From Recommendation to Generation: A Novel Fashion Clothing Advising Framework

Abstract

In the field of clothing recommendation, building a successful recommendation system means giving each user an optimal personalized recommending list. The top ranked clothing in the list are expected to meet a series of user's needs such as preference, taste, style, and consumption level. In online shopping, the most common way is to use user's explicit rating of items. However, user's implicit feedback such as browsing log, collection, and reviews may contains extra information to help model user's preference more accurately. In addition, the recommended clothing should also meet user's consumption level, which is an important factor easily overlooked in recommendation system. In this paper, we combine visual features of clothing images, user's implicit feedback and the price factor to construct a recommendation model based on Siamese network and Bayesian personalized ranking to recommend clothing satisfying user's preference and consumption level. Then on the basis of recommending clothing, we use Generative Adversarial Networks to generate new clothing images and use them to form a compatible collocation to provide fashion suggestions out of datasets.