Graph Construction Based on Nonnegative Sparse Representation for Semi-supervised Learning

Abstract

It is well known that the graph construction is the key part of graph-based semi-supervised learning algorithms, and the performance of algorithms relies heavily on the graph weight matrix given by graph construction process. In this paper, we propose two graph construction models based on nonnegative sparse representation. These two models accommodate small possible noise, and moreover, their solutions are sparse and nonnegative which can be used as the graph weights directly. Weights generated in such a way can reflect the point neighborhood structure well, thereby providing favorable similarity measures for the sample pairs. Numerical experiments on several UCI and face datasets indicate that in most cases the results yielded by the proposed algorithms are comparable even superior to the best ones yielded by the algorithms based on traditional graph construction methods and L1 graph.