Colloquium Artificial Intelligence - Giacomo Zamprogno, VU Amsterdam

Title: Extending Mined Rules for Link Prediction with Schema-based Exceptions

Abstract:

Link Prediction (LP) is the task of expanding information in Knowledge Graphs (KGs) by suggesting new relations between their entities. Mining and applying inference rules for this task can be used, with competitive performances, to maintain transparency with respect to machine learning methods for LP which behave as black-boxes. Yet, these rule-based methods suffer from a lack of expressivity and rarely leverage semantic information of the KG, focusing instead on entity-level triples, leading to inconsistencies when considering the KG schema restrictions.

In this work, we propose to extend rule expressivity with non-monotonicity. In particular, we leverage KG schema information to extend the mined rules with exception cases, in order to make the KG more consistent and semantically accurate. We automatically define exception cases based on the predicted relations, pruning the search space when the rule could produce semantically incorrect facts. We evaluate our approach by analyzing the quality of the materialized triples and the additional computing time required in four KGs of different size and curation level, and we further verify the change in rank-based LP metrics, using and extending rule sets mined by two rule miners. Our results suggest that the semantic quality of mined rules varies depending on the level of curation and schema-complexity of the KG. Further, we show that applying exceptions at single-rule level (thus ignoring between-rule interactions) allows to substantially reduce the number of inconsistent triples in all cases, at the cost of a limited increase in inference time. We finally report that rank-based LP metrics are influenced by the different rule expressivity in a limited way, further suggesting the need for integration with metrics that can take into account the semantic quality of predicted triples.