arXiv paper: Using skip and scroll-past data boosts CTR prediction by up to 9.6%

Current click-through rate prediction systems typically build user behavior sequences exclusively from positive interactions like clicks and purchases, ignoring the vast pool of implicit negative signals such as skips, low engagement, and scroll-past events. A new arXiv paper (arXiv:2606.15252) from researchers Zexuan Cheng, Yue Liu, Jun Zhang, and Jie Jiang challenges this convention by demonstrating that mixing both positive and negative behaviors chronologically within a fixed sequence length consistently outperforms positive-only sequences. Their experiments span five diverse model architectures (including target-aware attention and interest evolution networks) and three public benchmarks, yielding relative AUC improvements of +1.9% to +9.6%. The approach adds negligible computational overhead since it only changes input data, not model structure.

The paper also identifies a semantic indistinguishability problem with naive polarity embeddings and proposes Target-Aware Polarity Fusion (TAPF), a lightweight target-conditioned gating mechanism that differentiates positive and negative signals based on the item being predicted. However, even a simpler polarity bias baseline—which treats negative tokens with a learned offset—captures most of the lift, reinforcing that the key innovation is the mixed-polarity data paradigm itself. As gains from longer positive-only sequences plateau, this work offers a practical path to extracting more signal from existing user logs without additional data collection or heavy computation.