Description
Date depot: 5 avril 2024
Titre: Unsupervised Contrastive Learning for Attributed Networks
Directeur de thèse:
Mohamed NADIF (Centre Borelli (EDITE))
Encadrant :
Labiod LAZHAR (Centre Borelli (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: The main scientific objectives of this thesis project revolve around establishing a unified and adaptable framework for simultaneous co-clustering and embedding. These goals are outlined as follows: Initially, the aim is to construct a comprehensive mathematical model that seamlessly integrates data embedding and clustering, incorporating additional information such as graph links or data point constraints. This effort entails creatively formulating joint data representation and co-clustering within a contrastive context, directly applicable to real-world scenarios. Secondly, the project aims to devise robust and efficient algorithms to support data representation, contrastive learning, and clustering in attributed networks. This includes enhancing various objective functions and developing user-friendly software tools. Thirdly, the objective is to introduce novel evaluation metrics to assess the efficacy of co-clustering and embedding techniques, with evaluations conducted on intricate image and text datasets. Moreover, the thesis seeks to propose innovative unsupervised approaches tailored for multi-omics data, where contrastive co-clustering can alleviate the need for data augmentation. All proposed methodologies will undergo rigorous benchmark assessments to validate their effectiveness and performance.