I_ConvCF: Item-based Convolution Collaborative Filtering Recommendation

Abstract

Item-based collaborative filtering is widely used in industry to build recommendation systems because of its explanatory and efficiency in personalized recommendation. However, item-based collaborative filtering is mostly a shallow linear model, which cannot well mine the complex relationship between items. Therefore, in this work we propose a item-based convolution collaborative filtering model (I_ConvCF). Using a convolution neural network to extract the nonlinear relationship characteristics of Historical interaction/non-interactive items as a low dimensional latent factor. The target item is regarded as another low dimensional latent factor, and their product is regarded as the feature of the target item. We demonstrate their superiority in personalized ranking tasks on two real data sets.