Modeling Time Decay Effect in Temporal Knowledge Graphs via Multivariate Hawkes Process

• Published in: Proceedings of ... International Joint Conference on Neural Networks pp. 1 - 8
• Main Authors: Li, Qianyu, Chen, Jiebin, Tang, Xiaoli, Yu, Han, Song, Hengjie
• Format: Conference Proceeding
• Language: English
• Published: IEEE 30.06.2024
• Subjects: Attention mechanism; Attention mechanisms; Benchmark testing; Bridges; Graph embedding; Hawkes process; Knowledge engineering; Knowledge graph; Knowledge graphs; Merging; Neural networks
• ISSN: 2161-4407
• DOI: 10.1109/IJCNN60899.2024.10649972

Knowledge Graph Embedding (KGE) is attracting growing research interest because it offers great flexibility for the manipulation and application of Knowledge Graphs (KGs). However, most existing works focus on static KGE, while temporal KGE is still in its infancy. Recent temporal KGE methods attempt to obtain the long-term dependency of facts in consecutive timestamps by merging historical fact information. However, they ignore the different impacts of historical facts on the current facts due to the time decay effect and heterogeneity of historical facts. To bridge this gap, we formalize the concept of fact formation sequence to describe the evolution of an entity and propose the Modeling Time Decay Effect in Temporal Knowledge Graphs via Multivariate Hawkes Process method (TimeDE). TimeDE uses the Hawkes process to model the time decay effect of historical facts. It also incorporates the attention mechanism based on the score function of static KGE methods to better capture the impacts of heterogeneous historical facts on the current facts. Extensive experiments on the five commonly-used benchmark datasets demonstrate that TimeDE achieves significant improvements in terms of both Mean Reciprocal Rank and Hits@K compared to state-of-the-art methods.