Intra-View and Inter-View Supervised Correlation Analysis for Multi-View Feature Learning

Multi-view feature learning is an attractive research topic with great practical success. Canonical correlation analysis (CCA) has become an important technique in multi-view learning, since it can fully utilize the inter-view correlation. In this paper, we mainly study the CCA based multi-view supervised feature learning technique where the labels of training samples are known. Several supervised CCA based multi-view methods have been presented, which focus on investigating the supervised correlation across different views. However, they take no account of the intra-view correlation between samples. Researchers have also introduced the discriminant analysis technique into multiview feature learning, such as multi-view discriminant analysis (MvDA). But they ignore the canonical correlation within each view and between all views. In this paper, we propose a novel multi-view feature learning approach based on intra-view and inter-view supervised correlation analysis (ISCA), which can explore the useful correlation information of samples within each view and between all views. The objective function of ISCA is designed to simultaneously extract the discriminatingly correlated features from both inter-view and intra-view. It can obtain an analytical solution without iterative calculation. And we provide a kernelized extension of ISCA to tackle the linearly inseparable problem in the original feature space. Four widely-used datasets are employed as test data. Experimental results demonstrate that our proposed approaches outperform several representative multi-view supervised feature learning methods. Introduction In real world applications, datasets are usually described with different views or representations. Multi-view feature learning refers to learning with multiple feature sets that Copyright © 2014, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. reflect different characteristics or views of data, which is an vital research direction (Guo, 2013; Xu, Tao and Xu 2013; Wang, Nie and Huang 2013; Memisevic 2012; Wang, Li, and Ogihara 2012). Co-training and canonical correlation analysis are two representative and effective techniques in multi-view learning (Sun and Chao 2013). Co-training based methods (Kumar and Daume 2011a; Kumar and Daume 2011) are usually used for semisupervised classification that combines both labeled and unlabeled data under multi-view setting. Canonical correlation analysis (CCA, Hardoon, Szedmak, and ShaweTaylor 2004) has become an important technique in multiview learning, since it can fully utilize the inter-view correlation. Multi-view CCA (MCCA, Rupnik and ShaweTaylor 2010) is an unsupervised method. In this paper, we mainly study the CCA based multi-view supervised feature learning technique where the labels of training samples are known. Recently, several supervised CCA based multi-view feature learning methods have been presented, such as multiple discriminant CCA (MDCCA, Gao et al. 2012), multiple principal angles (MPA, Su et al. 2012). These methods focus on investigating supervised correlation across different views. To deal with the linearly inseparable problem, researchers extend CCA to be kernel CCA (Sun et al. 2007; Leurgans, Moyeed, and Silverman 1993; Lai and Fyfe 2000; Bach and Jordan 2002). All methods mentioned above only reveal the linear or nonlinear correlation relationship between features of multiple views. However, they take no account of the intra-view correlation between samples, which is also an important part when exploiting supervised correlation among the samples. Therefore, in this paper, we need to simultaneously extract the discriminatingly correlated features from both inter-view and intra-view. The Intra-View and Inter-View Supervised Correlation Analysis for Multi-View Feature Learning Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence