Frequency-Based Multi-Objective Feature Selection to Enhance the Generalization of Evolutionary Algorithms

Feature selection is a critical preprocessing task in machine learning, particularly with high-dimensional datasets and decision-making while handling big data which presents significant challenges. This paper introduces an innovative approach for multi-objective feature selection, aiming to minimize the number of features and classification error simultaneously. The method mitigates the generalization issues commonly faced when relying on the results of a single run of evolutionary algorithms. Our approach leverages the frequency of each feature across multiple runs of the optimization algorithm, applied to different portions of the data, as a key metric for ranking the features. This can reduce the risk of overfitting and enhances generalization by capturing more reliable features through repeated runs and different data subsets. To enhance the robustness of the selection, we incorporate the correlation between features and labels to determine the final feature set. To evaluate the proposed method, we selected fourteen datasets with varying numbers of features and instances. Experimental results demonstrate that this post-optimization processing technique significantly enhances generalization and consistently delivers superior performance across various datasets compared to the raw optimization results.