Personalized Image Recommendation with Photo Importance and user-item Interactive Attention
Wan Zhang, Zepeng Wang, Tao Chen · 2019
Human encounter a variety of images when browsing some websites, such as Flickr, Pinterest and Instagram. How to incorporate user and item attributes to give users a personalized recommendation is challenging. Aiming at this problem, we propose a new model based on Bayesian Personalized Ranking while combining photo importance and user-item interactive attention. Specifically, we define photo importance according to the average user 'favor' information, which will be further added to Bayesian Personalized Ranking model as a weighting factor. In addition, considering that the interaction between users and items is mutual, that is, one user may likes a series of images and one image may also be liked by a series of users, So we introduce attention mechanism(i.e., user-item interactive attention) to indicate users' different preference on their interested images, and different preference of different users on the same image they like. Finally, we construct a new user representation and item representation which contains much richer user-item interactive information. Experimental results on real-world datasets demonstrate the superiority of our proposed model.