Reproduce “COLD: Towards the Next Generation of Pre-Ranking System” 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
Zhe Wang, Liqin Zhao, Biye Jiang, Guorui Zhou, Xiaoqiang Zhu, Kun Gai
COLD reframes pre-ranking as co-design: it uses a richer model with user, ad, and user–ad interaction features, while tuning feature selection and serving computations to manage cost. In Alibaba’s live A/B test against a vector-product model, it reported revenue-per-thousand-impressions lifts of 6.5% on normal days and 10.8% during Double 11. This shows a tunable tradeoff, not lower serving cost.
Key insights
- An algorithm-system co-design approach can effectively overcome the traditional trade-off between model complexity and computational efficiency in pre-ranking, enabling more expressive models within latency budgets.
- Combining flexible deep architectures that use cross-features with aggressive engineering optimizations (e.g., Float16, MPS, column-based computation) is crucial for deploying sophisticated models at high throughput and low latency.
- A fully online learning and serving infrastructure is essential for continuous model adaptation to real-time data shifts, leading to improved performance during dynamic periods and faster development cycles.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.