Reproduce “Pantheon: Personalized Multi-objective Ensemble Sort via Iterative Pareto Policy Optimization” 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
Jiangxia Cao, Pengbo Xu, Yin Cheng, Kaiwei Guo, Jian Tang, Shijun Wang, Dewei Leng, Shuang Yang and 4 more
A neural-network-based ensemble sort mechanism, Pantheon, was developed by Kuaishou Technology to replace traditional formula-based methods in industrial recommender systems. This approach leverages iterative Pareto policy optimization and fine-grained representation inheritance, achieving an average +1.62% GAUC improvement offline and up to +1.671% gain in clicked users during live A/B tests on a platform serving 400 million daily users.
Key insights
- The Iterative Pareto Policy Optimization (IPPO) mechanism effectively automates multi-objective balancing, outperforming manual formula-based approaches even with simpler neural architectures.
- Leveraging high-dimensional hidden-state representations from the ranking model significantly enhances the ensemble sort's accuracy by providing richer, fine-grained information.
- More complex architectures like Transformers further improve ensemble sort performance, indicating potential for deeper neural integration into recommendation pipelines.
- Pantheon led to a positive ecological shift, reallocating traffic to increase overall effective usage and significantly improving diverse user behaviors like long-view and interaction rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.