Towards First-Order Random Walk Inference

Path Ranking Algorithm (PRA) addresses classification and retrieval tasks using learned combinations of labeled paths through a graph. Unlike most Statistical Relational Learning (SRL) methods, PRA scales to large data sets but uses a limited set of paths in its models—ones that correspond to short first order rules with no constants. We consider extending PRA in two ways—learning paths that correspond to first order rules with constants, and incorporating long paths. Backward random walks are applied to find candidate constant paths, and long paths are efficiently constructed using bidirectional random walks from both query and target nodes. We apply these techniques to the tasks of knowledge base inference and coordinate term extraction. Empirical results show that constant paths provide informative class priors for relational classification, significantly improving performance on both tasks. Bidirectional search significantly reduces path finding time, and improves model quality for coordinate term extraction.