Abductive Reasoning over Temporal Knowledge Graphs via Logical Hypothesis Generation

Rushan Geng, Cuicui Luo

Keywords: Temporal knowledge graph, Abductive reasoning, Temporal knowledge graph reasoning, Reinforcement learning

Abstract: Abductive reasoning (ABR) aims to infer plausible hypotheses that explain observed facts. Existing studies have mainly focused on abductive reasoning over static knowledge graphs, while the temporal setting remains underexplored. In this paper, we investigate \textbf{abductive reasoning on temporal knowledge graphs (ABTKG)} and propose a dedicated framework for this task. We first generate logical hypotheses that explain an observation at a given time, and then train a temporal hypothesis generator through supervised learning. However, supervision alone is insufficient to handle unfamiliar observations at specific time points. To address this limitation, we introduce a reinforcement learning objective defined on the TKG, which reduces the gap between the observation and the conclusion produced by the generated hypothesis. Experiments on four public TKG datasets demonstrate consistent improvements in both explanatory power and reasoning accuracy, with gains observed across all datasets.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 2509