p\u003e \u003cp\u003eThis paper investigates a machine learning approach that is an ensemble approach and aims to improve IMDb rating prediction by ensembling different models, advanced feature extraction methods, and meta-learning methods. This framework, which is built, is learned over a dataset of 33,600 movies of the years 1960\u0026ndash;2024 with compact heterogeneous features like director, crew, running time, budget, fan rating, gross revenue, and won awards. The information go through strenuous preprocessing in terms of TF-IDF vectorization of textual data, polynomial feature interaction for non-linear interaction, and dimensionality reduction techniques for achieving improved computational efficiency.\u003c/p\u003e \u003cp\u003eThe major contributions of this paper include:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e\u0026bull;Have suggested a feature engineering pipeline as an end-to-end framework, employing TF-IDF text metadata vectorization, polynomial expansion of numerical features, and dimensionality reduction for the sake of attaining higher computational efficiency.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Constructed a hierarchical ensemble learning system that combines a collection of machine learning models, including gradient boosting models (CatBoost, LightGBM, XGBoost) and deep neural networks (DNNs), along with a meta-learning strategy to drive optimal prediction.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Constructed and explored a high-dimensional 33,600-movie dataset with a wide collection of movie-focused features in order to achieve best predictive performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e\u0026bull; Was a better predictor than prior models with an RMSE of 0.4389, an R\u0026sup2; of 0.8511, and a 96.96% accuracy with a 1-point range.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe structure of the paper is as follows: section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents related work for the task of IMDb rating prediction, including prior methodologies and their inadequacies. Section \u003cspan refid=\"Sec3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the dataset, preprocessing methods, and feature engineering techniques in this study. Section \u003cspan refid=\"Sec7\" class=\"InternalRef\"\u003e4\u003c/span\u003e introduces the ensemble-based prediction system with the explanations of machine learning model combination and meta-learning strategies. Section \u003cspan refid=\"Sec10\" class=\"InternalRef\"\u003e5\u003c/span\u003e reports the experimental results, such as model evaluation and comparison with the current state-of-the-art techniques. Finally, Section 6 summarizes the research with major findings and suggestions for further research.\u003c/p\u003e"},{"header":"2. LITERATURE REVIEW","content":"\u003cp\u003eThe prediction of movie ratings has been an extensively researched subject over the last ten years with proposed techniques to maximize accuracy along with interpretability. Popular techniques include traditional approaches that have been widely used in CF and CBF. These suffer from data sparsity, cold-start problems, and weak generalization to new items and users. To overcome these limitations, machine learning (ML) and deep learning (DL) models have been incorporated into movie rating prediction systems, and this has shown remarkable performance gains.\u003c/p\u003e \u003cp\u003eOne of the researches investigated supervised machine learning algorithms in IMDb metadata-based movie rating prediction and identified Random Forest as the best algorithm with 92% accuracy. Although the research offered satisfactory results, it did not incorporate deep feature extraction techniques and meta-learning strategies, which constrained its generalization performance across different datasets [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Another method used a hybrid recommendation system, combining movie metadata and user behavior using weighted score aggregation and machine learning-based feature extraction. The model was better than simple recommendation methods but needed extensive hyperparameter tuning to achieve optimal prediction accuracy [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eGradient boosting algorithms like XGBoost, LightGBM, and CatBoost have been used as useful substitutes for working with structured data for rating prediction problems. One such application used a boost-based prediction model with 92.7% accuracy, which performed better than traditional machine learning algorithms. Still, problems continued to arise in rapid processing of categorical variables and computational complexity optimization [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. Another study used XGBoost on IMDb ratings with 82% accuracy as stated, limited by the size of the dataset and the use of manually chosen features [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Work also on the use of matrix factorization (MF) and recurrent neural networks (RNNs) did better but at higher computational expense for real-time use [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCurrent research has been focused on deep learning-based techniques, i.e., artificial neural networks (ANNs) and convolutional neural networks (CNNs) for latent feature representation extraction of visual and textual information. One such strategy employed a CNN-LSTM model for rating prediction on IMDb and its competence in extracting sequence patterns in review comments. Despite achieving the best results compared to conventional ML models, this approach was linked to high computational expense and large-scale hyperparameter search [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Also, one of the experiments conducted for testing feature engineering on hybrid ML models showed that inclusion of TF-IDF vectorization and polynomial feature transformations for improving prediction accuracy was constrained by availability of good quality metadata [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eSeveral recent works have written about ensemble learning techniques and feature engineering in domain-specific areas to further improve prediction. One work used an ensemble of decision trees and neural networks for predicting IMDb ratings with higher generalizability, reducing overfitting using regularization methods [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Another work used exploratory data analysis (EDA) of IMDb ratings to study features that affect rating distributions significantly. The results identified that genre, budget, and demographics have significant influence on movie ratings, underscoring the relevance of systematic feature selection for predictive models [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAlongside algorithm development, quality in the dataset also contributes to predictive accuracy. The IMDb dataset (1960\u0026ndash;2024) comprises an enormous collection of 33,600 movies, which allows researchers to train the models over a big representative dataset covering metadata, revenue data, and user ratings [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. This data has enabled the creation of stronger models that can effectively manage long-term trends in film ratings. It remains, however, challenging to make predictive models generalizable across times and changing audience tastes [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eLastly, fundamental machine learning concepts continue to influence movie rating prediction as a research topic. Experiments revealed how statistical learning methods like Lasso regression, decision trees, and Bayesian models remain essential to estimating feature importance and preventing model bias [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. These classic approaches, when combined with deep learning-inspired architecture, can provide a good balance of a hybrid model in movie prediction.\u003c/p\u003e \u003cp\u003eAnother study utilized Heterogeneous Information Network-based recommendation systems (HI2Rec) on MovieLens datasets and reported notable improvements in rating prediction accuracy. Scalability problems were, however, encountered, especially when the model was used on large datasets like IMDb [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. In the meantime, aesthetic-aware matrix factorization models have been introduced to improve content-based filtering by incorporating visual features of movie posters and trailers. These models have been more precise, but their dependency on multimodal data sources is challenging in computational efficiency [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMeta-learning methods have been increasingly popular in movie rating prediction within the domain. A research using Markov Factorization Process (MFMP) for recommendation indicated a notable performance improvement over conventional methods. However, the research mentioned that multimodal data fusion is an open problem [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Another research went as far as exploring long- and short-term information utilization in models from adversarial training, and they were able to make more content-aware recommendations with less overfitting problems [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe research landscape presents that ensemble learning techniques provide a robust methodology by combining strengths of individual models. Recent papers have attained state-of-the-art accuracy in the prediction of ratings for movies through the integration of boosting-based models and deep neural networks. Computational cost, model interpretability, and scalability are still issues of concern. The next section provides an overview of how the Meta-Ensemble Predictor follows up on such advancements to attain improved predictive results while showing a deficiency in current models.\u003c/p\u003e"},{"header":"3. Materials and Methods","content":"\u003cp\u003eThe Meta-Ensemble Predictor (MEP) framework is designed to leverage ensemble learning through fusing an ensemble of machine learning models and deep learning techniques towards improved IMDb Movie rating prediction. The approach that is being followed in the methodology begins with dataset gathering, preprocessing, feature engineering, model training, and ensemble stacking and extends to the best predictions. In Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e,It illustrates the workflow of the suggested framework, which highlights the sequential process of transforming raw movie metadata into predictive information.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset Description\u003c/h2\u003e \u003cp\u003eThe dataset utilized here is used from the Kaggle IMDb Movies from 1960 to 2023 publicly available dataset [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. The dataset comprises 33,600 movies distributed over a span of more than six decades that have an inclusive list of metadata details on every single film. The dataset incorporates crucial variables on which IMDb scores depend, i.e., production-oriented variables, audience measures of engagement, and business-oriented performance measures.\u003c/p\u003e \u003cp\u003eEach entry in the movies dataset has an official title, release date, and runtime, which are sparse context features for rating prediction. The addition of the Motion Picture Association (MPA) rating provides more information about a movie's target audience and possible restriction on viewing.\u003c/p\u003e \u003cp\u003eUser-generated attributes like the rating and vote on IMDb are associated with the prediction task because they capture the reception and engagement of the user. Economic attributes like budget, worldwide gross box office, U.S. and Canadian gross, and opening weekend gross provide useful information about the profitability of a film, a trait that is typically linked with critical success.\u003c/p\u003e \u003cp\u003eMovie crews such as directors, writers, and main actors are significant features of the dataset since they influence the expectations of audiences and the quality of the film. These attributes are preprocessed through encoding and vectorization to make them more convenient for model integration. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e illustrates average movie rating over time while Fig.\u0026nbsp;3 illustrates the distribution of Movie ratings where it is a normal distribution with its mode at mid-range scores.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eCategorical attributes such as genres, production country, shooting location, production companies, and spoken languages are crucial attributes to achieve the regional and cultural impact on the film. Award and award won, Oscars, wins, nominations, provide more richness to the dataset because award-winning films possess a higher IMDb rating. Release date is converted into numerical format so that time patterns that may affect rating prediction might be trapped. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e is language and genre-based rating.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe dataset heterogeneity allows a worldwide approach to predictive modeling, encompassing both objective monetary data and unmeasurable artistic attributes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Data Preprocessing\u003c/h2\u003e \u003cp\u003eThe raw data were preprocessed in detail to be consistent, minimize feature representation loss, and optimize model performance. Due to the heterogeneity of movie features\u0026mdash;ranging from numerical financial quantities to text-based categorical metadata\u0026mdash;a formal multi-step preprocessing pipeline was employed. Steps for missing value imputation, categorical encoding, feature transformation, and normalization were incorporated to pre-process the data for predictive modeling. Figure\u0026nbsp;3 illustrates the end-to-end preprocessing pipeline that converts raw movie metadata into machine-learning-friendly inputs.\u003c/p\u003e \u003cp\u003eOne of the first preprocessing problems was missing values on all attributes. Missing values in the IMDb rating column, the target column, were imputed with the global mean rating so that there was numerical consistency without adding too much bias. Money attributes like budget, worldwide gross, U.S./Canada gross, and opening weekend had many missing values because figures might be unreleased, and so it was imputed so that there was numerical consistency. Rather than dropping these films, missing values were imputed by zero so that the integrity of the dataset wasn't lost and bias from hypothetical imputations was prevented. Categorical attributes like directors, writers, and actors also had missing values as metadata was missing. Missing values in those were imputed with the placeholder \"Unknown\" so that the model was provided structured input instead of excluding informative points.\u003c/p\u003e \u003cp\u003eThe data was also needed to be thoroughly cleaned and normalized, especially for categorical data. Film names contained inconsistencies such as leading digits, extra spaces, and special characters, which were stripped out to make it consistent. The Votes column, which was initially stored in abbreviated format (e.g., \"687K\" for 687,000 votes), was converted to its full numerical value in order to facilitate appropriate numerical analysis. The genre column, being a set of genres for one movie, was transformed to multi-hot encoded vector representation in such a manner that every genre would contribute equally towards the learning process of the model. The model was thus capable of learning how the various sets of genres affected IMDb ratings by being transformed in this fashion.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eSince machine learning algorithms are programmed to take numerical inputs, the categorical variables were converted into numbers through some encoding schemes. For example, high-cardinality categorical features like MPA ratings, directors, writers, production companies, and locations employed label encoding where each category was represented as a numerical value. More sophisticated transformations were required for text features like actors, genres, and languages, though. Term Frequency-Inverse Document Frequency (TF-IDF) vectorization was applied to these features to convert text data into numerical data without losing the relative importance of words and terms. Since TF-IDF matrices are of large dimension, Truncated Singular Value Decomposition (SVD) was applied for reducing the dimension without losing valuable feature information while maintaining computational efficiency.\u003c/p\u003e \u003cp\u003eFeature transformation techniques were required to pre-process the data before it could be fed into machine learning algorithms. Polynomial feature interactions were formed in an attempt to reveal sophisticated interactions of powerful numerical features. Budget by gross revenue interactions, for instance, helped to reveal high-budget films that became box office hits, and vote by IMDb rating interactions helped to reveal user interaction. In addition, the vote column was also very imbalanced, with blockbusters registering unusually high vote counts, as is evident from Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThe last but most important preprocess step was feature standardization. In an improved preprocessing of the data, outlier removal techniques like interquartile range (IQR) filtering were used to eliminate outlier values from budget and revenue columns to avoid models learning noise but typical data. With the use of these preprocess steps, the data not only stabilized but enabled the model to learn more meaningful patterns.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e6\u003c/span\u003e displays word clouds for highly-rated movies, highlighting the most frequently occurring words in successful films.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Proposed Model Architecture\u003c/h2\u003e \u003cp\u003eThe Meta-Ensemble Predictor (MEP) is introduced for ensemble-improving IMDb rating prediction. MEP applies ensemble methods of a lot of machine and deep learning models in trying to develop a universally inclusive generalized and stable predict feature. MEP utilizes four chief models\u0026mdash;CatBoost, XGBoost, LightGBM, and Neural Network\u0026mdash;and winds down with the finishing touch that forms a meta-learning layer used in generalizing and summarizing prediction through Random Forest Regressor. Each model within the ensemble has something unique to contribute, so MEP can catch subtle feature interactions that a single model may miss.\u003c/p\u003e \u003cp\u003eCatBoost is a gradient boosting algorithm specifically designed for categorical feature handling. It is used here to imbed high-cardinality categorical columns such as film directors, stars, and categories that are highly significant in their prediction of viewer acceptability. XGBoost is used as another gradient boosting algorithm for the capture of non-linear interaction effects between numerical columns. It proves useful while performing economic calculations like budget, box office, and voting public and thus becomes the central part of MEP. LightGBM is specially designed for use with massive data and has computing speed with no compromise in accuracy. It proves very handy while performing TF-IDF vectorized text features and high-dimensional categorical features so that more expressiveness in the feature representation is achieved. Additionally, the Neural Network module is capable of introducing deep learning to the system such that it can learn abstract high-level feature representations which may be neglected by gradient boosting models. It uncovers underlying relationships between different features of the movie effectively and allows in-depth analysis of rating behavior.\u003c/p\u003e \u003cp\u003eWhen all base models have predicted, they are passed on to the meta-model, a Random Forest Regressor, which determines the best ensemble of base model predictions to make use of to make the end prediction. The meta-model takes the average across predictions, resists overfitting, and generalizes better by making the end prediction an optimal ensemble of the strengths of numerous models. Hierarchical learning maintains the model variance and bias at their minimum value, and MEP is highly insensitive to variation in movie data.\u003c/p\u003e \u003cp\u003eFigure 7. Proposed Model Diagram\u003c/p\u003e \u003cp\u003eThe Neural Network architecture in MEP architecture of the Neural Network is a multilayer feedforward fully connected neural network with many layers providing feature extraction as well as prediction enhancement.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eInput Layer\u003c/strong\u003e \u003cp\u003eThe numerical attributes and one-hot encoded attributes of categorical attributes are input to the neural network, where both are input as such. They are normalized by Z-score normalization so that they remain uniform in the instance of differing attribute scales.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eHidden Layers\u003c/strong\u003e \u003cp\u003eThe network has three hidden layers, and all of them are thoroughly optimized to increase learning capacity and generalization. The first hidden layer with 1024 neurons uses the ReLU (Rectified Linear Unit) activation function that brings non-linearity into the model. The activation function prevents the vanishing gradient problem and enables the model to learn complex relationships in the data efficiently. The second hidden layer consisting of 512 neurons uses Batch Normalization, speeding up training by normalizing weight distributions. The method stabilizes learning to enable the network to converge quickly and optimize optimally on diverse data distributions. The third hidden layer consisting of 256 neurons uses Dropout Regularization (30%), a method where randomly during training some of the neurons are cut. This method avoids overfitting as it forces the model to learn more regularized feature representations instead of depending on individual neurons. The employment of the three hidden layers enables the deep learning model to learn abstract high-level feature representations, enhancing prediction accuracy and enabling the model to generalize from new data.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eOutput Layer\u003c/strong\u003e \u003cp\u003eIt has a neuron with a linear activation function that provides the predicted final IMDb rating as a real actual value. Training optimization of the neural network is done with the Adam algorithm and the adaptive learning rate having dynamic ranges of values for free convergence. Furthermore, early stopping is applied, wherein training is terminated when no improvement is seen in validation loss. L2 Regularization is employed to combat overfitting even further, in deep layers, managing model complexity and generalization.\u003c/p\u003e \u003c/p\u003e \u003cp\u003eThe meta-learning phase of MEP improves the prediction performance with the help of knowledge from the individual base models. The process starts on the premise that the base models are individually trained on the same pre-processed data and produce unique data feature predictions. The predictions are stacked in a second dataset, which is employed to train the Random Forest meta-model. The meta-model is monitoring the interaction among base model\u003c/p\u003e \u003cp\u003eoutputs with one another and determining the best distribution of weights for fine-tuning the final estimate. The second level of training removes inconsistencies and improves the predictive accuracy by making the most accurate predictions of base models to announce the final decision.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eUnlike the traditional single-model approaches, MEP reaches the best prediction performance by stacking heterogeneous learning paradigms together. The gradient boosting models perform best on structured tabular data, the Neural Network learns high-level abstracted feature representations, and the meta-model stacks the best of every model in the best possible way. With the state-of-the-art stacking-based ensemble learning paradigm, MEP tends to make state-of-the-art IMDb rating predictions with improved robustness, decreased bias, and improved generalization on different movie datasets.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Results and Evaluation","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e4.1. Performance Analysis\u003c/h2\u003e \u003cp\u003eThe Meta-Ensemble Predictor (MEP) was comprehensively validated with the leading machine learning algorithms of the market, such as LightGBM, XGBoost, Gradient Boosting Regressor, and Neural Networks, to check MEP's prediction precision and overallizability. The performances of the models were assessed with four essential metrics, i.e., Root Mean Squared Error (RMSE), R\u0026sup2; Score, Mean Absolute Error (MAE), and Accuracy Within \u0026plusmn;\u0026thinsp;1-Point Deviation. RMSE was employed to estimate the standard deviation of the prediction error, with lower values meaning greater precision in prediction. R\u0026sup2; value was employed to calculate the variance proportion in IMDb ratings explained by the model, with the higher values reflecting good performance. MAE offered a clear measure of mean absolute error, which simplified the comparison of the accuracy of the models. In addition, Accuracy Within \u0026plusmn;\u0026thinsp;1-Point Deviation discussed how often the prediction was in 1-point range from the actual IMDb scores, thus ensuring the model output remained in effectively useful terms. These approximations as a whole provided a correct foundation to judge the efficiency of MEP in contrast to other methods.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the comparative accuracy of MEP compared to baseline models. From the results, it can be seen that MEP has the least RMSE value of 0.4389 and largest R\u0026sup2; of 0.8511 and thus is the top-performing model in terms of predicting IMDb rating. Its 96.96% accuracy under a 1-point error implies MEP's very good consistency of prediction, significantly better than others. LightGBM, being highly computationally efficient, performed an RMSE of 0.7342 and an R\u0026sup2; score of 0.5832, which is significantly lower than MEP.XGBoost, previously bragged about being able to model structured data, only managed a score of 0.3748 for R\u0026sup2;, which indicates low predictability. Gradient Boosting Regressor, while performing better than XGBoost, was unable to outperform MEP's accuracy and error minimization. The Neural Network model, although apt at feature extraction, was prone to overfitting and returned an R\u0026sup2; of only 0.4361, which is lower than MEP's ensemble model.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMODEL PERFORMANCE COMPARISON\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRMSE\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabb\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eR\u0026sup2; Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy (Within \u0026plusmn;\u0026thinsp;1-Point)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMEP (Proposed Model)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabc\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003e0.4389\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.8511\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e96.96%\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLightGBM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.7342\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5832\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e85.55%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.8992\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Tabd\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.3748\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76.88%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Boosting Regressor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.7417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e85.68%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeural Network\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.8540\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.4361\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e90.56%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe Comparative outcomes mirror the strength of MEP's ability to reduce prediction errors at retaining superior generalization on heterogeneous IMDb ratings. 40.2% of the RMSE reduction over LightGBM confirms MEP's ensemble learning approach in grasping intricate patterns in the data in an overall masterful manner. In contrast to XGBoost and LightGBM, which are more variant and less interpretable on categorical-dominant datasets, MEP's hybrid approach balances tree-based learning with deep neural feature extraction to produce a more robust prediction model. The considerably enhanced R\u0026sup2; score of 0.8511 proves that MEP captures more variation in IMDb ratings compared to other approaches. The precision in \u0026plusmn;\u0026thinsp;1-point difference to 96.96% also indicates MEP's practical usability in real life, in the sense that nearly all the predictions are very close to actual IMDb ratings. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e9\u003c/span\u003e illustrates the training loss vs valid loss graph and Fig.\u0026nbsp;10 illustrates Training R\u0026sup2; score vs valid R\u0026sup2; score.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eA deeper Residual analysis was carried out more rigorously to investigate the distribution of prediction errors for various levels of IMDb ratings. A well-calibrated predictive model would have residuals symmetrically spread around zero, reflecting unbiased predictions. Figure\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e11\u003c/span\u003e shows the residual distribution of MEP, and it can be seen that most residuals are tightly clustered around zero, which reflects that MEP is not adding systematic overestimation or underestimation biases. This contrasts with models like XGBoost and LightGBM, which exhibit higher residual variance, especially for well-rated movies. The residual spread remains low for mainstream ratings (4\u0026ndash;8), demonstrating the robustness of MEP in doing well with mainstream movie predictions. Residuals do open up a bit for extreme IMDb ratings (1\u0026ndash;2 and 8\u0026ndash;9), as would be expected from the subjective nature of audience preferences for extremely polarizing films.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure 12, is the predicted rating vs true ratings plot, The points are tightly clustered about the red dashed line, meaning that the predictions are very close to the actual ratings. The spread of the prediction is not highly skewed or biased, i.e., the model generalizes very well for different rating values.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Discussion\u003c/h2\u003e \u003cp\u003eThe Meta-Ensemble Predictor (MEP) shows a remarkable IMDb rating prediction using ensemble learning and feature engineering methods. In contrast to the conventional models like Random Forest, XGBoost, and LightGBM, which are not efficient when dealing with high-cardinality categorical features, feature interactions, and non-linear relationships, MEP combines several base models and a meta-model to offer more accurate predictions. Adding CatBoost, XGBoost, LightGBM, and a Neural Network enables MEP to identify complex patterns between the movie metadata, which has enhanced predictive accuracy.\u003c/p\u003e \u003cp\u003eOne of the strongest arguments MEP wins is that it boasts a complete engineering pipeline in the form of TF-IDF text feature vectorization, polynomial feature expansion, and Bayesian hyperparameter optimization. The prior work utilized rudimentary feature encoding methods and thus recorded accuracy levels of between 70% and 92.7%. Furthermore, much of the previous work was marred by small datasets and was thus non-generalizable. MEP, with 33,600 movies on training over the span of six decades, easily obviates this drawback with thorough generalization of a plenty of film characteristics.\u003c/p\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents a comparison of MEP's performance with existing research, demonstrating its higher accuracy and generalization capability.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePERFORMANCE COMPARISON OF MODELS\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eReference\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eModel\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eThis Work\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eMeta-Ensemble Predictor\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e96.96%\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eXGBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e82.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e92.70%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e70.00%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe results clearly demonstrate that MEP excels other standard machine learning models by achieving the highest accuracy (96.96%) while regulating overfitting and model robustness. Ensemble learning enables MEP to leverage strengths of different models, leading to stability across diverse rating distributions. In comparison to XGBoost-based studies (accuracy\u0026thinsp;~\u0026thinsp;82%), susceptible to categorical encoding, MEP is found to be robust with actor, director, and genre-dependent attributes with enhanced categorical support using CatBoost.\u003c/p\u003e \u003cp\u003eAnother most significant reason for the supremacy of MEP is that MEP can process giant-sized datasets. Most of the previous models are trained on comparatively much smaller datasets in terms of movies, thereby poor general representation of rating patterns. But MEP is trained on 33,600 movies, thereby a better general understanding of user rating patterns over a period of several decades. Second, the old models lack the ability to convincingly mimic top-level interactions between attributes, whereas MEP uses a Random Forest meta-model to average and smooth the predictions to remove variance and include generalization.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn addition, earlier research experiments have demonstrated high residual variance, especially when handling outlier IMDb ratings. MEP's residual analysis prefers uniform error distribution in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e13\u003c/span\u003e, as evidenced by the fact that the model accurately captures the pattern of the ratings without being overly affected by outliers. The ensemble architecture also prefers robustness, protecting against any of the base models individualizing predictions and thus creating lower margins for error and stabilizing ratings. Through the use of sophisticated ensemble learning, feature engineering, and hyperparameter tuning, MEP sets a new standard for IMDb rating prediction. By integrating conventional machine learning with deep learning techniques, MEP is more precise, generalizes better, and has more stable predictions, making it a very efficient predictive analytics framework in the entertainment industry.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Limitations and Future Work","content":"\u003cp\u003eWhile the Meta-Ensemble Predictor (MEP) is heavily supplemented to forecast IMDb ratings, it is not without its flaws. One of its major flaws is computational complexity. The ensemble learning method, particularly with many base models and a meta-model, is very computationally intensive and training is very time-consuming as opposed to single-model strategies. It is thus less scalable to real-time where inference speed is crucial.\u003c/p\u003e \u003cp\u003eAgain, the limitation is to be vulnerable to overfitting in the treatment of deep feature transformation and stacking-based learning. Bayesian hyperparameter optimization and dropout regularization mitigate the limitation but must be further optimized such that it generalizes much more to fresh data. Apart from this are certain extrinsic factors like movie critics' scepticism, audience sentiment, and social opinion not modelled in an explicit format and that drive the performance of prediction.\u003c/p\u003e \u003cp\u003eFuture work will involve maximizing computational cost with synergistically guiding effective model distillation and pruning methods towards optimizing maximum MEP for real-time application. Real-time powered sentiment analysis of social media and review websites can similarly be applied to improve rating prediction accuracy by tracking dynamic audience mood. Application of transformer models such as BERT or GPT models to text feature extraction can similarly include contextual movie metadata awareness. Finally, semi-supervised learning techniques from unlabelled data can be applied theoretically on MEP to integrate legacy recommendation systems.\u003c/p\u003e"},{"header":"Conclusion","content":" \u003cp\u003eThis study introduced the Meta-Ensemble Predictor (MEP), a new hybrid model that integrates gradient boosting models (CatBoost, XGBoost, LightGBM) with a deep learning-based Neural Network, meta-optimized using a Random Forest Regressor. The integration of heterogeneous learning strategies enabled MEP to learn intricate feature interactions, reduce overfitting, and improve generalization for IMDb rating prediction.\u003c/p\u003e \u003cp\u003eBy using sophisticated feature engineering methods, such as TF-IDF vectorization of text features, polynomial expansion of features, and Bayesian optimization of hyperparameters, MEP scored higher in its prediction accuracy with 96.96%\u0026mdash;far superior to that of traditional machine learning methods. Comparative study also established MEP's superiority over current methodologies by being able to model high-dimensional data precisely and predict more uniformly with a high diversity of movie features.\u003c/p\u003e \u003cp\u003eThe results emphasize the significance of ensemble learning for predictive modeling, especially where user ratings based on subjective opinions are driven by multiple influencing factors. With its success in combining ordered numerical data with text and category features, MEP provides a robust and scalable solution for further breakthroughs in IMDb rating prediction and other wider uses in entertainment analytics.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eM.F.A.R. (Md. Faishal Ahmed Rudro) conceptualized and designed the study, developed the Meta-Ensemble Predictor (MEP) framework, and performed the primary machine learning model implementation. He also wrote the main manuscript text and conducted data preprocessing, feature engineering, and model evaluation.M.S.R.B. (Md. Shahriar Rahman Bhuiyan) contributed to dataset acquisition, preprocessing, and exploratory data analysis. He also assisted in hyperparameter tuning and the implementation of gradient boosting models (CatBoost, LightGBM, XGBoost).A.D. (Abu Dojana) worked on deep neural network model integration and contributed to performance comparison with state-of-the-art methods. He was also involved in fine-tuning the Random Forest meta-learner.M.R.H.S. (Md. Reduanul Haque Shakib) assisted in statistical analysis, residual evaluation, and model validation. He prepared figures and tables summarizing model performance metrics and contributed to result interpretation.A.A.O. (Arisha Ashraf Oishi) contributed to literature review compilation, manuscript formatting, and references organization. She also assisted in ensuring compliance with journal submission requirements.All authors reviewed the manuscript and approved the final version for submission.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe dataset used in this study, titled \"IMDb Movies from 1960 to 2023\", is publicly available on Kaggle at the following link:\u0026#128279; https://www.kaggle.com/datasets/raedaddala/imdb-movies-from-1960-to-2023This dataset contains metadata on 33,600 movies spanning from 1960 to 2024, including attributes such as genre, director, cast, runtime, box office revenue, and IMDb user ratings. The data was preprocessed and analyzed for model training and validation in this study.Researchers interested in replicating the results can access the dataset through the provided Kaggle link. Any additional preprocessing scripts and model configurations used in this study can be shared upon reasonable request to the corresponding author.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eInternet Movie Data Base URL: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.imdb.com/\u003c/span\u003e\u003cspan address=\"https://www.imdb.com/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePersson K (2015) Predicting movie ratings: A comparative study on random forests and support vector machines. Bachelor Degree Project in Informatics, University of Sk\u0026ouml;vde\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTashman M (2015) The Association Between Film Industry Success and Prior Career History: A Machine Learning Approach, Master's thesis, Harvard Extension School\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiddique A, Ahmed M, Rahman S (2023) Movies Rating Prediction Using Supervised Machine Learning Techniques. Int J Inform Syst Comput Technol (IJISCT) 6(2):45\u0026ndash;51\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar S, De K, Roy PP (2020) IEEE Trans Comput Soc Syst 7(4):915\u0026ndash;923. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TCSS.2020.2993585\u003c/span\u003e\u003cspan address=\"10.1109/TCSS.2020.2993585\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Movie Recommendation System Using Sentiment Analysis from Microblogging Data\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbidin D, Bostanci C, Site A (2018) Movie Rating Prediction with Machine Learning Algorithms on IMDB Data Set, in \u003cem\u003eProceedings of the International Conference on Advanced Technologies, Computer Engineering and Science (ICATCES'18)\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDixit P, Verma R, Tiwari A (2021) Predicting the IMDB Rating by Using EDA and Machine Learning Algorithms, \u003cem\u003eInternational Journal of Scientific Research in Computer Science, Engineering and Information Technology (IJSRCSEIT)\u003c/em\u003e, vol. 7, no. 1, pp. 55\u0026ndash;63\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J, Wang Y, Yuan Z, Jin Q (2020) Personalized Real-Time Movie Recommendation System: Practical Prototype and Evaluation, pp. 180\u0026ndash;191\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrownlee J (2019) A Gentle Introduction to Gradient Boosting Algorithms for Machine Learning, \u003cem\u003eMachine Learning Mastery\u003c/em\u003e, [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://machinelearningmastery.com/gradient-boosting-algorithms/\u003c/span\u003e\u003cspan address=\"https://machinelearningmastery.com/gradient-boosting-algorithms/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen K, Wu Z (2022) Feature Engineering for Movie Rating Prediction Using Hybrid Machine Learning Models. IEEE Trans Knowl Data Eng 34(7):2501\u0026ndash;2515\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSivakumar P, Kumar R, Sharma SK (2020) Movie Success and Rating Prediction Using Data Mining Algorithms, International Conference on Computational Intelligence and Data Science (ICCIDS)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson T, Patel M (2022) Machine Learning-Based Analysis of Movie Ratings Using IMDb Data, \u003cem\u003eInternational Conference on Artificial Intelligence and Data Analytics (ICAIDA)\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIMDb D (1960\u0026ndash;2024) Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.kaggle.com/datasets/raedaddala/imdb-movies-from-1960-to-\u003c/span\u003e\u003cspan address=\"https://www.kaggle.com/datasets/raedaddala/imdb-movies-from-1960-to-\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cem\u003e2023.\u003c/em\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHastie T, Tibshirani R, Friedman J (2009) The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edn. Springer\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBishop C (2006) Pattern Recognition and Machine Learning, 1st edn. Springer\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe M, Wang BO, Du X (2019) HI2Rec: Exploring Knowledge in Heterogeneous Information for Movie Recommendation. IEEE Access 7:30276\u0026ndash;30284. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2019.2902398\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2019.2902398\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X et al (2019) Exploiting Aesthetic Features in Visual Contents for Movie Recommendation, IEEE Access, vol. PP, no. c, p. 1. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/ACCESS.2019.2910722\u003c/span\u003e\u003cspan address=\"10.1109/ACCESS.2019.2910722\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X, Liu D, Member S, Xiong Z (2020) Learning and Fusing Multiple User Interest Representations for Micro-Video and Movie Recommendations, vol. 9210, no. c, pp. 1\u0026ndash;13. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TMM.2020.2978618\u003c/span\u003e\u003cspan address=\"10.1109/TMM.2020.2978618\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao W et al (2019) Leveraging Long and Short-Term Information in Content-Aware Movie Recommendation via Adversarial Training. IEEE Trans Cybern PP:1\u0026ndash;14. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1109/TCYB.2019.2896766\u003c/span\u003e\u003cspan address=\"10.1109/TCYB.2019.2896766\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":true,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"IMDb Rating Prediction, Meta-Ensemble Learning, Gradient Boosting, Deep Neural Networks, Feature Engineering, Hybrid Models, Machine Learning","lastPublishedDoi":"10.21203/rs.3.rs-6153928/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6153928/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePrecise IMDb movie ratings predictions are vital to the stakeholders of the film industry since it will dictate investment, marketing, and content recommendation. Standard machine learning techniques are not capable of handling high-cardinality categorical columns, bad interactions between features, and non-linear relations between reviews and metadata. This paper proposes the Meta-Ensemble Predictor (MEP), a new state-of-the-art hybrid framework that integrates various gradient boosters (CatBoost, LightGBM, XGBoost) among themselves and with a deep neural network and tuned using a meta-learning algorithm with a Random Forest as the final predictor. With a database of 33,600 movies from 1960 to 2024 having metadata details like genre, director, cast, runtime, box office, and voter ratings, MEP model uses TF-IDF text feature vectorization, polynomial interaction of features, and dimensionality reduction methods for improved feature representation. The model presented here has achieved RMSE of 0.4389 and accuracy of 96.96% at a difference of 1-point from actual IMDb ratings when compared to other state-of-the-art models. The study showcases the strength of ensemble learning to learn the sophisticated patterns of the ratings of the movies in becoming a good film industry predictive analytics tool.\u003c/p\u003e","manuscriptTitle":"Meta-Ensemble Learning for IMDb Ratings: A Stacked Hybrid Model Integrating Gradient Boosting and Deep Neural Networks","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-07 11:24:35","doi":"10.21203/rs.3.rs-6153928/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"adaf2cf4-093a-4083-8c37-faaec1024d03","owner":[],"postedDate":"March 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-03-12T12:23:34+00:00","versionOfRecord":[],"versionCreatedAt":"2025-03-07 11:24:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6153928","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6153928","identity":"rs-6153928","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}