Multi-view Subspace Learning for Large-Scale Multi-Modal Data Analysis

Abstract

Dimensionality reduction methods play a big role within the modern machine learning techniques, and subspace learning is one of the common approaches to it. Although various methods have been proposed over the past years, many of them suffer from limitations related to the unimodality assumptions on the data and low speed in the cases of high-dimensional data (in linear formulations) or large datasets (in kernel-based formulations). In this work, several methods for overcoming these limitations are proposed. In this thesis, the problem of the large-scale multi-modal data analysis for single- and multi-view data is discussed, and several extensions for Subclass Discriminant Analysis (SDA) are proposed. First, a Spectral Regression Subclass Discriminant Analysis method relying on the Graph Embedding-based formulation of SDA is proposed as a way to reduce the training time, and it is shown how the solution can be obtained efficiently, therefore reducing the computational requirements. Secondly, a novel multi-view formulation for Subclass Discriminant Analysis is proposed, allowing to extend it to data coming from multiple views. Besides, a speed-up approach for the multi-view formulation that allows reducing the computational requirements of the method is proposed. Linear and nonlinear kernel-based formulations are proposed for all the extensions. Experiments are performed on nine single-view and nine multi-view datasets and the accuracy and speed of the proposed extensions are evaluated. Experimentally it is shown that the proposed approaches result in a significant reduction of the training time while providing competitive performance, as compared to other subspace-learning based methods.