Learning with Relational Data: Sequential Models and Propagation Models for Structured Classification and Labeling
The content of my research is mainly motivated by two facts. The first one is that classical machine learning models are not well adapted to the complexity of real-life problems. They have been developped to handle simple data, and cannot be applied to complex structures. Moreover, the domain of Machine Learning (ML) clearly ignores many practical problems and the generic tasks that have been studied in the literrature, even the most famous one which is classification, are in fact very far from what happens concretly. For example, almost all ML models consider that the full knowledge of each input to classify is known, or that the objective function is easy to minimize through convex optimization. These assumptions are very strong and rarely discussed. The second fact is that, under these assumptions, and for some real applications, the domain is mature and the existing ML methods are very effective, robust and able to learn complex information.
In this context, the objective of my work is to build on top of these well-known models in order to propose effective learning methods that are able to better handle the complexity of real-life applications. My research tries to answer questions like how can we propose methods able to classify any complex data without manual adaptation ? how can we deal with tasks where the acquisition of information is expensive, slow, incomplete, etc ? These questions have been neglected during the last decades while they are fundamental for making learning usable by non-specialists, and thus to make computer able to solve problems "by themselves". The two main contributions presented here are directly inspired by these key questions and aims at providing one (partial) answer.
As a more prospective aspect of my research, my motivations are also to explore the use of Machine Learning techniques in the perspective of creating learning "living" agents able to decide what to do and when to act which are keys problems for developing an Artificial Intelligence. This means that the work presented has been made with the objective to bring new solutions for solving actual, industrial data mining problems, but also to open new research directions in the domain of Artificial Intelligence. This more prospective research is illustrated for example by the investigation of sequential learning methods that aim at learning how to do and by the development of models that are able to learn and infer on a complex knowledge.
Defence : 12/07/2012 - 10h - Campus Jussieu, grande salle de visioconférence, Atrium RdC, porte jaune entrée 2
Jury members :
M. Patrick Gallinari Président du jury
M. Eric Gaussier Rapporteur
M. Marco Gori Rapporteur
M. Philippe Preux Examinateur
M. Marc Sebban Examinateur
M. Louis Wehenkel Rapporteur
1 PhD student (Supervision / Co-supervision)
- GAYA Jean-Baptiste : Reinforcement Learning for Human-Machine Collaboration
- MAAG Maria : Apprentissage automatique de fonctions d’anonymisation pour les graphes et les graphes dynamiques.
- DULAC-ARNOLD Gabriel : Modèles Sequentiels pour la Classification Multiclasse, Sparse et Budgetée.
- JACOB Yann : Classification dans les graphes hétérogènes et multi-relationnels: application aux réseaux sociaux.
- GAO Sheng : Prédiction de liens par modèles à facteurs latents.