Bridging Ontologies and Symbolic Learning - Toward Constraint-Aware Pattern Extraction

Knowledge Graphs (KGs) offer a simple yet flexible way of data representation and interconnection by relating entities via predicates in triples of the format (subject, predicate, object). However, they struggle with completeness and correctness of their data. With a rising overall interest in KGs, there is a growing research area focused on tackling these problems by producing Inductive Knowledge. One method for this is Rule Mining, which describes the process of finding recurring patterns in a KG that can be formulated into rules, which state the existence of specific structures in a KG that can imply new facts. Many approaches do not incorporate external knowledge, disregarding ontological and data constraints. Presented is RON-a-CON, a new approach that does consult metadata. Employing a graph traversal strategy similar to RuDiK [1], preceded by normalisation and validation of the given KG, and only considering triples in line with provided domain and range information for properties, RON-a-CON produces compact sets of constraint-aware rules. In the downstream task of Link Prediction with Knowledge Graph Embedding models, equal or better results are achieved in most cases when enriching KGs with predictions made by rules from RON-a-CON compared to those from AMIE3 [2] and AnyBURL [3].