Multi-view feature extraction based on canonical correlation analysis

Abstract

Canonical Correlation Analysis (CCA) aims at identifying linear dependencies between two sets of variables. CCA has recently become popular in the eld of machine learning with the increase in the number of multi-view datasets, which consist of di erent representations coming from di erent sources or modalities. This thesis presents our e orts to improve the robustness and discriminative ability of CCA. CCA uses the views as complex labels to guide the search of maximally correlated projection vectors (covariates). Therefore, CCA can over t the training data. Although, ensemble approaches have been e ectively used to avoid such over ttings of classi cation and clustering techniques, an ensemble approach has not yet been formulated for CCA. In this thesis, we propose an ensemble method for obtaining a nal set of covariates by combining multiple sets of covariates extracted from subsamples. Experimental results on various datasets demonstrate the usefulness of ensemble CCA approach. The correlated features extracted by CCA may not be class-discriminative since it does not utilize the class labels in its implementation. This thesis introduces a method to explore correlated and also discriminative features. Our proposed method utilizes two (alternating) multi-layer perceptrons, each with a linear hidden layer, learning to predict both the class-labels and the outputs of each other. The experimental results show that the features found by the proposed method accomplish signi cantly higher classi- cation accuracies. Another contribution of this thesis is the use of CCA to improve a lter feature selection algorithm. We also present our works on ensemble classi cation and clustering for multi-view datasets.