A novel technique of using coupled matrix and greedy coordinate descent for multi-view data representation

The challenge of clustering multi-view data is to learn all latent features embedded in multiple views accurately and efficiently. Existing Non-negative matrix factorization based multi-view methods learn the latent features embedded in each view independently before building the consensus matrix. Hence, they become computationally expensive and suffer from poor accuracy. We propose to formulate and solve the multi-view data representation by using Coupled Matrix Factorization (CMF) where the latent structure of data will be learned directly from multiple views. The similarity information of data samples, computed from all views, is included into the CMF process leading to a unified framework that is able to exploit all available information and return an accurate and meaningful clustering solution. We present a variable selection based Greedy Coordinate Descent algorithm to solve the formulated CMF to improve the computational efficiency. Experiments with several datasets and several state-of-the-art benchmarks show the effectiveness of the proposed model