Block Decomposition with Multi-granularity Embedding for Temporal Knowledge Graph Completion

Abstract

Temporal knowledge graph (TKG) completion is the mainstream method of inferring missing facts based on existing data in TKG. Majority of existing approaches to TKG focus on embedding the representation of facts from a single-faceted low-dimensional space, which cannot fully express the information of facts. Furthermore, most of them lack the comprehensive consideration of both temporal and non-temporal facts, resulting in the inability to handle the two types of facts simultaneously. Thus, we propose BDME, a novel Block Decomposition with Multi-granularity Embedding model for TKG completion. It adopts multivector factor matrices and core tensor em-bedding for fine-grained representation of facts based on the principle of block decomposition. Moreover, it captures interaction information between entities, relationships, and timestamps in multiple dimensions. By further constructing a temporal and static interaction model, BDME processes temporal and non-temporal facts in a unified manner. Besides, we propose two kinds of constraint schemes, which introduce time embedding angle and entity bias component to avoid the overfitting problem caused by a large number of parameters. Experiments demonstrate that BDME achieves sub-stantial performance against state-of-the-art methods on link prediction.