Efficient Feature Selection Algorithm Based on Counterfactuals
Diwen Liu, Xiaodong Yue · 2024
Feature selection is crucial for model construction, involving the identification and selection of the most predictive subset of features from the original set. While information theory-based feature selection algorithms are broadly utilized, they primarily assess correlations and fail to uncover causality, which is essential for pinpointing key decision-influencing features. Additionally, the computational efficiency of these algorithms is often challenged by vast, high-dimensional datasets prevalent today. This chapter introduces a counterfactual-based feature selection algorithm that incorporates counterfactual reasoning to analyze and evaluate features. This innovative approach not only integrates causal analysis into the computation process but also enhances the selection effectiveness of the feature subset. The proposed algorithm includes submodular properties, allowing it to be optimized using greedy techniques, thereby significantly boosting computational efficiency. Experimental results confirm the efficacy of the algorithm, demonstrating its potential to improve feature selection outcomes in complex data environments.