.png){kind=icon}
.png){kind=icon}
Return
Temporal knowledge graph extrapolation with subgraph information bottleneck
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.126226.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the realm of temporal knowledge graph (TKG) extrapolation, subgraph-based reasoning methods offer clearer insights than those based on individual paths, effectively capturing local evidence. However, these subgraphbased methods face particular challenges, such as obtaining subgraph-level annotations and maintaining a balance between keeping enough information for accurate predictions and discarding unnecessary details. To overcome these obstacles, we introduce a novel reasoning method known as Subgraph Information Bottleneck based Reasoning (SIBR) for TKG extrapolation. Based on information bottleneck theory, SIBR aims to find subgraphs that are full of predictive value yet concise, leading to an efficient learning and inference process. SIBR is designed to capture the key temporal dynamics and evolution within TKGs by identifying subgraphs that are just large enough to represent the TKG's structure and temporal changes. It skillfully handles the computational complexities related to mutual information using variational techniques, offering a practical optimization strategy for TKG analysis. Our method's effectiveness is substantiated by comprehensive experiments on six public datasets, which demonstrate its superiority from multiple perspectives: it showcases improvement in predictive accuracy, robustness against data sparsity and data noise, and sensitivity to parameters, highlighting its strength in the sophisticated domain of TKG extrapolation.
Keywords:
Temporal knowledge graph extrapolation
Knowledge representation and reasoning
Subgraph information bottleneck
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W