Automatic tracking and control for web recommendation New approaches for web recommendation

International audienceRecommender systems provide users with pertinent resources according to their context and their profiles, by applying statistical and knowledge discovery techniques. This paper describes a new approach of generating suitable recommendations based on the active user's navigation stream, by considering long distance resources in the history. Our main idea to solve this problem is the following: we consider that users browsing web pages or web contents can be seen as objects moving along trajectories in the web space. Having this assumption, we derive the appropriate description of the so-called recommender space to propose a mathematical model describing the behavior of the users/targets in the web/along the trajectories inside the recommender space. The second main assumption can then be expressed as follow: if we are able to track the users/targets along their trajectories, we are able to predict the future positions in the sub-spaces of the recommender space i.e., we are able to derive a new method for web recommendation and behavior monitoring. To achieve these objectives, we use the theory of the dynamic state estimation and more specifically the theory of Kalman filtering. We establish the appropriate model of the target tracker and we derive the iterative formulation of the filter. Then, we propose a new recommender system formulated as a control loop. We validate our approach on data extracted from online video consumption and we derive a users monitoring approach. Conclusions and perspectives are derived from the analysis of the obtained results and focus on the formulation of a topology of the recommender space