Optimize Rule Mining Based on Constraint Learning in Knowledge Graph

Knowledge graph completion (KGC) involves inferring unknown triples from known ones to improve the completeness and richness of information and semantics in knowledge graphs (KG). Most existing KGC works are dominated by the embedding method, but it fails to explain the predicted triples. Rule mining methods can achieve this, but it is challenging to mine rules from a large KG under the open-world assumption. Although several mining methods have already addressed this issue, those based on search lack the ability to learn during the search process, which leads to the generation of ineffective rules. In this paper, answer set programming is used to convert search-failed rules into constraints to avoid rule specializations. Through constraint learning, the search process is continuously improved to optimize mining. Additionally, this paper utilizes the embedding method to pre-enhance KG and introduces a comprehensive evaluation approach to tackle the weak generalization of the rule mining method. Extensive experiments show that the proposed method achieves the best accuracy and efficiency.