Knowledge Graph Reasoning with Self-Supervised Reinforcement Learning

Abstract

The University of Central Florida invention is a new knowledge graph completion (KGC) technique that provides reasoning paths and infers missing facts—functionality not available in other self-supervised reinforcement learning (SSRL) approaches. Reinforcement learning (RL) is an effective method to find reasoning pathways in incomplete knowledge graphs (KGs). To overcome the challenges of sparse rewards and the explore-exploit dilemma, the UCF invention comprises a self-supervised pretraining method to warm up the policy network before the RL training stage. The seeding paths used in the supervised pretraining stage are generated by searching the 3-hop neighborhoods of start entities in a subset of training facts. The UCF SSRL method, with partial labels, combines the fast learning speed of RL and wide coverage of supervised learning (SL). It meets or exceeds current state-of-the-art results for all KG reasoning tasks.

Partnering Opportunity

The research team is seeking partners for licensing, research collaboration, or both.

Stage of Development

Prototype available.

Benefit

Market Application