11th March 2019
Speaker: Benjamin Piwowarski, researcher CNRS, France.
Abstract: Learning to represent users and items in a homogeneous vector space is central to collaborative filtering systems. Most approaches suppose that the representation is reduced to a single point, that is, that a representation is certain. Uncertainty can have various causes related to the lack of information (isolated users/items) or because of the contradiction between neighboring nodes. Our hypothesis is that, because of these different factors, training will result in learned representations with different confidence, and that this uncertainty is important for many inference problems. For that, we leverage Gaussian embeddings, which have been firstly proposed for learning word embeddings (Vilnis et al., 2015), and show their application in collaborative filtering and other areas. We discuss how this can help to integrate heterogeneous meta-data information (e.g. plot of a movie).