We are going to present one novel parallel algorithm to perform alternating least squares for tensor completion applied to recommendation systems.
Parallel Higher Order Alternating Least Square for Tensor Recommender System (Romain Warlop, Alessandro Lazaric, and Jérémie Mary) [pdf]
Many modern recommender systems rely on matrix factor- ization techniques to produce personalized recommendations on the basis of the feedback that users provided on different items in the past. The feedback may take different forms, such as the rating of a movie, or the number of times a user listened to the songs of a given music band. In some situ- ations, the user can perform several actions on each item, and in this case the feedback is multidimensional. For in- stance, the user of an e-commerce website can either click on a product, add the product to their cart or buy it. Another example is restaurant rating, where the user may rate not only the quality of the food but also service and decoration. When dealing with multiple actions, one cannot view the recommendation problem as a matrix complexion unless the problem is considered as a series of multiple independent problems. In this work we propose to use a tensor approach to learn all this feedback simultaneously and benefit from transferring knowledge between each kind of feedback. Our work can be seen as a transfer learning application to recom- mender system as well as a multi-task recommender system learning where each task represents each action a user can perform. The proposed approach perform effective paral- lel tensor factorization in order to complete the tensor and make recommendation that we validate experimentally.