Seasonality Aware Topological Graph Neural Network for Recommender Systems

Traditional recommender systems often attempt to model relationships between users and products through techniques such as matrix factorization. However, these methods have limited success in capturing the complex relationships between users and products. Graph Neural Networks (GNN), which have begun to be studied intensively by researchers in recent years, have a remarkable potential in recommendation systems thanks to their power to model complex relationships. In this study, it is aimed to present a new modeling system by using Graph Neural Networks comparatively with traditional recommendation systems on the Movielens dataset. In this regard, the existing data set was enriched in order to create a richer information graph, and the learning transfer method with Bert transformer was used to vectorize the new data obtained. A topological layer was added to capture distant relationships between nodes in the data set. Finally, the seasonality variable was added to the data node, taking into account the effect of the season factor on the choice of watching movies. According to the study results, the proposed Graph Neural Network Method stands out as a strong recommendation system alternative, surpassing other models with an AUC success of over 93%.