Yannis Karmim - PhD Student

News

  • (29/09/2024): Looking for a postdoc position for september 2025 on machine learning on graphs.
  • (26/09/2024): Our paper on dynamic graph transformer is accepted at NeurIPS2024!

  • (13/09/2024): Presentation of our work on the temporal receptive field at the ECML2024 conference in Vilnius.
  • (15/07/2024): Our work on temporal receptive field is accepted at ECML2024 Machine Learning on Graphs Workshop !
  • (01/07/2024): Our work on improving GNN for top-k recommandation is accepted at TMLR (Transaction in Machine Learning Research) !

Phd Abstract

In many situations, graphs represent data that are intrinsically dynamic (communications between people, financial exchanges, etc.). However, due to a lack of consensus on modeling, static graphs are often used, aggregating more or less important periods of time and thus destroying information. Graph Neural Networks have also been designed from static graphs (possibly series of snapshots). And it is only recently that models have emerged to take into account a continuous time and have proposed applications. Several questions are therefore open, in order to move towards better GNN models able to integrate the temporal dimension of data, to efficiently combine spatial and temporal diffusion of information. The transition to large scale, the explicability of the models, as well as a standardization of the evaluation are also issues for the coming years in the field.