Efficient Multi-View Clustering via Greedy Automatic View Selection and Diverse Feature Integration

Multi-view clustering leverages complementary information from multiple feature representations, yet its success relies on selecting optimal feature combinations and clustering algorithms. We propose a Greedy Automatic View Selection (GAVS) algorithm to identify the most informative subset of feature views that maximize clustering performance. GAVS iteratively adds feature views based on their contribution to clustering quality, measured by normalized mutual information (NMI). We evaluate GAVS on Coil20, UCI Digits, Movies, and Caltech 7 datasets using Spectral, Agglomerative, and Affinity Propagation clustering with diverse features (GIST, LBP, HOG, CENTRIST). Results show optimal combinations vary across datasets, with GAVS achieving peak NMIs of 1.000 (Coil20), 0.9351 (UCI Digits), 0.6937 (Movies), and 0.9806 (Caltech 7). This adaptive strategy offers practical guidance for improving clustering accuracy in real-world applications.