Learning Logical Rules from Knowledge Graphs

Abstract

Expressing and extracting regularities in multi-relational data, where data points are interrelated and heterogeneous, requires well-designed knowledge representation. Knowledge Graphs (KGs), as a graph-based representation of multi-relational data, have seen a rapidly growing presence in industry and academia, where many real-world applications and academic research are either enabled or augmented through the incorporation of KGs. However, due to the way KGs are constructed, they are inherently noisy and incomplete. In this thesis, we focus on developing logic-based graph reasoning systems that utilize logical rules to infer missing facts for the completion of KGs. Unlike most rule learners that primarily mine abstract rules that contain no constants, we are particularly interested in learning instantiated rules that contain constants due to their ability to represent meaningful patterns and correlations that can not be expressed by abstract rules. The inclusion of instantiated rules often leads to exponential growth in the search space. Therefore, it is necessary to develop optimization strategies to balance between scalability and expressivity. To such an end, we propose GPFL, a probabilistic rule learning system optimized to mine instantiated rules through the implementation of a novel two-stage rule generation mechanism. Through experiments, we demonstrate that GPFL not only performs competitively on knowledge graph completion but is also much more efficient then existing methods at mining instantiated rules. With GPFL, we also reveal overfitting instantiated rules and provide detailed analyses about their impact on system performance. Then, we propose RHF, a generic framework for constructing rule hierarchies from a given set of rules. We demonstrate through experiments that with RHF and the hierarchical pruning techniques enabled by it, significant reductions in runtime and rule size are observed due to the pruning of unpromising rules. Eventually, to test the practicability of rule learning systems, we develop Ranta, a novel drug repurposing system that relies on logical rules as features to make interpretable inferences. Ranta outperforms existing methods by a large margin in predictive performance and can make reasonable repurposing suggestions with interpretable evidence.