On Monday, December 15th, at 2 pm, Modou Gueye will defend his thesis at Telecom ParisTech, in room B312. Here is the abstract:
In this thesis, we address the scalability problem of recommender systems. We propose accurate and scalable algorithms.
We first consider the case of matrix factorization techniques in a dynamic context, where new ratings are continuously produced. In such case, it is not possible to have an up to date model, due to the incompressible time needed to compute it. This happens even if a distributed technique is used for
matrix factorization. At least, the ratings produced during the model computation will be missing. Our solution reduces the loss of the quality of the recommendations over time, by introducing some stable biases which track users’ behavior deviation. These biases are continuously updated with the new
ratings, in order to maintain the quality of recommendations at a high level for a longer time.
We also consider the context of online social networks and tag recommendation. We propose an algorithm that takes into account the popularity of the tags and the opinions of the users’ neighborhood. But, unlike common nearest neighbors’ approaches, our algorithm does not rely on a fixed number of neighbors while computing a recommendation. It uses a heuristic that bounds the network traversal in a way that enables computing the recommendations on the fly, with a limited computation cost, while preserving the quality of the recommendations.
Finally, we propose a novel approach that improves the accuracy of the recommendations for top-k algorithms. Instead of a fixed list size, we adjust the number of items to recommend in a way that optimizes the global accuracy of the recommendations. We other words, we optimize the likelihood that all the recommended items will be chosen by the user, and find the best candidate sublist (i.e., the most accurate one) to recommend to the user.