Visual analytics for high reliable knowledge graph fact prediction

Fact prediction aims to complete a knowledge graph by predicting missing facts based on existing facts. Reinforcement learning based approaches are the mainstream methods for fact prediction. However, existing reinforcement learning based approaches derive rules solely from reachability between entities, which may introduce logical errors to rules. These rules result in false positives when predicting missing facts, thus hindering the reliability of the knowledge graph. Identifying false positives is time-consuming and laborious, requiring the examination of numerous fact prediction results and analysis of their contextual information to judge whether suspected results conform to the real world. To help users complete the identification of false positives, this paper thus proposes a visual analytics approach. A reliability indicator is proposed to quantify the reliability of fact prediction results from the perspectives of rule generation and rule application, indicating those that are likely to be false positives. A visual interface is designed to present the subgraph information, matching rule information, and case triple information of fact prediction results, assisting users in making review decisions. A series of experiments are conducted to demonstrate the effectiveness of the approach, including a performance evaluation experiment, a case study, and a user study.