Guénaël Cabanes

Associate Professor - University Paris XIII

Research Interest

Machine Learning

  • Data mining
  • Unsupervised learning (clustering)
  • Sequential data and data stream analysis
  • Autonomous learning
  • Neural Networks
  • Complex data

Human and Animal Behaviour

  • Self-organisation, emergent behaviour
  • Complex structure and dynamic of social groups
  • Individual decision-making and collective intelligence
  • Spatio-temporal behaviour
  • Agent based modelling

Post-doctoral project - University of Sydney

Subject : "Optimization in a competitive context".

Abstract: Solving problems in dynamic environments has relevance to many systems, both biological and artificial. Most organisms live in dynamic environments where conditions are constantly changing and where survival may depend on the ability to respond to these changes. Given the richness of problem solving in Nature, can't ‘biology-inspired’ algorithms profit from more extensive biological underpinnings? The main goal of this project is to use nest-site selection and foraging by ant colonies as a paradigm for investigating how complex systems adapt their decision-making in a competitive environment. Understanding the mechanisms underlying such adaptations will provide the basis for the development of a new class of data-mining algorithms.

Ph.D. - University Paris XIII

Under supervision of Younès Bennani (LIPN, Apprentissage Artificiel & Applications).

Subject : "Two-level Unsupervised Clustering driven by neighborhood and density".

Abstract: The research outlined in this thesis concerns the development of approaches based on Self-Organizing Map (SOM) for the discovery and the monitoring of class structures in the data through unsupervised learning. We propose a simultaneously two levels clustering method. This method is based on the estimate, from the data, of connectivity and density values of the SOM's prototypes. The number of clusters is detected automatically. Moreover, the complexity is linear with the number of data. We show that it is relatively simple and efficient to adapt these algorithms to variants of the SOM in order to obtain a versatile method capable of analyzing different data types. We also propose an improvement of the quality of the SOM using the connectivity values during the learning of the prototypes. In addition, we combine the clustering algorithm to measure similarity between distributions for the analysis of evolutionary data, and we propose an algorithm for monitoring data stream. These algorithms are all based on an estimate of the underlying density for learning a modified SOM.