New Methodologies for Fashion Recommender Systems

Sammanfattning: Traditional Recommender Systems rely on finding similarities between users and/or between items. In its broadest definition, a Recommender System tries to predict the preference a user would give to an item. While Content-Based approaches try to discover similarities between items and then predict a user’s preference based on its past interactions with the items, Collaboritve-Filtering approaches try to find similarities between users in order to recommend to an user what similar users already bought. In the fashion domain, though, users may not buy an item per se, but rather because it would fit in an ideal outfit that they may want to wear. This behaviour implies that the items content similarity between the items already bought by the user is not enough to make accurate predictions. Thus, it would be more reasonable to suggest the purchase of compatible clothes, rather than similar ones. The problem of scoring the compatibility of different outfits and learning a concept of style has already been tackled in the research community by the application of different machine learning techniques. However, the tasks and the datasets used to evaluate state-of-the-art models make some unrealistic assumptions that would not hold in a real-case scenario. This thesis introduces a novel algorithm to tackle the problem of learning outfit styles, in order to classify ensemble of clothes as fashionable outfits and complete them in a fashionable manner. Moreover, this work presents a proper comparison with the state-of-the-art models on the most used public academic datasets in this domain and on a real industrial dataset, provided by H&M as industrial partner of this research. In addition to this, a novel evaluation task, that releases some of the constraints existing in the tasks presented in literature, is introduced in order to asses the potentials of the different algorithms when dealing with problems more similar to those faced in real-case scenarios. Finally, this thesis attempts to move the problem of outfit completion from a general classification problem, into the recommender Systems domain and evaluates the performances of these algorithms using some of the typical metrics used in Information Retrieval problems.