Multi-view Ensemble Feature Selection via SemiDefinite Programming

Abstract

Multi-view learning faces significant challenges in selecting discriminative features while managing redundancy and noise across heterogeneous data sources. To address these issues, this paper introduces Multi-view Ensemble Feature Selection (MEFS), a novel framework that systematically integrates view generation (VG) and view selection (VS) through a unified optimization paradigm. By reformulating feature selection as a MaxCut problem and leveraging SemiDefinite Programming (SDP) relaxation, MEFS dynamically balances the generalization capability of individual views with their pairwise diversity, eliminating the need for manual parameter tuning. A key innovation is the proposed pairwise diversity metric, which quantifies inter-view dissimilarity using between-class scatter matrices to ensure complementary feature subsets. Extensive experiments on ten benchmark datasets demonstrate that MEFS consistently outperforms state-of-the-art methods in accuracy, robustness, and computational efficiency. Ablation studies validate the synergistic effect of combining VG and VS modules.