Recommandation et analyse de sentiments dans un espace latent textuel

Recommender systems were developed to cherry-pick interesting content in an always growing environment to help users discover new products or information that they might like. Traditional collaborative filtering methods mainly focus on the ratings to establish user profiles, ignoring the joint review text. Our work does the opposite. We use the written comments as the primary source of information and propose, in addition to rating forecast, to predict review text. Our approch also enables us to elegantly tackle the cold-start problem. Using state-of-the-art text embedding technique, we build a latent review space where all the components of a review is mapped. The created vectors are used to predict both rating and review text. MOTS-CLÉS : Recommandation, Analyse de sentiments, Apprentissage de représentations textuelles