Reproduce “UniCon: A Unified Context-Centric Modeling Paradigm for CTR Prediction” 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
Jiajun Cui, Zhengqi Xu, Fan Zhang, Zhangteng, Gu Tang, Honghong Zhu, Mengxi Wu, Yulin Liang and 1 more
UniCon models click-through rate by representing each display event as a context containing jointly shown items, intent, and environmental signals. It applies attention within contexts and across historical and target contexts, with target-aware compression for long histories. On Meituan search advertising, UniCon-Large reaches 0.8697 AUC versus 0.8558 for production Base; a seven-day online test reported gains with latency within SLA.
Key insights
- Unifying historical actions and current prediction targets as homogeneous 'context units' aligns more closely with the user's decision-making process in context-rich environments.
- Explicitly modeling both intra-context item relationships (locality) and inter-context temporal dynamics is crucial for capturing nuanced user preferences and interactions.
- Context-aware sequence compression and variable-length attention are effective for industrial-scale deployment, maintaining prediction quality while significantly improving serving efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.