Reinforcement Learning of Graph Matching

Siqi Tang, Congying Han, Tiande Guo, Mingqiang Li

Abstract: Graph matching not only has widespread application prospect and practical value, but also provides a broader perspective and new technologies for fundamental researches. The recently presented idea to learn heuristics or approximation algorithms often require significant specialized knowledge and trial-and-error. Therefore, we need better models and better ways to meet practical implementation. In this paper, a unique feature extraction combined with reinforcement learning procedure is proposed to tackle this challenge. There are three main contributions: 1. we introduce matrix symmetric compression to obtain global feature and Bi-directional Recurrent Neural Network (Bi-RNN) to extract local feature; 2. we transform graph matching to sequence-to-sequence problem based on the above feature; 3. we optimize parameters using Actor-Critic framework. Our experiments on synthetic and real databases reveal that reinforcement learning compares favorably to supervised case and traditional methods, both in terms of efficiency and quality.