Fast and direct co-clustering via adaptive anchor graph learning

Spectral clustering faces three main challenges: high computational costs arising from regular graph construction and eigenvalue decomposition, important information loss and solution deviation due to the ‘relaxation and re-discretization’ strategy, and the inflexibility of fixed graphs that fail to adapt to unreliable similarities. To address these issues, Fast and Direct Co-Clustering via adaptive anchor graph learning (FDC²) is proposed. This method dynamically learns the anchor graph and efficiently solves the co-clustering problem. FDC² iteratively refines the anchor graph using raw data, anchors, and an evolving indicator matrix, thus reducing computational complexity while enabling the modification of unreliable similarities. The learned anchor similarity matrix is used as the weight matrix in a bipartite graph, facilitating the simultaneous co-clustering of raw data and anchors. To directly and efficiently solve the discrete indicator matrix, a coordinate descent method is employed. Extensive experimental results demonstrate that FDC² significantly reduces computational time and delivers superior clustering results.