Multi-Modal Knowledge Graph Reasoning with Causal Relation Modeling

Abstract

This paper addresses the limitations of traditional knowledge graph reasoning, which primarily relies on correlational relationships, leading to potentially biased inferences. We propose a novel approach to multi-modal knowledge graph reasoning by explicitly modeling causal relationships between entities. Our core claim is that incorporating causal reasoning significantly enhances both the accuracy and interpretability of knowledge graph inferences. We leverage Graph Neural Networks (GNNs) to model the knowledge graph, coupled with causal inference algorithms, specifically causal graph learning methods, to identify and infer these underlying causal links. This framework allows for a more robust and reliable understanding of complex relationships within the graph. The results demonstrate improved reasoning performance compared to methods solely based on correlation. This work contributes to a more nuanced and accurate representation of knowledge, paving the way for more intelligent applications in areas such as decision support and automated reasoning.