Multi-granularity Histories Merging Network for Temporal Knowledge Graph Reasoning

Temporal Knowledge Graphs (TKGs) are essential for modeling complex event relations between real-world entities. TKG reasoning aims to predict facts at future timestamps. The key to TKG reasoning is capturing historical information as comprehensively as possible for prediction. Different types of historical information provided in a multi-granular form (such as local history and global history) have different effects on forecasting and are indispensable without each other. However, many current methods fail to fully capture multi-granularity historical facts or consider complex associations between entity-relation pairs during the decoding. A novel Multi-Granularity Histories Merging (MGHM) network for TKG reasoning is proposed to address this issue. This method involves a Local History Encoder that captures temporal dependencies between adjacent temporal subgraphs. A History-Gate (HisGate) mechanism is also designed for effectively merging histories of varying granularity. Further, an attention-based decoder is incorporated to excavate the complex interplay of entity-relation information. Extensive experimental analysis on benchmark datasets highlights the enhanced performance of the proposed model and validates the effectiveness of multi-granularity history modeling in TKG reasoning.