A novel multi-view classifier based on Nystrِm approximation

The existing multi-view learning (MVL) is learning from patterns with multiple information sources and has been proven its superior generalization to the conventional single-view learning (SVL). However, in most real-world cases, researchers just have single source patterns available in which the existing MVL is uneasily directly applied. The purpose of this paper is to solve this problem and develop a novel kernel-based MVL technique for single source patterns. In practice, we first generate different Nystrِm approximation matrices Kps for the gram matrix G of the given single source patterns. Then, we regard the learning on each generated Nystrِm approximation matrix Kp as one view. Finally, different views on Kps are synthesized into a novel multi-view classifier. In doing so, the proposed algorithm as a MVL machine can directly work on single source patterns and simultaneously achieve: (1) low-cost learning; (2) effectiveness; (3) the same Rademacher complexity as the single-view KMHKS; (4) ease of extension to any other kernel-based learning algorithms.