SHAP and MIC based Hybrid Feature Selection Method and Classification of Gene Expression Data

In the realm of machine learning, feature selection plays a pivotal role in enhancing model interpretability and performance. This research paper introduces an innovative feature selection algorithm leveraging Shapley values and Maximum Information Coefficient (MIC) for an iterative and dynamic approach. The algorithm aims to iteratively select features based on their Shapley values and update the importance of the remaining features considering the mutual information captured by MIC. The initial phase employs the Minimum Redundancy Maximum Relevance (mRMR) algorithm to select a set of features. Subsequently, Shapley values are calculated to gauge the individual contributions of each feature to model performance. The Dynamic Feature Index (DFI) is then introduced, combining Shapley values and MIC to iteratively select features with the highest DFI. Until a predetermined number of characteristics are chosen, the procedure is repeated, optimizing the feature subset for interpretability and model accuracy. Experimental validation on diverse datasets demonstrates the efficacy of the proposed algorithm in improving model interpretability while maintaining competitive predictive accuracy. The algorithm presents a novel and flexible framework for feature selection, offering a balance between interpretability and performance in machine learning applications.