Combinatorial optimization based recommender systems.

Recommender systems exploit a set of established user preferences to predict topics or products that a new user might like [2]. Recommender systems have become an important research area in the field of information retrieval. Many approaches have been developed in recent years and the interest is very high. However, despite all the efforts, recommender systems are still in need of further development and more advanced recommendation modelling methods, as these systems must take into account additional requirements on user preferences, such as geographic search and social networking. This fact, in particular, implies that the recommendation must be much more “personalized” than it used to be.
In this paper, we describe the recommender system used in the “DisMoiOu”(“TellMeWhere” in French) on-line service (http://dismoiou.fr), which provides the user with advice on places that may be of interest to him/her; the definition of “interest” in this context is personalized taking into account the geographical position of the user (for example when the service is used with portable phones such as the Apple iPhone), his/her past ratings, and the
ratings of his/her neighbourhood in a known social network.
Using the accepted terminology [6], DisMoiOu is mainly a Collaborative Filtering System (CFS): it employs opinions collected from similar users to suggest likely places. By contrast with existing recommender systems, ours puts
together the use of a graph theoretical model [4] and that of combinatorial optimization methods [1]. Broadly speaking, we encode known relations between users and places and users and other users by means of weighted graphs. We then define essential components of the system by means of combinatorial optimization problems on a reformulation of these graphs, which are finally used
to derive a ranking on the recommendations associated to pairs (user,place). Preliminary computational results on the three classical evaluation parameters for recommender systems (accuracy, recall, precision [3]) show that our system performs well with respect to accuracy and recall, but precision results need to be improved.