Machine learning algorithms for user interactions analysis
The research described in this habilitation document proposes new machine learning methods dedicated to the study of users interactions on three main levels.
First, we are interested in the analysis of user interaction in the form of users’ activity traces. To this aim, we develop new incremental clustering algorithms for large scale data sets or data streams. Unlike existing approaches, our methods are not limited to numerical data processing and are paired with tools to research typical user profile and that provide dynamic web paths visualization to facilitate interpretation of the analyzes.
Secondly, we are interested in semi-supervised clustering methods which allow the integration of expert knowledge in the clustering process. In this context, we have proposed active learning algorithms that optimize interactions between the human expert and the classification algorithms to improve their performance.
Finally, we are interested in the problem of automatic extraction of metadata from structured corpus in the context of information retrieval. The specificity of this work is twofold: on one hand we have introduced methods that are independent of the document's content and, on the other hand, we have proposed methods based on the content that combine statistical learning algorithms and contextual, stylistic and linguistic features.
Defence : 12/13/2012 - 10h30 - Site Jussieu 25-26/105
Jury members :
Boughanem Mohamed Université Paul Sabatier / IRIT PR [rapporteur]
Gançarski Pierre Université de Strasbourg / LSIIT PR [rapporteur]
Poncelet Pascal Université de Montpellier 2 / LIRMM PR [rapporteur]
Artières Thierry UPMC / LIP6 PR
Bouchon-Meunier Bernadette CNRS / LIP6 DR