Sequence Separation-Based Modeling of Denoised Implicit Feedback Behavior

This paper analyzes Click-through rate prediction (CTR), a critical component within recommender systems aiming to forecast the personalized probability of user-item click events. Recent advancements have shown that incorporating user behavior sequences into CTR prediction models can yield significant performance improvements. However, CTR prediction models primarily rely on implicit positive feedback, such as clicks, from user-item interactions while overlooking the negative feedback, such as unclicks. Moreover, the implicit feedback obtained from users often contains noisy data, which hampers the accuracy of user interest modeling. As a solution, we propose a novel framework for estimating click-through rates, leveraging the modeling of Denoised Implicit feedback Behavior (DIB). DIB integrates the complete implicit feedback behavior of users into the click-through rate estimation task and aims to mitigate the influence of noise in implicit feedback on the model’s accuracy. Through extensive experiments conducted on real-world, largescale datasets, we demonstrate that DIB outperforms state-of-the-art models by a substantial margin, resulting in an approximate 5% improvement in Area Under the Curve (AUC).