A Hybrid Movie Recommendation System Using BERT-Based Semantic Embeddings and SVD Collaborative Filtering

A Hybrid Movie Recommendation System Using BERT-Based Semantic Embeddings and SVD Collaborative Filtering | 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 A Hybrid Movie Recommendation System Using BERT-Based Semantic Embeddings and SVD Collaborative Filtering Ahmad Yasser Alnahhas This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8829543/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 On modern digital platforms, recommender systems are crucial in assisting users in finding pertinent content. Users frequently find it difficult to find products that fit their individual interests as the amount of information available keeps expanding. While content-based methods rely on item information like titles and genres, traditional collaborative filtering techniques learn from past user ratings. Despite their widespread use, both strategies have definite drawbacks. While content-based systems frequently struggle to fully comprehend the meaning of text, collaborative filtering suffers when data is sparse or when new users or items appear. The hybrid movie recommendation system proposed in this paper combines the advantages of both methods. A pre-trained DistilBERT model that can extract contextual meaning from text is used to convert movie titles and genres into semantic vectors. Simultaneously, Singular Value Decomposition (SVD) is used to model user preferences based on rating data from the past. Using a weighted linear fusion strategy, the outputs of the two models are combined to determine the final recommendation score. Standard Top-N evaluation metrics were used to conduct the experiments on the MovieLens dataset. The hybrid model consistently performs better than standalone BERT-based and SVD-based systems, according to the results. The model achieves an NDCG@10 value near 0.91 at the ideal fusion weight, indicating robust and realistic ranking performance. These results show that, particularly in cold-start and sparse-data scenarios, combining semantic content understanding with collaborative signals results in more precise, reliable, and customized movie recommendations. This study presents a lightweight hybrid recommendation framework that balances semantic understanding and collaborative learning. Information Retrieval and Management Recommender Systems Hybrid Recommendation BERT SVD Collaborative Filtering Deep Learning Text Analysis Figures Figure 1 1. Introduction Online platforms provide users with access to large volumes of multimedia and commercial content. Although this abundance provides a wide range of options, it also poses a significant problem known as information overload. Without extra assistance, users frequently struggle to find content that suits their interests. Because recommender systems make it easier for users to find relevant content quickly, they have become an essential part of contemporary digital services. One of the best methods for making recommendations is collaborative filtering. To discover latent user preferences, techniques like Singular Value Decomposition (SVD) examine user-item rating patterns. When sufficient interaction data is available, these methods work well. However, in cold-start scenarios, such as when new users or items enter the system or when rating data is sparse, their performance drastically declines. A different strategy is used by content-based recommendation systems. They are able to suggest new or unrated items by concentrating on item attributes and textual descriptions. However, traditional content-based approaches frequently rely on keyword matching and basic text representations, which restricts their capacity to capture deeper semantic meaning. Transformer-based models, like BERT, have been introduced by recent developments in natural language processing. Text can be meaningfully and contextually represented by these models. In order to develop a more efficient and well-rounded recommendation system, this paper investigates the integration of collaborative filtering with BERT-based semantic representations. Unlike recent neural and graph-based recommender systems that require complex architectures, the proposed approach focuses on a lightweight and interpretable hybrid design. Research Contributions This study proposes a practical and efficient hybrid movie recommendation framework that integrates transformer-based semantic embeddings with collaborative filtering. The main contributions of this work can be summarized as follows. First, it introduces a lightweight hybrid recommendation approach that combines BERT-based contextual text representations with SVD-based collaborative filtering to improve recommendation accuracy while maintaining computational efficiency. Second, it demonstrates how semantic information derived from movie metadata can complement user rating patterns, particularly in cold-start and sparse-data scenarios. Third, the study provides an empirical evaluation of the hybrid framework using standard recommendation metrics, highlighting its robustness and consistent performance across different fusion configurations. Finally, the proposed approach offers an interpretable and deployable solution that balances performance, simplicity, and scalability for real-world recommendation environments. 2. Related Work Matrix factorization techniques, particularly Singular Value Decomposition (SVD), were popularized by Koren et al. and remain a strong baseline for collaborative filtering due to their simplicity, scalability, and effectiveness. These methods extract latent representations of users and items from rating matrices and have been widely adopted in real-world recommender systems. Despite the emergence of more complex deep learning approaches, SVD-based models continue to provide competitive performance and are frequently used as reference baselines in recommendation research. Content-based recommendation has evolved from traditional TF–IDF and bag-of-words approaches to deep learning-based methods. Transformer models such as BERT have been adopted for semantic feature extraction in recommendation systems, demonstrating improved contextual understanding [ 1 ]. More recent methods extract richer representations from text using deep learning models. In a variety of natural language tasks, transformer-based models like BERT have demonstrated excellent performance. BERT has been used in a number of studies to enhance recommendation systems' semantic comprehension of item descriptions. These strategies outperform more conventional text-based techniques. Hybrid recommender systems aim to integrate collaborative filtering with content-based signals to improve recommendation accuracy and robustness. Recent approaches include neural architectures, graph-based models such as LightGCN [ 3 ], and generative frameworks like GenRec [ 8 ], which achieve strong performance but often require substantial computational resources and complex system design. In contrast, lightweight hybrid strategies that combine transformer-based semantic modeling with classical collaborative filtering remain relatively underexplored. This study addresses this gap by proposing an efficient and interpretable hybrid framework that balances performance with computational practicality. 3. Methodology 3.1 Dataset and Preprocessing The MovieLens dataset, widely used in recommendation research, is employed to evaluate the proposed system. The MovieLens dataset, consisting of: (1) movies.csv : movieId, title, genres. (2) ratings_small.csv : userId, movieId, rating, timestamp. User ratings and basic movie details are included in the dataset. The movie identifier is used to combine the datasets. The title and genre details are combined into a single text field to create a textual representation for each film. The BERT model uses this text as input. Since temporal modeling is not taken into account in this study, the timestamp attribute has been removed. In order to ensure fair evaluation, the data are finally split 70/30 between training and testing sets. The system was implemented in Python using standard machine learning and natural language processing libraries, including the Surprise library for collaborative filtering and transformer-based frameworks for semantic modeling. 3.2 Content-Based Recommendation Using BERT A pre-trained DistilBERT model is used to model semantic content information. DistilBERT is a lighter version of BERT that lowers computational costs without sacrificing performance. The model is used to create a dense vector embedding from each movie's text representation. In a continuous vector space, these embeddings represent the semantic meaning of film titles and genres. The similarity between movies is then calculated using cosine similarity. Stronger semantic relationships are indicated by higher similarity values. This makes it possible for the system to suggest films that are conceptually similar even if they don't have the same keywords. The final embedding representation is obtained using the pooled output of the DistilBERT model and used as the semantic vector for each movie. 3.3 Collaborative Filtering Using SVD Singular Value Decomposition from the Surprise library is used to implement collaborative filtering. Lower-dimensional latent factors that reflect user preferences and item attributes are extracted from the user–item rating matrix. By reducing the Root Mean Squared Error between expected and actual ratings, the model gains knowledge of these variables. With this method, the system can identify hidden trends in user behavior. SVD offers strong personalization and accurate predictions when there is enough rating data available. Its combination with BERT-based features is motivated by the fact that it does not utilize any content information. 3.4 Hybrid Recommendation Model Using a weighted linear formula, the hybrid model integrates the collaborative filtering prediction with the content-based similarity score. Each component's contribution is managed by a single parameter. The system can prioritize semantic content or collaborative information by varying this weight. The model is easy to understand and adjust thanks to this straightforward fusion technique. Additionally, it enables the system to take advantage of semantic information without overpowering the collaborative signal. Achieving balanced performance under various data conditions is the goal of the hybrid design. The final recommendation score is computed using a weighted linear combination: $$\:Scor{e}_{final}=\alpha\:\cdot\:Scor{e}_{BERT}+(1-\alpha\:)\cdot\:Scor{e}_{SVD}$$ where \(\:\alpha\:\in\:\left[\text{0,1}\right]\) controls the contribution of content-based versus collaborative signals. This formulation enables controlled integration of semantic and collaborative signals while maintaining model interpretability. 4. Experimental Setup 4.1 Evaluation Metrics Commonly used Top-N metrics, such as Precision@k, Recall@k, and NDCG@k, are used to assess the recommendation system's performance. These metrics assess how well the system places pertinent items close to the top of the list of recommendations. Additionally, the collaborative filtering component's rating prediction accuracy is assessed using RMSE. Because it takes into account the position of pertinent items and gives earlier-ranked items more weight, NDCG is especially significant. 4.2 Compared Models Three recommendation models were compared in this study. Only BERT-based content similarity is used in the first model. The second model only uses collaborative filtering based on SVD. The suggested hybrid system, which blends the two methods, is the third model. This comparison demonstrates the advantages of the hybrid design and enables a clear analysis of how each component contributes to overall performance. 5. Results According to the experimental findings, the SVD model demonstrates strong performance in rating prediction, achieving an RMSE of 0.875. This outcome is in line with earlier research on matrix factorization techniques [ 2 ]. The BERT-based model is less successful as a stand-alone personalized recommender, but it shows strong capability in handling cold-start scenarios and identifying semantically related movies. Overall, the hybrid model achieves the best performance. When a small amount of semantic content information from BERT is added to the collaborative signal generated by SVD, the highest performance is obtained. The model achieves an NDCG@10 value of approximately 0.91 at the ideal fusion weight. Performance gradually declines as the weight of the content-based (BERT) component increases. This demonstrates that when sufficient rating data are available, collaborative filtering continues to be the dominant signal while BERT-based features offer useful supplementary information. In conclusion, the results indicate that the hybrid model achieves optimal performance at \(\:\alpha\:=0.1\) , reflecting a higher contribution from SVD than BERT. Table 1 reports the sensitivity analysis of the fusion weight ( \(\:\alpha\:\) ) across multiple evaluation metrics. Table 1 Performance of the hybrid recommendation system for different fusion weights (α). α Precision@10 Recall@10 NDCG@5 NDCG@10 0.0 0.6916 0.5351 0.7902 0.8647 0.1 0.7688 0.6141 0.8571 0.9105 0.2 0.5574 0.4639 0.7330 0.8018 0.3 0.4246 0.3136 0.6089 0.6930 0.4 0.2922 0.2274 0.4424 0.5258 0.5 0.1517 0.1411 0.2760 0.3585 0.6 0.1076 0.1067 0.2155 0.2841 0.7 0.0777 0.0719 0.1550 0.2097 0.8 0.0655 0.0643 0.1441 0.1955 0.9 0.0635 0.0567 0.1332 0.1812 1.0 0.0600 0.0535 0.1239 0.1711 To further illustrate this behavior, Fig. 1 visualizes the sensitivity of recommendation performance to the fusion weight \(\:\alpha\:\) on Precision@10, Recall@10, NDCG@5, and NDCG@10. The hybrid approach outperforms both standalone models and demonstrates competitive performance compared with recent state-of-the-art methods such as LightGCN [ 3 ] and GenRec [ 8 ]. 6. Discussion The experimental findings show that, in the presence of adequate user–item interaction data, collaborative filtering continues to be the main factor influencing recommendation accuracy. The SVD model's impressive performance attests to the ongoing efficacy of matrix factorization methods in identifying latent item relationships and user preferences. These results are in line with earlier studies that demonstrate the resilience of collaborative filtering in settings with a lot of data. Nonetheless, quantifiable gains in recommendation quality are achieved through the incorporation of semantic content representations produced by BERT, especially in ranking-based metrics. The hybrid model gains from the complementary nature of the two methods: BERT adds contextual knowledge of movie metadata, whereas SVD records behavioral patterns gleaned from past interactions. Because of this combination, the system can suggest semantically related items that might not be highly connected based just on user ratings. The findings also show that scenarios with cold-start conditions or sparse data benefit most from the contribution of semantic features. Semantic similarity derived from textual features helps maintain recommendation relevance when rating information is scarce. On the other hand, collaborative signals take center stage as interaction data volume rises, and content-based features' relative contribution becomes secondary rather than essential. The hybrid model's stability across a range of fusion weights is another noteworthy finding. In a variety of configurations, the hybrid framework consistently outperforms the standalone models, even though performance peaks at a particular balance between collaborative and semantic signals. This implies that integrating heterogeneous recommendation signals enhances resilience and lessens susceptibility to changes in the distribution of data. The suggested system offers a useful trade-off between computational complexity and performance when compared to more recent deep learning-based recommendation models, such as large language model-driven recommenders and graph neural network approaches. Even though state-of-the-art models frequently attain high accuracy, they usually call for sophisticated architectures and substantial computational resources. On the other hand, the suggested hybrid framework is still computationally efficient, interpretable, and easier to deploy in practical systems. From the standpoint of the application, dynamic web environments with regularly added content are best suited for the hybrid design. Even with limited historical interaction data, the system can produce meaningful recommendations thanks to its ability to incorporate semantic information. Because of this, the strategy is applicable to e-commerce platforms, streaming platforms, and new digital services that are constantly growing their content libraries. Notwithstanding these benefits, the results also draw attention to a number of factors. The prevalence of collaborative filtering suggests that recommendation performance is still heavily influenced by the density and quality of rating data. Furthermore, the semantic component's representational richness may be constrained by its reliance on comparatively simple textual features. Richer content sources might be incorporated into future improvements to fortify the semantic modeling element even more. All things considered, the findings demonstrate that hybrid recommendation techniques offer a sensible and practical path forward for contemporary recommender systems. The proposed method maintains computational efficiency and practical applicability while improving personalization, ranking quality, and resilience to sparse-data conditions through the integration of behavioral and semantic signals. 7. Conclusion, Limitations, and Future Work This study introduced a hybrid movie recommendation system that combines collaborative filtering based on SVD with semantic text embeddings based on BERT. The suggested method produces reliable and accurate recommendations by fusing latent user preference modeling with deep content understanding. Overall, the study demonstrates that lightweight hybrid recommender systems can offer a practical balance between accuracy, interpretability, and deployment feasibility in real-world environments. The proposed approach highlights the practical potential of lightweight hybrid recommender systems for real-world deployment in dynamic content platforms. Limitations The system depends on a small amount of textual data, namely the genres and titles of movies. The richness of semantic representations is thus limited. The fusion strategy does not adjust to specific users; instead, it employs a fixed global weight. Additionally, sequential behavior and temporal dynamics are not taken into account. Future Work Richer textual sources, like user reviews and plot summaries, will be investigated in subsequent work. Adaptive fusion strategies, temporal modeling, graph-based collaborative filtering techniques, and fine-tuning BERT on domain-specific data are possible additional enhancements. References Devlin J, Chang M-W, Lee K, Toutanova K (2019) BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, Proc. NAACL-HLT Koren Y, Bell R, Volinsky C (2009) Matrix Factorization Techniques for Recommender Systems. Computer 42(8):30–37 He X, Deng K, Wang X, Li Y, Zhang Y, Wang M (2020) LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation, Proc. SIGIR Jing Y, Wang Y, Feng Y, Yu PS (2023) Contrastive Self-supervised Learning in Recommender Systems: A Survey. IEEE Trans Knowl Data Eng Zhao WX, Mu S, Hou Y, Lin Z, Li Y, Wang X, Wen J-R Recommender Systems in the Era of Large Language Models, arXiv:2307.02046, 2023. McAuley J, Leskovec J (2013) Hidden Factors and Hidden Topics: Understanding Rating Dimensions with Review Text, Proc. ACM RecSys Tang J, Wang K (2018) Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding (Caser), Proc. WWW Ji J, Li Z, Xu S, Hua W, Ge Y, Tan J, Zhang Y GenRec: Large Language Model for Generative Recommendation, arXiv:2307.00457, 2023. Zhou X, Ma Y, Wang Y, Liu W, Li H (2023) A Comprehensive Survey of Recommender Systems Based on Deep Learning. ACM Comput Surv Gheewala A, Shah K, Patel D (2025) In-depth Survey of Deep Learning Techniques in Recommender Systems. Inf Syst Front Additional Declarations The authors declare no competing interests. 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-8829543","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":588211444,"identity":"4d870fec-8c68-443f-85b6-73003e3eae5d","order_by":0,"name":"Ahmad Yasser Alnahhas","email":"data:image/png;base64,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","orcid":"","institution":"Damascus University","correspondingAuthor":true,"prefix":"","firstName":"Ahmad","middleName":"Yasser","lastName":"Alnahhas","suffix":""}],"badges":[],"createdAt":"2026-02-09 11:05:42","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-8829543/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8829543/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102291323,"identity":"36d25553-8631-40a3-901c-54b7d8e6d81e","added_by":"auto","created_at":"2026-02-10 09:20:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":47411,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eSensitivity of recommendation performance to the fusion weight α.\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-8829543/v1/18e1cab14b0ffa49f496f07d.png"},{"id":102291325,"identity":"e5773e57-6c63-44cf-9b0e-b12664e0d0b4","added_by":"auto","created_at":"2026-02-10 09:20:42","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":565428,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8829543/v1/4fcd333d-aa82-4763-9933-2fbfde004ecf.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eA Hybrid Movie Recommendation System Using BERT-Based Semantic Embeddings and SVD Collaborative Filtering\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eOnline platforms provide users with access to large volumes of multimedia and commercial content. Although this abundance provides a wide range of options, it also poses a significant problem known as information overload. Without extra assistance, users frequently struggle to find content that suits their interests. Because recommender systems make it easier for users to find relevant content quickly, they have become an essential part of contemporary digital services.\u003c/p\u003e \u003cp\u003eOne of the best methods for making recommendations is collaborative filtering. To discover latent user preferences, techniques like Singular Value Decomposition (SVD) examine user-item rating patterns. When sufficient interaction data is available, these methods work well.\u003c/p\u003e \u003cp\u003eHowever, in cold-start scenarios, such as when new users or items enter the system or when rating data is sparse, their performance drastically declines.\u003c/p\u003e \u003cp\u003eA different strategy is used by content-based recommendation systems. They are able to suggest new or unrated items by concentrating on item attributes and textual descriptions. However, traditional content-based approaches frequently rely on keyword matching and basic text representations, which restricts their capacity to capture deeper semantic meaning.\u003c/p\u003e \u003cp\u003eTransformer-based models, like BERT, have been introduced by recent developments in natural language processing. Text can be meaningfully and contextually represented by these models. In order to develop a more efficient and well-rounded recommendation system, this paper investigates the integration of collaborative filtering with BERT-based semantic representations. Unlike recent neural and graph-based recommender systems that require complex architectures, the proposed approach focuses on a lightweight and interpretable hybrid design.\u003c/p\u003e \u003cp\u003e \u003cb\u003eResearch Contributions\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThis study proposes a practical and efficient hybrid movie recommendation framework that integrates transformer-based semantic embeddings with collaborative filtering. The main contributions of this work can be summarized as follows. First, it introduces a lightweight hybrid recommendation approach that combines BERT-based contextual text representations with SVD-based collaborative filtering to improve recommendation accuracy while maintaining computational efficiency. Second, it demonstrates how semantic information derived from movie metadata can complement user rating patterns, particularly in cold-start and sparse-data scenarios. Third, the study provides an empirical evaluation of the hybrid framework using standard recommendation metrics, highlighting its robustness and consistent performance across different fusion configurations. Finally, the proposed approach offers an interpretable and deployable solution that balances performance, simplicity, and scalability for real-world recommendation environments.\u003c/p\u003e"},{"header":"2. Related Work","content":"\u003cp\u003eMatrix factorization techniques, particularly Singular Value Decomposition (SVD), were popularized by Koren et al. and remain a strong baseline for collaborative filtering due to their simplicity, scalability, and effectiveness. These methods extract latent representations of users and items from rating matrices and have been widely adopted in real-world recommender systems. Despite the emergence of more complex deep learning approaches, SVD-based models continue to provide competitive performance and are frequently used as reference baselines in recommendation research.\u003c/p\u003e \u003cp\u003eContent-based recommendation has evolved from traditional TF\u0026ndash;IDF and bag-of-words approaches to deep learning-based methods. Transformer models such as BERT have been adopted for semantic feature extraction in recommendation systems, demonstrating improved contextual understanding [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. More recent methods extract richer representations from text using deep learning models.\u003c/p\u003e \u003cp\u003eIn a variety of natural language tasks, transformer-based models like BERT have demonstrated excellent performance. BERT has been used in a number of studies to enhance recommendation systems' semantic comprehension of item descriptions. These strategies outperform more conventional text-based techniques.\u003c/p\u003e \u003cp\u003eHybrid recommender systems aim to integrate collaborative filtering with content-based signals to improve recommendation accuracy and robustness. Recent approaches include neural architectures, graph-based models such as LightGCN [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], and generative frameworks like GenRec [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e], which achieve strong performance but often require substantial computational resources and complex system design. In contrast, lightweight hybrid strategies that combine transformer-based semantic modeling with classical collaborative filtering remain relatively underexplored. This study addresses this gap by proposing an efficient and interpretable hybrid framework that balances performance with computational practicality.\u003c/p\u003e"},{"header":"3. Methodology","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Dataset and Preprocessing\u003c/h2\u003e \u003cp\u003eThe MovieLens dataset, widely used in recommendation research, is employed to evaluate the proposed system. The MovieLens dataset, consisting of: (1) \u003cb\u003emovies.csv\u003c/b\u003e: movieId, title, genres. (2) \u003cb\u003eratings_small.csv\u003c/b\u003e: userId, movieId, rating, timestamp. User ratings and basic movie details are included in the dataset.\u003c/p\u003e \u003cp\u003eThe movie identifier is used to combine the datasets. The title and genre details are combined into a single text field to create a textual representation for each film. The BERT model uses this text as input. Since temporal modeling is not taken into account in this study, the timestamp attribute has been removed. In order to ensure fair evaluation, the data are finally split 70/30 between training and testing sets. The system was implemented in Python using standard machine learning and natural language processing libraries, including the Surprise library for collaborative filtering and transformer-based frameworks for semantic modeling.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Content-Based Recommendation Using BERT\u003c/h2\u003e \u003cp\u003eA pre-trained DistilBERT model is used to model semantic content information. DistilBERT is a lighter version of BERT that lowers computational costs without sacrificing performance. The model is used to create a dense vector embedding from each movie's text representation.\u003c/p\u003e \u003cp\u003eIn a continuous vector space, these embeddings represent the semantic meaning of film titles and genres. The similarity between movies is then calculated using cosine similarity. Stronger semantic relationships are indicated by higher similarity values. This makes it possible for the system to suggest films that are conceptually similar even if they don't have the same keywords.\u003c/p\u003e \u003cp\u003eThe final embedding representation is obtained using the pooled output of the DistilBERT model and used as the semantic vector for each movie.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Collaborative Filtering Using SVD\u003c/h2\u003e \u003cp\u003eSingular Value Decomposition from the Surprise library is used to implement collaborative filtering. Lower-dimensional latent factors that reflect user preferences and item attributes are extracted from the user\u0026ndash;item rating matrix. By reducing the Root Mean Squared Error between expected and actual ratings, the model gains knowledge of these variables.\u003c/p\u003e \u003cp\u003eWith this method, the system can identify hidden trends in user behavior. SVD offers strong personalization and accurate predictions when there is enough rating data available. Its combination with BERT-based features is motivated by the fact that it does not utilize any content information.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Hybrid Recommendation Model\u003c/h2\u003e \u003cp\u003eUsing a weighted linear formula, the hybrid model integrates the collaborative filtering prediction with the content-based similarity score. Each component's contribution is managed by a single parameter. The system can prioritize semantic content or collaborative information by varying this weight.\u003c/p\u003e \u003cp\u003eThe model is easy to understand and adjust thanks to this straightforward fusion technique. Additionally, it enables the system to take advantage of semantic information without overpowering the collaborative signal. Achieving balanced performance under various data conditions is the goal of the hybrid design.\u003c/p\u003e \u003cp\u003eThe final recommendation score is computed using a weighted linear combination:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Scor{e}_{final}=\\alpha\\:\\cdot\\:Scor{e}_{BERT}+(1-\\alpha\\:)\\cdot\\:Scor{e}_{SVD}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:\\in\\:\\left[\\text{0,1}\\right]\\)\u003c/span\u003e\u003c/span\u003e controls the contribution of content-based versus collaborative signals.\u003c/p\u003e \u003cp\u003eThis formulation enables controlled integration of semantic and collaborative signals while maintaining model interpretability.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Experimental Setup","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Evaluation Metrics\u003c/h2\u003e \u003cp\u003eCommonly used Top-N metrics, such as Precision@k, Recall@k, and NDCG@k, are used to assess the recommendation system's performance. These metrics assess how well the system places pertinent items close to the top of the list of recommendations.\u003c/p\u003e \u003cp\u003eAdditionally, the collaborative filtering component's rating prediction accuracy is assessed using RMSE. Because it takes into account the position of pertinent items and gives earlier-ranked items more weight, NDCG is especially significant.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Compared Models\u003c/h2\u003e \u003cp\u003eThree recommendation models were compared in this study. Only BERT-based content similarity is used in the first model. The second model only uses collaborative filtering based on SVD. The suggested hybrid system, which blends the two methods, is the third model.\u003c/p\u003e \u003cp\u003eThis comparison demonstrates the advantages of the hybrid design and enables a clear analysis of how each component contributes to overall performance.\u003c/p\u003e \u003c/div\u003e"},{"header":"5. Results","content":"\u003cp\u003eAccording to the experimental findings, the SVD model demonstrates strong performance in rating prediction, achieving an RMSE of 0.875. This outcome is in line with earlier research on matrix factorization techniques [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The BERT-based model is less successful as a stand-alone personalized recommender, but it shows strong capability in handling cold-start scenarios and identifying semantically related movies.\u003c/p\u003e \u003cp\u003eOverall, the hybrid model achieves the best performance. When a small amount of semantic content information from BERT is added to the collaborative signal generated by SVD, the highest performance is obtained. The model achieves an NDCG@10 value of approximately 0.91 at the ideal fusion weight.\u003c/p\u003e \u003cp\u003ePerformance gradually declines as the weight of the content-based (BERT) component increases. This demonstrates that when sufficient rating data are available, collaborative filtering continues to be the dominant signal while BERT-based features offer useful supplementary information.\u003c/p\u003e \u003cp\u003eIn conclusion, the results indicate that the hybrid model achieves optimal performance at \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:=0.1\\)\u003c/span\u003e\u003c/span\u003e, reflecting a higher contribution from SVD than BERT. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e reports the sensitivity analysis of the fusion weight (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:\\)\u003c/span\u003e\u003c/span\u003e) across multiple evaluation metrics.\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\u003ePerformance of the hybrid recommendation system for different fusion weights (α).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eα\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePrecision@10\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRecall@10\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNDCG@5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eNDCG@10\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.6916\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5351\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7902\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8647\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.7688\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.6141\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.8571\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.9105\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.5574\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.4639\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.7330\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.8018\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.4246\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.3136\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.6089\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.6930\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.2922\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.2274\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.4424\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.5258\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.1517\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.1411\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.2760\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.3585\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.1076\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.1067\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.2155\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.2841\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.0777\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0719\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1550\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.2097\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.0655\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0643\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1441\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1955\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.0635\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0567\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1332\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1812\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.0600\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.0535\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.1239\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.1711\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\u003eTo further illustrate this behavior, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e visualizes the sensitivity of recommendation performance to the fusion weight \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\alpha\\:\\)\u003c/span\u003e\u003c/span\u003e on Precision@10, Recall@10, NDCG@5, and NDCG@10.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe hybrid approach outperforms both standalone models and demonstrates competitive performance compared with recent state-of-the-art methods such as LightGCN [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e] and GenRec [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e"},{"header":"6. Discussion","content":"\u003cp\u003eThe experimental findings show that, in the presence of adequate user\u0026ndash;item interaction data, collaborative filtering continues to be the main factor influencing recommendation accuracy. The SVD model's impressive performance attests to the ongoing efficacy of matrix factorization methods in identifying latent item relationships and user preferences. These results are in line with earlier studies that demonstrate the resilience of collaborative filtering in settings with a lot of data.\u003c/p\u003e \u003cp\u003eNonetheless, quantifiable gains in recommendation quality are achieved through the incorporation of semantic content representations produced by BERT, especially in ranking-based metrics. The hybrid model gains from the complementary nature of the two methods: BERT adds contextual knowledge of movie metadata, whereas SVD records behavioral patterns gleaned from past interactions. Because of this combination, the system can suggest semantically related items that might not be highly connected based just on user ratings.\u003c/p\u003e \u003cp\u003eThe findings also show that scenarios with cold-start conditions or sparse data benefit most from the contribution of semantic features. Semantic similarity derived from textual features helps maintain recommendation relevance when rating information is scarce. On the other hand, collaborative signals take center stage as interaction data volume rises, and content-based features' relative contribution becomes secondary rather than essential.\u003c/p\u003e \u003cp\u003eThe hybrid model's stability across a range of fusion weights is another noteworthy finding. In a variety of configurations, the hybrid framework consistently outperforms the standalone models, even though performance peaks at a particular balance between collaborative and semantic signals. This implies that integrating heterogeneous recommendation signals enhances resilience and lessens susceptibility to changes in the distribution of data.\u003c/p\u003e \u003cp\u003eThe suggested system offers a useful trade-off between computational complexity and performance when compared to more recent deep learning-based recommendation models, such as large language model-driven recommenders and graph neural network approaches. Even though state-of-the-art models frequently attain high accuracy, they usually call for sophisticated architectures and substantial computational resources. On the other hand, the suggested hybrid framework is still computationally efficient, interpretable, and easier to deploy in practical systems.\u003c/p\u003e \u003cp\u003eFrom the standpoint of the application, dynamic web environments with regularly added content are best suited for the hybrid design. Even with limited historical interaction data, the system can produce meaningful recommendations thanks to its ability to incorporate semantic information. Because of this, the strategy is applicable to e-commerce platforms, streaming platforms, and new digital services that are constantly growing their content libraries.\u003c/p\u003e \u003cp\u003eNotwithstanding these benefits, the results also draw attention to a number of factors. The prevalence of collaborative filtering suggests that recommendation performance is still heavily influenced by the density and quality of rating data. Furthermore, the semantic component's representational richness may be constrained by its reliance on comparatively simple textual features. Richer content sources might be incorporated into future improvements to fortify the semantic modeling element even more.\u003c/p\u003e \u003cp\u003eAll things considered, the findings demonstrate that hybrid recommendation techniques offer a sensible and practical path forward for contemporary recommender systems. The proposed method maintains computational efficiency and practical applicability while improving personalization, ranking quality, and resilience to sparse-data conditions through the integration of behavioral and semantic signals.\u003c/p\u003e"},{"header":"7. Conclusion, Limitations, and Future Work","content":"\u003cp\u003eThis study introduced a hybrid movie recommendation system that combines collaborative filtering based on SVD with semantic text embeddings based on BERT. The suggested method produces reliable and accurate recommendations by fusing latent user preference modeling with deep content understanding. Overall, the study demonstrates that lightweight hybrid recommender systems can offer a practical balance between accuracy, interpretability, and deployment feasibility in real-world environments. The proposed approach highlights the practical potential of lightweight hybrid recommender systems for real-world deployment in dynamic content platforms.\u003c/p\u003e \u003cp\u003e \u003cb\u003eLimitations\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe system depends on a small amount of textual data, namely the genres and titles of movies. The richness of semantic representations is thus limited. The fusion strategy does not adjust to specific users; instead, it employs a fixed global weight. Additionally, sequential behavior and temporal dynamics are not taken into account.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFuture Work\u003c/b\u003e \u003c/p\u003e \u003cp\u003eRicher textual sources, like user reviews and plot summaries, will be investigated in subsequent work. Adaptive fusion strategies, temporal modeling, graph-based collaborative filtering techniques, and fine-tuning BERT on domain-specific data are possible additional enhancements.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eDevlin J, Chang M-W, Lee K, Toutanova K (2019) BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding, Proc. NAACL-HLT\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoren Y, Bell R, Volinsky C (2009) Matrix Factorization Techniques for Recommender Systems. Computer 42(8):30\u0026ndash;37\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe X, Deng K, Wang X, Li Y, Zhang Y, Wang M (2020) LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation, Proc. SIGIR\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJing Y, Wang Y, Feng Y, Yu PS (2023) Contrastive Self-supervised Learning in Recommender Systems: A Survey. IEEE Trans Knowl Data Eng\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao WX, Mu S, Hou Y, Lin Z, Li Y, Wang X, Wen J-R Recommender Systems in the Era of Large Language Models, arXiv:2307.02046, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMcAuley J, Leskovec J (2013) Hidden Factors and Hidden Topics: Understanding Rating Dimensions with Review Text, Proc. ACM RecSys\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang J, Wang K (2018) Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding (Caser), Proc. WWW\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJi J, Li Z, Xu S, Hua W, Ge Y, Tan J, Zhang Y GenRec: Large Language Model for Generative Recommendation, arXiv:2307.00457, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou X, Ma Y, Wang Y, Liu W, Li H (2023) A Comprehensive Survey of Recommender Systems Based on Deep Learning. ACM Comput Surv\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGheewala A, Shah K, Patel D (2025) In-depth Survey of Deep Learning Techniques in Recommender Systems. Inf Syst Front\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Damascus University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"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":"Recommender Systems, Hybrid Recommendation, BERT, SVD, Collaborative Filtering, Deep Learning, Text Analysis","lastPublishedDoi":"10.21203/rs.3.rs-8829543/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8829543/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOn modern digital platforms, recommender systems are crucial in assisting users in finding pertinent content. Users frequently find it difficult to find products that fit their individual interests as the amount of information available keeps expanding. While content-based methods rely on item information like titles and genres, traditional collaborative filtering techniques learn from past user ratings. Despite their widespread use, both strategies have definite drawbacks. While content-based systems frequently struggle to fully comprehend the meaning of text, collaborative filtering suffers when data is sparse or when new users or items appear.\u003c/p\u003e \u003cp\u003eThe hybrid movie recommendation system proposed in this paper combines the advantages of both methods. A pre-trained DistilBERT model that can extract contextual meaning from text is used to convert movie titles and genres into semantic vectors. Simultaneously, Singular Value Decomposition (SVD) is used to model user preferences based on rating data from the past. Using a weighted linear fusion strategy, the outputs of the two models are combined to determine the final recommendation score.\u003c/p\u003e \u003cp\u003eStandard Top-N evaluation metrics were used to conduct the experiments on the MovieLens dataset. The hybrid model consistently performs better than standalone BERT-based and SVD-based systems, according to the results. The model achieves an NDCG@10 value near 0.91 at the ideal fusion weight, indicating robust and realistic ranking performance. These results show that, particularly in cold-start and sparse-data scenarios, combining semantic content understanding with collaborative signals results in more precise, reliable, and customized movie recommendations.\u003c/p\u003e \u003cp\u003eThis study presents a lightweight hybrid recommendation framework that balances semantic understanding and collaborative learning.\u003c/p\u003e","manuscriptTitle":"A Hybrid Movie Recommendation System Using BERT-Based Semantic Embeddings and SVD Collaborative Filtering","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-10 09:20:25","doi":"10.21203/rs.3.rs-8829543/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":"daf75c8e-55fc-421f-a79e-bf094c863f8e","owner":[],"postedDate":"February 10th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":62576314,"name":"Information Retrieval and Management"}],"tags":[],"updatedAt":"2026-02-10T09:20:26+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-10 09:20:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8829543","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8829543","identity":"rs-8829543","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}