CFF: combining interactive features and user interest features for click-through rate prediction

Lin Zhang, Fang'ai Liu, Hongchen Wu, Xuqiang Zhuang, Yaoyao Yan

Abstract: Click-through rate is a central issue in ad recommendation and has recently received extensive research attention in academia and industry. Research shows that the accuracy of prediction results in CTR prediction is closely related to interactive features and user interest features. However, existing models usually focus on one aspect of features, i.e., interactive features or interest features, and few studies have attempted to learn both interactive features and interest features simultaneously. In this paper, a novel model called CFF as an abbreviation for Combining interactive Features and interest Features is proposed to learn interactive features and user interest features simultaneously. To efficiently learn fine-grained interactive features, an attention-based squeeze equal interaction network (ASENet) is constructed to select salient feature information at the level of equal interactive features. A bi-directional attention-target item gated recurrent unit (Bi-ATGRU) is designed to learn the dependencies between user interests and items. Specifically, it refines and integrates interest features by incorporating context information, historical behaviors, and target item. Extensive experiments on four public datasets indicate CFF outperforms other baselines in terms of evaluation metrics (the Logloss decreases by 1.97% on Frappe and 1.85% on MovieLens).