(AIMS 2024) Incorporating Feature Refinement Enhancement and Cross Network for Click-Through Rate Prediction

Click-through rate prediction is an important task in personalized advertising and recommender systems. Currently, many approaches model feature interactions to improve their performance. DeepFM as a classical approach takes care of both high-order and low-order feature interactions in a parallel way, but it has some limitations. Firstly, it ignores the fact that the importance of the same feature may be different in different contexts, and secondly, it has a limited number of orders of explicit feature interactions and the importance of feature interactions is not differentiated. To address these issues, we improve DeepFM and propose FRGCN, a click rate prediction model that combines feature refinement enhancement with cross network; specifically, we add an attention-based feature refinement enhancement layer FRNet-A after the embedding layer to achieve context-aware representation of features and attention enhancement. In addition, we introduce gated cross network instead of FM modules in the feature interaction layer to capture higher-order explicit feature interactions. Comprehensive experimental results on Frappe and Malware datasets demonstrate the effectiveness and superiority of FRGCN.