A Novel Efficient Unclick Behavior Modeling Framework for Click-Through Rate Prediction

Abstract

Click-Through Rate (CTR) prediction is a crucial task in many applications, such as recommendation systems. It predicts a personalized click probability for each user-item pair. Recently, researchers have found that the performance of CTR prediction model can be significantly improved by considering user behavior sequence. However, most existing CTR prediction models take only implicit positive feedback (e.g., click) from user-item interactions as input. They do not consider the implicit negative feedback (e.g., unclick), resulting in incomprehensive user representations and sub-optimal model performance since they ignore the complete user behavior data. Furthermore, un click behavior usually contains heavy noise and greatly in-terferes with the accuracy of user interest modeling. To address these issues, we propose an efficient Unclick Behavior Modeling framework (UBM) to model the implicit negative feedback based on the click behavior modeling to learn users' complete and unbiased preferences for CTR prediction. Specifically, UBM utilizes a novel implicit negative feedback modeling to enrich user interest representations. Then, the Interest Denoising Layer and Explicit Interest Contrastive Layer are proposed, which use the representation of explicit positive feedback to distill accurate user interests from implicit negative feedback. Based on that, an Interest Extraction Layer is designed to effectively model and fuse with different types of user interest representations. Extensive experiments on real-world and large-scale datasets demonstrate that UBM significantly outperforms the state-of-the-art models, bringing 8.05% AUC lift.