.png){kind=icon}
.png){kind=icon}
Return
Learning entity-query dependency and query classification for temporal knowledge graph prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.128016.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Temporal knowledge graph prediction (TKGP), as a hotspot in temporal knowledge graph reasoning research, has been extensively explored in recent years. Most TKGP methods aim to predict events with recurrent, periodic, or sequential patterns, so it is challenging for them to predict new events. Some existing TKGP methods integrate information from historical repetitive events and underlying non-historical factors for new events prediction. However, the majority of new events can be successfully predicted based on the essential information contained in historical non-repetitive events and historical multi-hop events. Current research categorizes these two types of events into historical unrelated events in a coarse-grained manner, which limits the ability of model to predict new events. Therefore, we fine-grainedly partition historical events according to their association with queries and propose a novel TKGP model called CMHE-NET to better predict new events in this paper. To distinguish the matching degree between entities and queries at the entity level, we encode the historical repetitive dependency, historical non-repetitive dependency, historical multi-hop dependency, and non-historical dependency between queries and entities through multiple categories of historical events based on the copy mechanism to obtain the probability distribution of entities. Moreover, to focus on key entities at the query level under fine-grained historical event division, the query contrastive representation encoder based on supervised contrastive learning is combined with the query multi-class classifier with alpha-balanced focal loss to achieve multi-class historical contrastive learning. Experimental results demonstrate that CMHE-NET surpasses existing TKGP methods and achieves state-of-the-art performance on three event-based datasets (ICEWS14, ICEWS18, and ICEWS05-15) and two general knowledge datasets (WIKI and YAGO).
Keywords:
Copy mechanism
Entity-query dependency
Historical contrastive learning
Temporal knowledge graph prediction
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W