SPARQL-based relaxed rules for learning over knowledge graphs

In today’s world where explainable AI is gaining importance, rule learning is a classic but favourable machine learning approach that can be enhanced using advances in semantic web technologies. The expressiveness of rules learnt over Knowledge Graphs (KGs) is largely dependent on the language bias of a rule learner which in turn determines the complexity of the search space for models. In this paper, we propose a relaxed rule specification
approach that allows rules to contain branches and tree shapes, negated graph patterns, and SPARQL style filtering which adds numerical and temporal comparisons into the rule’s vocabulary. By targeting compatibility with SPARQL, we can tap into the advanced query optimisation features of triple stores which implement SPARQL to evaluate and apply weakly-constrained rules as SPARQL queries. We use an expert-guided rule-template based approach to generate candidate rules and we extend standard rule quality measures for such relaxed rules. We introduce a further extension that can be used to extract numeric attribute values in order to include numeric attributes within our symbolic rule learning framework.