Research

Methodological areas

The work presented below spans several methodological areas, organized around complementary machine learning problems.

1. 01

   Learning from heterogeneous, multi-view and multimodal data

   I focus on developing methods able to jointly exploit several heterogeneous sources of information — images, text, signals, relational data. The goal is to leverage the complementarity between these modalities to improve clustering quality, model robustness and their ability to adapt to different application contexts. My contributions deal in particular with information fusion and collaborative learning across multiple representations.

   fusion d'informations · apprentissage multi-vues · données multimodales

2. 02

   Attributed, multiplex and multi-layer networks

   Part of my work deals with the analysis of complex relational data, where the topological structure must be studied jointly with the attributes carried by nodes and edges. In this context, I have proposed several approaches for community detection and clustering in attributed, multiplex or multi-layer graphs. This work aims to better characterize the hidden structures in complex networks and to effectively integrate relational and attribute information.

   détection de communautés · clustering · graphes attribués

3. 03

   Transfer learning and domain adaptation

   I develop domain adaptation methods, mainly unsupervised and semi-supervised, to handle situations where the training and test data do not follow the same distribution. A significant part of this work relies on hierarchical optimal transport, used both as a methodological tool and as a theoretical framework. I have thus contributed to approaches that align structures between source and target domains, exploit latent clusters in the target domain, and provide theoretical guarantees for single-source, multi-source and semi-supervised adaptation.

   transport optimal hiérarchique · adaptation de domaine · apprentissage par transfert

4. 04

   Deep neural networks and explainability

   My recent work also addresses the explainability of deep learning models, particularly in contexts where the transparency of decisions is a central concern. I am more specifically interested in approaches based on interpretable concepts, which link a model's predictions to understandable semantic representations.

   explicabilité · concepts interprétables · apprentissage profond

These methodological areas are developed in connection with several application domains: e-health, mobility, multimodal emotion analysis, chronic pain research, embedded systems, and more recently the integration of AI models in constrained environments.