Personalized Fashion Recommendation with Visual Feature based on Conditional Hierarchical VAE

Abstract

Recently, with the increasing of online shopping services, there has been a lot of research on fashion item recommendation. Unlike standard recommendation systems, a recommendation for fashion items needs to take into account not only external factors such as season, temperature, and weather, but also fashion-specific image features such as color and design, and mutual dependencies in which important features change depending on the item. In this study, we propose a recommendation system that takes these factors into account. First, discrete features such as labels are insufficient to represent the colors and designs of items. Therefore, we propose a conditional hierarchical VAE that captures the continuous latent space of color and design to solve this problem. Second, important attributes such as color and size change depending on the item category, etc. Therefore, we propose a mechanism to capture the interdependence between the item attributes using the attention mechanism. In our experiments, we show that our model outperforms existing methods and achieves the same or better performance as human stylists.