Kamrul Islam (Capsid) will defend his thesis, entitled “Explainable link prediction in large complex graphs – application to drug repurposing”, on Friday, 16 December at 9.30 am in room B013.
Composition du jury:
Rapporteurs:
Luc Brun, Professeur, ENSICAEN, France
Paolo Merialdo, Professeur, Université Rome III (Roma Tre University), Italie
Examinateurs:
Miguel Couceiro, Professeur, Université de Lorraine, France
Fatiha Saïs, Professeure, Université Paris Saclay, France
Invité : Marie-Dominique Devignes, Chargée de Recherches, CNRS, France
Encadrants:
Malika Smaïl-Tabbone, Maître de conférences, Université de Lorraine, HDR, France
Sabeur Aridhi, Maître de conférences, Université de Lorraine, France
Abstract:
Link prediction is one of the most interesting and long-standing problems in the field of graph mining; it predicts the probability of a link between two unconnected nodes. This thesis presents several contributions for link prediction in simple graphs and knowledge graphs. Firstly, we compare a few similarity-based and embedding-based link prediction approaches in different simple graphs with diverse properties and analyze their interesting connections to alleviate the “black-box” limitation of embedding-based approaches. Secondly, we develop an explainable supervised link prediction approach for simple graphs. Thirdly, we develop a negative triple sampling method which are useful for training of embedding methods for knowledge graphs. Fourthly, we develop a rule mining method for knowledge graphs and an explanation strategy using mined rules to explain embedding-based link predictions. Fifthly, we apply our explainable link prediction approach to a biological knowledge graph for drug repurposing of COVID-19. Finally, we present a new framework for distributed training of knowledge graph embedding methods.