Découverte évolutive de connaissance à partir de graphes de données RDF
Rémi Felin · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
Knowledge graphs are collections of interconnected descriptions of entities (objects, events or concepts). They provide context for the data through semantic links, providing a framework for integrating, unifying, analysing and sharing data. Today, we have many factual data-rich knowledge graphs, and building and enriching them is relatively straightforward. Enriching these graphs with schemas, rules or constraints that allow us to check their consistency and infer implicit knowledge by reasoning is more difficult and costly. This thesis presents an approach based on the Grammatical Evolution technique for automatically discovering new knowledge from the factual data of a data graph expressed in RDF. This approach is based on the idea that candidate knowledge is generated from a heuristic mechanism (exploiting the graph data), is tested against the graph data, and evolves through an evolutionary process so that only the most credible candidate knowledge is kept. First, we focused on discovering OWL axioms that allow, for example, the expression of relationships between concepts and the inference of new facts previously unknown from these relationships. Candidate axioms are evaluated using an existing heuristic based on possibility theory, which makes it possible to consider the incompleteness of information in a data graph. This thesis presents the limitations of this heuristic and a series of contributions allowing an evaluation that is significantly less costly in computation time, thus opening up the discovery of candidate axioms using this heuristic. Second, we propose discovering SHACL shapes that express constraints that RDF data must respect. These shapes are useful for checking the data graph's consistency (e.g., structural) and facilitating new data integration. The evaluation of candidate shapes is based on the SHACL evaluation mechanism, for which we proposed a probabilistic framework to take into account errors and the inherent incompleteness of the data graphs. Finally, we present RDFminer, an open-source Web application that executes our approach to discovering OWL axioms or SHACL shapes from an RDF data graph. Through an interactive interface, the user can also control the execution and analyse the results in real-time. The results show that the proposed approach can be used to discover a wide range of new, credible and relevant knowledge from large RDF data graphs.