Future-aware and High-order representation learning for cold-start recommendation

The Cold-start problem is critical but challenging for dynamic recommender systems since new entities (users/items) are added dynamically without any purchasing behavior. Most existing methods solve the problem by building the relationship between cold-start entities and existing entities. However, due to several challenges, such approaches can not effectively handle the cold start in dynamic recommender systems. It is hard to learn and predict dynamically for constantly added entities without any historical interactions. Moreover, it is challenging to characterize cold-start entities precisely for indicating future purchasing with limited information. This paper formalizes the dynamic recommender systems as a time-evolving graph to handle the challenges of modeling the dynamic relations between users and items. Mainly, we design a unified learning framework that can learn future-aware representations for newly added entities. Additionally, we build a novel mapping function to model high-order interactions between attributes and further convey obtained expressive information to the high-order neighbors on the graph. Extensive experiments were conducted, and the results demonstrate the outstanding performance of the proposed method on MovieLens 1M and LastFM datasets, providing at least 16.69% and 11.11% relative performance gains, respectively.