Learning and inference engine applied to ubiquitous recommender system

Djallel Bouneffouf
Institut Telecom, Telecom SudParis, CNRS UMR Samovar 9, rue Charles Fourier, 75004 Evry Djallel.bouneffouf@it-sudparis.eu

The need for adapting information systems to the user context has been accentuated by the extensive development of mobile applications that provide a considerable amount of data of all types (images, texts, sounds, videos, etc.). It becomes thus crucial to help users by guiding them in their access to information. Systems should be able to recommend information helping the user to fulfill his/her goal. The information given by the system depends on the users situation, i.e. an instance of the context. Possible situations and the associated actions reflect the users work habits. Major difficulties when applying techniques to adapt a system to the user follow: - Avoiding the intervention of experts: on one hand, experts are not sure of the interest of the user, may define wrong ideas about him; on the other hand, an expert is not always available [9, 26]. - Starting from scratch: in the initial state, the systems behavior should not be incoherent for the user to not refuse it quickly [7, 27, 29]. - A slow learning process: the learning process has to be quick to avoid bothering the user with incorrect recommendation [11, 22, 23, 24]. -The evolution of the users interest: the interest of the user may change with the time. The system has to be continuously adapted to this dynamic change using the users context information to provide the relevant recommendations because, if the system behavior is incoherent, the user refuses it quickly [9, 19, 31]. We sum up all of these problems in the following scenario. Knowing the high mobility of its employees and their dependencies to the information contained in their corporate databases, the Nomalys company has equipped all mobile phones with the NS application. This application allows them to adapt to the nomadic life by consulting the companys database from their mobile phones. Knowing the high mobility of its employees and their dependencies to the information contained in their corporate databases, the Nomalys company has equipped all mobile phones with the NS application. This application allows them to adapt to the nomadic life by consulting the companys database from their mobile phones. Because of the diversity of jobs existing in the company, Nomalys decides provide the application with a generic recommender system, which has to retrieve the relevant information to users without any set of actions predefined by an expert.

Paul, John and Lauren are new employees of the company integrating different teams of the company (marketing, commercial, and technique resp.). Regarding their agendas, they have a meeting with clients in Paris at midday. When they arrive at their meetings, the system should recommend them the relevant information which would help them to better manage their meetings. The system recommends to Paul the register of complaints, to John the register of factures and to Lauren the technical registers. To do these recommendations without the need of an expert and avoiding starting from scratch, the recommender system has to infer them from the actions of the users team. The recommender system finds that Paul often opens the register of complaints two hours before his meeting and not at the meeting. Moreover, John always tries to find companies which are near and do the same work as the one he will visit the next day. Using this knowledge, one month later, the system is able to recommend the register of complaints to Paul two hours before his meeting; it also recommends John companies which are near and do the same work as the one he will visit the next day. To do these recommendations, the learning process of the system has to be quick and has to follow the evolution users interest. In summary, to solve the problem of the scenario, I study during my PHD thesis the possibility to start with a predefined set of actions, not defined by an expert, but by the users social group (in the scenario we talk about job teams) and adapts it progressively to a particular user. This default behavior allows the system to be ready-to-use and the learning is a lifelong process. Thus, the system will, at first, be only acceptable to the user, and will, as time passes, give more and more satisfying results. 8. Bibliographie
[Assad et al., 2007] Mark Assad, David Carmichael, Judy Kay et Bob Kummerfeld. PersonisAD : Distributed, Active, Scrutable Model Framework for Context-Aware Services . Dans Anthony La-Marca, Marc Langheinrich et Khai N. Truong, rdacteurs, Proceedings of the 5th International Conference on Pervasive Computing, PERVASIVE 2007, tome 4480 de Lecture Notes in Computer Science, pages 55 - 72. Springer, Toronto, Ontario, Canada, mai 2007. [Godoy et Amandi, 2005] Daniela Godoy et Analia Amandi. User proling for Web page filtering , Internet Computing, IEEE, tome 9, n 4,pages 56-64, juillet - aout 2005. [Christopher, 2006] Christopher M. Bishop, Pattern Recognition And Machine Learning, Springer, 2006.