Grant: Discovery Grants Program - Individual

Recent advances in data modeling, storage and analytics have fostered the creation of large datasets by organizations, e.g. as a decision-making support. Often, the corresponding analytical process is knowledge-intensive, i.e. involves knowledge sources such as knowledge graphs and domain ontologies. The respective data are also organized in graph formats, typically as linked data. This program aims at helping such organizations acquire explicit knowledge from their data: For instance, by revealing fine-grained associations between inter-related entity types --e.g. particular medical conditions, patients of specific genetic profile, drug classes and adverse effects-- useful insights might be gleaned to support further domain-related reasoning, in this case, for pharmacovigilance or personalized medicine. Our overall objective of this program is to enable the unsupervised mining of interesting patterns from linked data using the two expressive and complementary pattern languages, one based on logic and the other one on graphs. This objective splits into three technical ones: 1) design logic-based pattern miners of better efficiency, focus and user support while reusing bases of existing approaches, 2) design efficient graph pattern miners that discover large patterns yet avoid combinatorial explosion, and 3) elaborate methods for combining both sorts of patterns to mutually enhance their respective results and thus provide a more holistic view of the regularities in the data. In pursuing these objectives, we shall conduct research on graphs, pattern mining, algorithmic design, etc. Major challenges include the complexity of the target pattern structures, e.g. the multidimensional nature of graph data which involve costly matching, the combinatorial explosion of patterns, the irregular space induced by a domain ontology, etc. The target methods will provide fully interpretable outputs and therefore they qualify as « explainable AI ». A In particular, our methods will fit well in the context of clinical studies where the complexity of the patient records and their relatively small number would seriously impede deep learning-based solutions. Through my collaborations the developed methods will be tested and validated in a variety of practical situations from manufacturing, agriculture, education, etc. Thus, our tools should help a number of Canadian companies, public services and non-profit organizations properly analyze -hence more thoroughly master- their data while taking advantage of the available computerized domain knowledge.