見送った項目も時系列に加えるクリック率予測：TAPFの仕組みと検証結果

Beyond Positive Signals: Unlocking Implicit Negative Behaviors for Enhanced Sequential User Modeling

User behavior sequence modeling has become a central component in modern click-through rate (CTR) prediction. Over the past years, the community has invested substantial effort into improving how sequences are encoded, from target-aware attention and interest evolution networks to unified architectures that jointly process sequential and non-sequential features. However, a more fundamental question remains under-explored: what should constitute the behavior sequence? Current practice constructs sequences exclusively from positive interactions (clicks, purchases, completions), while the far more abundant implicit negative behaviors (skips, low engagement, scroll-past) are largely underutilized. As gains from longer positive sequences approach diminishing returns, we revisit this underutilized data source within the sequential modeling framework. In this paper, we demonstrate that mixed-polarity behavior sequences, which chronologically interleave positive and negative tokens within a fixed length budget, consistently outperform positive-only sequences across diverse model architectures with negligible additional computational overhead. We further identify a semantic indistinguishability problem inherent to naive polarity embeddings and propose Target-Aware Polarity Fusion (TAPF), a lightweight target-conditioned gating mechanism that provides additional gains by differentiating behavioral evidence. Notably, even the simpler polarity bias baseline captures the majority of improvement, underscoring that the primary contribution is the mixed-polarity data paradigm itself. Experiments on three public benchmarks demonstrate consistent improvements of +1.9% to +9.6% relative AUC across five architectures, which validate the practical value of our approach.