An optimal feature selection method for text classification through redundancy and synergy analysis

Feature selection is an essential step in text classification tasks to enhance model performance, reduce computational complexity, and mitigate the risk of overfitting. Filter-based methods have gained popularity for their effectiveness and efficiency in selecting informative features. However, these methods often overlook feature correlations, resulting in the selection of redundant and irrelevant features while underestimating others. To address this limitation, this paper proposes FS-RSA (Feature Selection through Redundancy and Synergy Analysis), a novel method for text classification. FS-RSA aims to identify an optimal feature subset by considering feature interactions at a lower computational cost. It achieves this by evaluating features to optimize their synergy information and minimize redundancy within small subsets. The core principle of FS-RSA is that features offering similar classification information to the class variable are likely to be correlated and redundant, whereas features with high and low classification information can provide synergistic information. In the conducted experiments on five public datasets, FS-RSA was compared to five effective filter-based methods in text classification. It consistently achieved higher F1 scores with NB and SVM classifiers, highlighting its effectiveness in feature selection while significantly reducing dimensionality.