Abstract: Learning effective vector representations of users and items are the core goal of the recommendation system. Most of the existing recommendation models use deep neural networks or specially designed feature crossing to learn the feature crossing between users and projects to generate user (item) vector representations, but the cross information between user (item) features is not coded into the embedding vector to make full use of the feature cross information, and multiple feature cross information has different effects on the generation of the final user (item) vector representation. Based on this, two graph neural network modules are constructed to learn the intersection information between user (item) features and the feature intersection information between user and item, following the feature cross information weighted by calculating the attention score to obtain the user (item) feature information. Then the original feature information and the feature cross information learned from the network layer are aggregated through the gated recurrent neural network (GRU) to obtain the final user (item) vector expression. Finally, the final recommendation result is obtained through the element product of user vector and item vector. The validity of the model is verified on the datasets MovieLens 1M, Book-Crossing, and Taobao.

Key words: collaborative filtering, graph neural networks, GRU, bilinear characteristic crossover, attention