Weiqiu Wang, Runfeng Zhang, Yifei Liu, Anqi Zhu, Jie Fang, Mingming Pan, Haihong Jiang, Wen Xu, Yipeng Yu, Jiahao Yu, Yijiao Wang, Yeqiu Yang, Fei Sun, Jian Wu, Yuning Jiang, Xu Liu, Chengxiao Feng, B. Zheng · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3840628
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Click-through Rate (CTR) prediction and Conversion Rate (CVR) prediction, as core tasks in modern e-commerce recommender systems, are conventionally formulated as binary classification tasks using click/purchase labels. However, this approach suffers from Intention-Action Divergence (IAD) bias, where users' explicit actions (e.g., click/purchase) misalign with their implicit interest levels inferred from fine-grained interactions (e.g., dwell time, swipe velocity). In this paper, we focus on mitigating the IAD bias by leveraging users' implicit intentions. Specifically, we first mine more than one hundred kinds of user-device interactions in their decision-making process, termed Micro-Behaviors (MBs). Instead of constructing resource-consuming sequential features, we compress MBs into a lightweight sample-level engagement score that scalably quantifies users' intentions for billions of samples. Second, considering the inaccessibility of MBs in online serving, we propose a Pointwise-and-Pairwise Engagement Debiasing (P2ED) approach to mitigate the IAD bias in the CVR task, which utilizes the predefined engagement scores only in model training process. Finally, we extend the P2ED method to the CTR task with a novel Engagement-aware Bidirectional Debiasing (EBD) framework to address the limitation of missing engagement scores in non-click samples. Our method improves CVR and CTR performance, delivering +2.2‰ +1.4‰ AUC gains offline and driving 4.27% GMV growth and 1.38% IPV lift in online A/B tests. In addition, we also release a sampling version of our product dataset to enable future research, the first one containing abundant MB signals for the CTR task, which is available at https://taobao-ebd.github.io/.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.