Hybrid Recommendation Model Fused with Compressed Cross Network

Abstract

In recent years, with the successful application of Deep Neural Network (DNN) in various fields, researchers have also proposed several DNN-based recommendation models to learn low-level and high-level feature interactions. Although ordinary DNN has powerful learning capabilities, it lacks vector-level interactive operations and increases the time complexity of calculation. In this paper, a new hybrid recommendation model (DeepFM & CIN, DeepFC) is proposed based on the Compressed Interaction Network (CIN) method, whose purpose is to generate interactive features in a display manner at the vector level. The model can not only perform low-order and high-order point-level interactive operations, but also combines high-order vector-level operations. Through simulation experiments on the Criteo and Avazu datasets and comparison with the baseline model, the experimental results show that the model has improved both the logic loss and the AUC value. Finally, the method of structural pruning is used to further reduce the computational time complexity of the model while maintaining good performance.