A Feature Evolution Aware Classification Framework for Streaming Data using Dynamic Autoencoder and Ensembled Learning

Recent advancements in the field of data mining and knowledge discovery have opened up a multitudeof research opportunities related to streaming data. One major challenge in handling streaming data isbuilding efficient machine-learning models that handle the dynamics of features and concepts. Focusing more on features, feature drift and evolving features are among the most unaddressed issues. It isdifficult to deal with the evolution of features in a stream; since most machine learning algorithms arerestricted to learning a fixed number of features. In the proposed work, a machine learning frameworkfor data streams with feature evolution is introduced. This methodology utilizes a dynamic autoen-coder to translate varying features into feature spaces with fixed dimensions. For classification, we constructed an ensemble model with Logical Regression, a Decision Tree, a Support Vector Machine,and K-Nearest Neighbor(KNN), which preserves past concepts. Based on experimental results, themodel was found to be promising for a variety of datasets, including Weather data (accuracy86%), Electricity data (94%), and Forest Cover types data (95%). By effectively combining deep-learning techniques with traditional approaches, many streaming data challenges can be addressed