Recursive RNN Based Shift Representation Learning for Dynamic User-Item Interaction Prediction

Accurately predicting user-item interactions is critically important in many real applications including recommender systems and user behavior analysis in social networks. One limitation of most existing approaches is that they use the sparse user-item interaction relationships directly, but ignore the second order user-user and item-item relationships. Another limitation is that they generally embed users and items into different embedding spaces in a static way, but cannot capture the dynamic and evolving dependency between users and items and embed them into a unified latent space. In this paper, we aim to learn dynamic embedding vector trajectories rather than static embedding vectors for users and items simultaneously. A Recursive RNN based Shift embedding method called RRNN-S is proposed to learn the continuously evolving embeddings of users and items for more accurately predicting their future interactions. Specifically, we first propose to quantize the user-user and item-item relationships from the original user-item interaction graph, which can be used as auxiliary information to enrich the sparse user-item interaction graph. A recursive RNN is proposed to iteratively and mutually learn the dynamic user and item embeddings in the same latent space based on their historical interactions. A shift embedding module is next proposed to predict the future user embedding. To predict the item which a user will interact with, we innovatively output the item embedding instead of the pairwise interaction probability between users and items, which is much more efficient. Through extensive experiments on two real-world datasets, we demonstrate that RRNN-S achieves superior performance by comparison with several state-of-the-art baseline models.