Evolutionary Transfer Optimization Assisted by Unselected Features for Multiobjective Feature Selection

Feature selection plays a crucial role in classification tasks, particularly in high-dimensional datasets where identifying relevant features while minimizing redundancy is challenging. Traditional multiobjective feature selection (MOFS) methods face challenges due to random initialization and focusing solely on the classification accuracy of selected features. This often results in the selection of many irrelevant features, while important features may be overlooked because they remain hidden among the unselected features, which are typically ignored in later optimization stages. To address these issues, this paper proposes a specific preference task-assisted multiobjective feature selection (SPTMFS) method, an advanced evolutionary transfer optimization framework assisted by unselected features to improve the performance of optimizing the main MOFS task. SPTMFS incorporates two complementary auxiliary MOFS tasks: one focusing on minimizing classification accuracy on unselected features, and the other on optimizing the size of the feature subset. These tasks integrate information from both selected and unselected features. A knowledge transfer mechanism leverages task-specific preferences to enhance the main task's ability to identify improved Pareto-optimal feature subsets. Personalized environmental selection ensures solutions align with each task's objectives, enhancing overall population performance. Experimental results across 16 datasets demonstrate SPTMFS's superiority over state-of-the-art MOFS methods. Rigorous statistical analyses using Friedman and Wilcoxon tests validate SPTMFS's robustness and competitive advantage. Comparative evaluations with single-objective multitasking methods underscore SPTMFS's effectiveness in addressing diverse feature selection challenges across datasets.