Reproduce “Click-aware Structure Transfer with Sample Weight Assignment for Post-Click Conversion Rate Estimation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kai Ouyang, Wenhao Zheng, Chen Tang, Xuanji Xiao, Hai-Tao Zheng
Researchers from Tsinghua University and Shopee Inc. introduced Click-aware Structure Transfer with Sample Weight Assignment (CSTWA), a multi-task learning framework designed to improve post-click conversion rate (CVR) prediction. The model addresses the 'curse of knowledge' problem by selectively transferring task-independent structural information from the click-through rate (CTR) task, calibrating CVR representations with click-specific bias, and adaptively re-weighting training samples, leading to superior CVR AUC on industrial and public datasets.
Key insights
- Transferring latent, task-independent structural information (like user/item relationships) from CTR is more effective for CVR than directly sharing raw feature representations, mitigating the 'curse of knowledge'.
- Explicitly modeling 'click bias' and differentiating between clicked and non-clicked samples within the CVR model helps to refine CVR predictions by providing click-specific context.
- Dynamically assigning higher weights to samples where CTR predictions contradict CVR outcomes (e.g., high predicted CTR but no conversion) guides the CVR model to learn true conversion patterns, safeguarding CVR-specific knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.