GraphEBM: Energy-based Graph Construction for Semi-Supervised Learning

Zhijie Chen, Hongtai Cao, Kevin Chen-Chuan Chang

Abstract: With the rapid improvement of various techniques in graph-based semi-supervised learning, the call for higher-quality graphs becomes more intensive. However, such affinity graphs are not naturally existing in most semi-supervised learning tasks. In this paper, we propose a learning-based approach, GraphEBM, for the graph construction problem. GraphEBM is designed to address three main requirements in graph construction: 1) supporting dynamic update; 2) providing interpretable metrics; 3) tailoring to tasks. Specifically, in GraphEBM, we adopt a probabilistic view, Edge Probability Space, to model a graph construction process as constituted of events from the space. Our objective is thus to learn, by our Energy-Based Model (EBM), the latent sampling distribution. Experimental results show that our proposed GraphEBM outperforms the existing graph construction methods in improving the semi-supervised learning tasks on various datasets and it can learn global properties of a target graph only with direct local guidance.