Multi-Granularity History Graph Network for temporal knowledge graph reasoning

Abstract

Reasoning on knowledge graphs (KGs) can be categorized into two main categories: predicting missing facts and predicting unknown facts in the future. However, when it comes to future prediction, it becomes crucial to incorporate temporal information and add timestamps to KGs, thereby forming temporal knowledge graphs (TKGs). The key aspect of reasoning lies in treating a TKG as a sequence of static KGs in order to effectively grasp temporal information. Additionally, it is equally important to consider the evolution of facts from various perspectives. Existing models tend to replicate the original time granularity of data while modeling TKGs, often disregarding the impact of the minimum time period in the evolution process. Furthermore, historical information is typically perceived as a single sequence of facts, with a lack of diversity in strategies (e.g., modeling sequences with varying granularities or lengths) to capture complex temporal dynamics. This unified approach may lead to the loss of critical information during the modeling process. However, the process of historical evolution often exhibits complex periodic transformation characteristics, and associated events do not necessarily follow a fixed time period. Therefore, a single granularity is insufficient to model periodic events with dynamic changes in history. Consequently, we propose the Multi-Granularity History Graph Network (MGHGN), an innovative model for TKG reasoning. MGHGN dynamically models various event evolution periods by constructing representations with multiple time granularities, and integrates various modeling methods to reason the potential facts in the future. Our model adeptly captures valuable insights from the history of multi-granularity and employs diverse approaches to model historical information. The experimental results on six benchmark datasets demonstrate that the MGHGN outperforms state-of-the-art methods.