ML-Based Feature Selection Technique for Imbalanced Data Streams

In the data stream classification domain, efficient feature selection (FS) plays an ornamental role in improving the overall performance of machine learning (ML) frameworks. Also, the imbalanced data significantly affects the classifier's performance. This study proposes an efficient ML framework with a hybrid FS based on the filter and wrapper approach. Another primary goal of this rigorous study is to show the efficacy of ML using the proposed FS in the classification of imbalanced data. First, feature subsets are rapidly selected using statistical (filter) FS methods. Modified correlation-based feature subset selection with greedy stepwise search and chi-square method has been implemented as filter-based methods. Then the selected feature subset is fine-tuned using the wrapper approach recursive feature elimination (RFE) and an ensemble of classifiers. The framework is built and tested on imbalanced datasets from the healthcare domain. The proposed framework has been compared to three datasets used in disease prediction. The framework effectively achieves better diagnostic and prediction performance with the optimal number of features without computational overheads.