Fast multi-view twin hypersphere support vector machine with consensus and complementary principles

摘要

Multi-view learning (MVL) is an emerging machine learning approach which concentrates on the problems of learning from data represented by multiple distinct feature sets. Various support vector machine (SVM) based MVL algorithms have been proposed and showed excellent performance in learning tasks. However, multi-view SVMs face two major challenges, one being their immense computational complexity and the other being difficult to achieve views’ consistency and views’ complementary simultaneously. In this paper, we firstly propose a fast training MVL algorithm based on twin hypersphere support vector machine (THSVM), termed as multi-view THSVM (MvTHSVM). MvTHSVM trains two independent hyperspheres on two views under a pair of co-regularization constraints, which maximizes the consensus of two distinct views. We further simplify the dual problems of MvTHSVM into a pair of smaller-scaled ones based on the idea of matrix reduction, and adopt the alternating direction multipliers method (ADMM) to boost its solving speed. Secondly, we extend MvTHSVM to a general MVL framework which realizes consensus principle and complementary principle at the same time, termed as MvTHSVM-2C. The integration of complementary principle renders MvTHSVM-2C better classification performance. Benchmark results conducted on 55 binary data sets confirm the generalization capacity and training efficiency of the proposed algorithms.