Personalized style recommendation via reinforcement learning

Pinterest fashion and home decor searchers often have different style tastes. Some existing work adopts users’ past engagement to infer style preference. These methods cannot help users discover new styles. Other work requires users to provide text or visual signals to describe their style preference, but users often are not familiar with style terms and do not have the right image to start with. In this paper, we propose a reinforcement learning (RL) method to help users explore and exploit style space without requiring extra user input. Experimental results show that our method improves the success rate of Pinterest fashion and home decor searches by 34.8%.