Constrained least squares regression for semi-supervised learning

Abstract

The core tasks of graph based semi-supervised learning (GSSL) are constructing a proper graph and selecting suitable supervisory information. The ideal graph is able to outline the intrinsic data structure, and the ideal supervisory information could represent the whole data. In this paper, we propose a new graph learning method, called constrained least squares regression (CLSR), which integrates the supervisory information into graph learning process. To learn a more adaptive graph, regression coefficients and neighbor relations are combined in CLSR to capture the global and local data structures respectively. Moreover, as byproduct of CLSR, a new strategy is presented to select the high-quality data points as labeled samples, which is practical in real applications. Experimental results on different real world datasets demonstrate the effectiveness of CLSR and the sample selection strategy. © 2014 Springer International Publishing.