A0704
Title: Canonical correlation analysis for multimodal labeled data
Authors: Mitsuhiro Hashiguchi - Doshisha University (Japan) [presenting]
Masaaki Okabe - Doshisha University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: Canonical correlation analysis (CCA) is a method of dimensionality reduction for two multivariate data. It is effective in reflecting the characteristics of each class in a low-dimensional space when one of the data represents a class label. On the other hand, CCA cannot reflect the structures of multimodal data in a low-dimensional space. The multimodal data has multiple unknown cluster structures within a class. This problem arises because CCA cannot take the local structure of the data into account. The entropic-regularized Wasserstein distance can capture the local structure. This distance corresponds to the squared Euclidean distance with a weighting. Therefore, to reflect the multiple structures within a class in multimodal data in a low-dimensional space, we incorporate the Wasserstein distance into CCA. Simulations and real data examples demonstrate the effectiveness of this method for data with multimodality in classes. Results suggest that the proposed method can reflect characteristics of multiclass data with multimodal classes in a low-dimensional space.
