Dimensionality reduction via representation and affinity learning

Abstract

Because graphs are crucial in modeling the data distribution or structure, graph-based dimensionality reduction methods have been widely employed in many machine learning and pattern recognition applications. However, in most existing graph-based dimensionality reduction techniques, the graph is usually specified empirically or learned from raw data, which may deteriorate their performances. In this paper, we propose a novel algorithm named Dimensionality Reduction via Representation and Affinity Learning (DRRAL), which adaptively constructs graph and learns the projection matrix for dimensionality reduction. Our algorithm takes noise and different local structure of each sample into account. Moreover, an efficient optimization strategy based on the iterative updating Augmented Lagrange Multiplier (ALM) and eigenvectors is developed to solve the proposed DRRAL. Extensive experimental results on five databases are carried out to verify the feasibility and effectiveness of proposed approach.