Fashion coordinates recommendation based on user behavior and visual clothing style

Fashion coordinates recommendation focuses on coordinating the pattern and colors of fashion items. For example, which shirt out of the candidates matches the pants best? The traditional recommender system based on Latent Factor Model (LFM) uses user behavior features to deal with the recommendation of two matching item, but it is weak in recommending more than two items because the problem of cold start becomes significant. Furthermore, fashion coordinates are correlated with visual features of items directly, which is another factor that can be imported into the recommendation calculation. In this paper, we propose a fashion coordinates system which considers both user behaviors and visual fashion styles. We extend LFM to dealing with user behavior features, and a deep learning model called Denoising Autoencoder is used to process visual features. Combining those two methods, our system can recommend multi-items more accurate than traditional methods and support cold start. Initial experiments on coordinating three kinds of items (tops, trousers and shoes) are demonstrated in detail.