Temporal Rule Mining for Knowledge Graph Completion.

Abstract

This thesis introduces StaTeR (Static to Temporal Rule learner), the first dynamic symbolic rule learner designed to generate temporal rules from static rules. StaTeR combines symbolic reasoning with temporal logic to enable link and time prediction on temporal facts. The process begins with learning static rules using a symbolic rule learner. For each static rule, five different temporal confidences are computed: Naive, Naive Overlap, Same Interval, Intersection Over Union, and Temporal Alignment Coefficient. These temporal rules are evaluated for their effectiveness in link and time prediction tasks. Furthermore, this thesis generated refined datasets for temporal datasets Wikidata12k and YAGO11k. Both datasets were cleaned in such a way that identical static facts do not have any overlap in their time interval. Furthermore, temporal ranges from before the year 0 were removed, ensuring compatibility with temporal knowledge graph tools. The performance of the StaTeR model was compared against state-of-the-art models on the two datasets for link and time prediction. We show that for some quality measures, StaTeR achieves competitive results against the state-of-the-art-systems for link and time prediction tasks.