TideFC: Learning Temporal Interaction for Dynamic Embedding via Feature Crossing

Recommendation system is becoming more and more important. Specially in social network, such as Twitter, Facebook, and YouTube. In the beginning, a number of studies learn the relationship of users and items from a bipartite graph. They embed each user and item in an embedding space. However, they ignore temporal properties. They regard users and items as static embeddings. As time passed, user preferences and item concepts can change, and node embeddings should be dynamically adjusted in the embedding space. Nevertheless, most of existing models update embeddings only when user takes an interaction with an item. In this paper, we propose TideFC, a novel model based on the state-of-the-art dynamic embedding model JODIE. TideFC can predict the future trajectories of users' and items' embeddings. Besides, we take advantage of t-batch that creates time-consistent batches to make the training stage more efficient. More importantly, we incorporate feature crossing to generate high-order feature interactions in our TideFC. Experiments conducted on multiple real datasets demonstrate the promising performance of TideFC, compared with the state-of-the-art JODIE.