Self-paced learning-based multi-graphs semi-supervised learning

Graph-based semi-supervised learning has received considerable attention in machine learning community. The performance of existing methods highly depends on the input weight graph, which is related to the choice of hyper-parameters, such as the number of neighbors and the weight function. Multiple weights graphs can be easily built by selecting different hyper-parameters. In addition, there are multiple weights graphs representations for the same data set in many real-world applications. A key challenge is how to obtain better performance for graph-based semi-supervised learning by introducing multiple graphs. In this paper, we focus on multi-modal semi-supervised learning. A novel graph-based multi-modal semi-supervised learning framework, self-paced multi-modal label propagation learning (SPLP), is proposed. The main idea of SPLP is to take each type of weights graph as one modality and train “easy” modality first, then “complex” ones by self-paced learning. SPLP has the ability to select different weight coefficients in a purely self-paced way for multiple graphs and has good robustness against noisy or irrelevant graphs which cannot be rightly achieved by existing methods. We apply the proposed learning framework to solve the classification task on benchmark datasets including non-network datasets (MNIST, PIE, Sector and Reuters-21578) and network datasets (Cora and Citeseer). Experimental results show that our method achieves more than 2% improvement with regard to the several graph based state-of-the-art semi-supervised learning methods.