Alibaba's Three Papers Accepted at NeurIPS 2022

Alibaba's technical team has three papers accepted at the 36th Conference on Neural Information Processing Systems (NeurIPS 2022), a top international conference in machine learning and computational neuroscience organized by the NeurIPS Foundation. The conference will be held from November 28 to December 9.

The three accepted papers are:

1. APG: Adaptive Parameter Generation Network for Click-Through Rate Prediction - Proposes an efficient and effective module called Adaptive Parameter Generation Network (APG) that dynamically generates different model parameters for CTR models based on different samples. APG has been deployed in Alibaba's search advertising system, achieving 3% CTR improvement and 1% revenue increase.

2. GBA: A Tuning-free Approach to Switch between Synchronous and Asynchronous Training for Recommendation Models - Proposes a Global Batch gradients Aggregation (GBA) scheme that enables seamless switching between synchronous and asynchronous training modes without hyperparameter tuning. GBA achieves 2.4x training speedup in resource-constrained clusters while maintaining convergence properties.

3. Sustainable Online Reinforcement Learning for Auto-bidding - Proposes a Sustainable Online Reinforcement Learning (SORL) framework that directly trains auto-bidding strategies through online advertising system interaction, addressing the offline-online inconsistency problem. SORL includes safe exploration algorithms and V-CQL training methods, outperforming existing auto-bidding algorithms in both simulation and online experiments.