Faithful Rule Learning and Extraction with Applications

Machine learning (ML) models have demonstrated strong performance across a variety of tasks over knowledge graphs (KGs) and databases, such as knowledge graph completion and recommendation. However, their limited interpretability makes them unsuitable for safety-critical domains or scenarios requiring legal or ethical compliance. This has motivated growing interest in extracting logical rules from trained ML models to explain their predictions. To this end, our works aim to establish formal guarantee on the precise relationship between a model and the rules extracted from it. Specifically, we aim to ensure the extracted rules exhibit soundness (or completeness), which means that the results obtained by applying the model to any input dataset always contain (or are contained in) the results obtained by applying the rules to the same dataset. In practice, we explore the tasks of knowledge graph completion and tabular data cell completion. In either case, we propose rule extraction approaches for the ML model with faithfulness guarantee. In the future, we will investigate broader applications of our work such as entity resolution, investigate other feature of extracted rules such as multiplicity, and more comprehensive benchmarking for rule learning.