Coupling privileged kernel method for multi-view learning

Abstract

Multi-view learning concentrates on fully using the data collected from diverse domains or obtained from various feature extractors to learn effectively. The consensus and complementarity principles provide significant guidance in multi-view modeling. Many support vector machine (SVM)-based multi-view learning models have been proposed, which mainly follow the consensus principle through exploiting the label correlation with regularization terms. In this paper, we propose a simple yet effective coupling privileged kernel method for multi-view learning, termed as MCPK. The coupling term included in the primal objective allows the combination of the errors from all views to be minimized, which guarantees the consensus principle. Similar to our previous work PSVM-2V, MCPK realizes the complementarity principle by applying the learning using privileged information (LUPI) paradigm. The proposed model not only fully integrates the information from all views in the learning process, but also maintains the characteristic of different views to some extent. We employ the standard quadratic programming solver to solve MCPK. Further more, we theoretically analyze the performance of MCPK from the viewpoints of the generalization capability and the PSVM-2V and SVM-2K models. Experimental results demonstrate that MCPK compares more favorably than other state-of-the-art multi-view algorithms in terms of classification accuracy and efficiency.