Semi-supervised Probability Propagation on Instance-Attribute Graphs

Abstract

Graph-based methods have become one of the most active research areas of semi-supervised learning (SSL). Typical SSL graphs use instances as nodes and assign weights that reflect the similarity of instances. In this paper, we propose a novel type of graph, which we call instance-attribute graph. On the instance-attribute graph, we introduce another type of node to represent attributes, and we use edges to represent certain attribute values. The instance-attribute graph thus moreexplicitly expresses the relationship between instances and attributes. Typical SSL graph-based methods are nonparametric, discriminative, and transductive in nature. Using the instance-attribute graph, we propose a nonparametric and generative method, called probability propagation, where two kinds of messages are defined in terms of corresponding probabilities. The messages are sent and transformed on the graph until the whole graph become smooth. Since a labeling function can be returned, the probability propagation method not only is able to handle the cases of transductive learning, but also can be used to deal with the cases of inductive learning. From the experimental results, the probability propagation method based on the instance-attribute graph outperforms the other two popular SSL graph-based methods, Label Propagation (LP) and Learning with Local and Global Consistency (LLGC).