Causal Inference KGC: A Multi-Hop Reasoning Framework Over Sparse Knowledge Graph

Compared to traditional Knowledge Graphs (KGs), the data sparsity and incompleteness of sparse knowledge graphs more significantly impact the performance of downstream applications. In order to seek an effective and interpretable method for Knowledge Graph Completion (KGC), particularly focusing on multi-hop inference, considerable attention has been devoted to research in KGs. However, existing methods usually use a pre-cropped action space of the Markov Decision Process (MDP), which leads to insufficient learning and inference ability of the inference model and difficulty in converging intelligence training. To address these issues, we propose a causal inference knowledge graph completion framework, named CIKGC, which exploits causal relationships between entities to predict some of the missing-tailed entities. We utilize the strong reasoning ability of reinforcement learning (RL) to accomplish multi-hop reasoning tasks in KGs. In the sampling process, we utilize bidirectional rejection sampling for negative sampling. The experimental results show that CIKGC framework performs well on multiple benchmark datasets and effectively addresses the lack of interpretability in sparse knowledge graph complementation.