Universum Discriminant Canonical Correlation Analysis

Over the past decades, extensive studies on multi-view learning and Universum learning have been witnessed in pattern recognition and machine learning. Incorporating multi-view learning and Universum learning together, we propose a novel supervised dimensionality reduction method for multi-view data accompanied by Universum data, termed Universum discriminant canonical correlation analysis (UDCCA). UDCCA exploits inter-view information by means of the within-class correlation as well as the within-Universum correlation between different views, and at the same time, utilizes intra-view discriminant information captured from the target samples and Universum data of each view. In the low-dimensional discriminant space, the within-class correlation of the target samples is maximized, and the correlation of the Universum data is minimized, and simultaneously the scatter among target samples and Universum data is also maximized. Experimental results on real-world multi-view datasets show its effectiveness compared to other related state-of-art dimensionality reduction methods.