Learning to Efficiently Propagate for Reasoning on Knowledge Graphs

Path-based methods are more appealing solutions than embedding methods for knowledge graph reasoning, due to their interpretability and generalization ability to unseen graphs. However, path-based methods usually suffer from the problem of scalability, as the time complexity grows exponentially w.r.t. the length of paths. While recent methods compute reasoning paths with the Bellman-Ford algorithm in polynomial time, the time and memory cost remains very high, as they need to propagate through all the nodes and edges in the graph. In this paper, we propose A*Net, an efficient model for path-based reasoning on knowledge graphs. Inspired by the classical A* algorithm for shortest path problems, our A*Net prioritizes important nodes and edges at each propagation step, to reduce the time and memory footprint. Unlike the classical A* algorithm that uses a heuristic function, we propose to learn the priority function for each node to capture the complex semantics in knowledge graphs. The priority function and the propagation steps are jointly optimized through backpropagation. Experiments on both transductive and inductive knowledge graph reasoning benchmarks show that A*Net achieves competitive performance with existing state-of-the-art path-based methods, and meanwhile reduces the number of messages, the time and the memory cost up to 7.2$\times$, 3.4$\times$ and 4.9$\times$ respectively.