Semi-supervised kernel matrix learning using adaptive constraint-based seed propagation

Abstract

In this paper, we propose semi-supervised kernel matrix learning (SS-KML) using adaptive constraint-based seed propagation (ACSP). Conventional SS-KML methods such as pairwise constraint propagation (PCP) and kernel propagation (KP) have achieved outstanding performance in data classification. However, they are likely to distort the global data structure because of using hard constraints in their semi-definite problems (SDPs) for constraint propagation. Moreover, given a large number of pairwise constraints and a large amount of samples, they tend to be incredibly complex, thus being hard to be applied to real-life complex problems such as internet-scale image categorization. To address this problem, we utilize adaptive constraints to effectively maintain the inherent coherence of samples and successfully propagate constraint information into all samples. Moreover, we adopt seed propagation to remarkably reduce the computational complexity of SS-KML. Experimental results demonstrate that ACSP achieves a significant improvement in performance over PCP and KP in terms of both effectiveness and efficiency.