Flow to Candidate: Temporal Knowledge Graph Reasoning With Candidate-Oriented Relational Graph.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Temporal knowledge graph (TKG) reasoning is crucial for inferring future events from historical data.
• Current state-of-the-art (SOTA) subgraph-based methods excel at local information but struggle with complex topological patterns.
• Path-based methods capture relation sequences but may lose information compared to subgraphs.

Purpose of the Study:

• To propose a novel subgraph-based approach for temporal knowledge graph reasoning.
• To effectively capture complex relational and topological patterns within TKGs.
• To improve the accuracy and efficiency of future event prediction.

Main Methods:

• Constructs candidate-oriented relational graphs to represent local TKG structures.
• Employs a variant graph neural network with a recursive propagation architecture and self-attention mechanism.
• Utilizes a prior directed temporal edge sampling method and a scoring function for prediction.

Main Results:

• The proposed approach achieves stronger inference capabilities compared to SOTA methods on benchmark datasets (ICEWS14, ICEWS18, ICEWS0515, YAGO).
• Demonstrates faster convergence rates in experimental evaluations.
• Provides interpretable relational graphs for each query-candidate pair, enhancing result transparency.

Conclusions:

• The novel subgraph-based method effectively captures complex relational patterns in TKGs.
• The approach offers significant improvements in inference accuracy and convergence speed.
• The interpretability of the generated relational graphs aids in understanding TKG prediction outcomes.