Pierre-Edouard Osche (Kiwi) will defend his thesis on Friday, 26th February at 9 am.

His thesis is entitled “Sequence-based recommendations in a multidimensional space” and supervised by Anne Boyer and Sylvain Castagnos.

Abstract:
Recommender systems are a fundamental research topic at the intersection of several major disciplines such as machine learning, human-computer interaction and cognitive sciences. They also constitute an ambitious application framework for the community of researchers in Artificial Intelligence by their great complexity and the numerous constraints they generate.

The purpose of these systems is to improve the interaction between the general audience and the systems of search and access to information. It has become difficult to identify the most relevant items in the context of big data. The goal is thus to assist users in their explorations (whether in a virtual or physical environment), but also to propose items that may interest them but that they would not consult spontaneously.

Current systems have largely proven their added value and are based on various machine learning techniques (numerical or symbolic, supervised or not, etc.) [Castagnos, 2008]. Nevertheless, they still suffer from limitations when making recommendations of sequences (recommending items in a specific order may depend on requirements, progressiveness, context, time constraints, etc.). Some models, such as the DANCE model [Castagnos, 2015], integrate this temporal dimension by following in real time the evolution in diversity of resources consulted by users to better understand the exploration context. In [Bonnin, 2010], the author also proposes a temporal model capable of detecting frequent consultation patterns in a history of consultations, in order to provide a priori resource recommendations related to the same context. Nevertheless, while temporal and spatial modeling have been made possible [Zheng, 2015], state-of-the-art models that focus on sequence recommendations or on the overall quality of the sequence are still too rare.

In the framework of this thesis, we will focus on defining a new formalism and a methodological framework allowing : (1) the definition of human factors leading to decision making and user satisfaction; (2) the construction of a generic and multi-criteria model (physical or temporal constraints, diversity, progressiveness, etc.), integrating these human factors in order to recommend relevant resources in a coherent sequence; (3) a holistic evaluation of user satisfaction with its recommendation path. The evaluation of recommendations, all domains included, is currently done recommendation by recommendation with each evaluation metric taken independently (accuracy, diversity, novelty, coverage, …). Thus, we expect a more comprehensive evaluation framework, measuring the progressiveness and the completeness of the path.

Such a multi-criteria recommendation model has many application frameworks. As an example, it can be used in the context of online music listening with the recommendation of adaptive playlists (recommendation of music sequences to change the atmosphere in a place such as a bar, to raise or lower the emotion felt by the audience progressively, or to adapt to the complementary/similar/different expectations of a group). It can also be useful to adapt the recommendation path to the learner’s progress and the teacher’s pedagogical scenario in an e-education context. Let us also mention the tourism field where this model could integrate the spatial and temporal constraints of a physical environment (cities, museums, etc.).

Keywords: Recommender systems, Multi-agent systems, User modeling.

Committee:

Rewievers:
– Mme Sylvie Calabretto, Professeur, INSA de Lyon, France
– M. Laurent Vercouter, Professeur, INSA de Rouen, France

Examiner:
– M. Laurent Vigneron, Professeur, Université de Lorraine, France

Supervisor:
– M. Sylvain Castagnos, Maître de conférences, Université de Lorraine, France