Reproduce “Generative Click-through Rate Prediction with Applications to Search Advertising” 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
Lingwei Kong, Lu Wang, Changping Peng, Zhangang Lin, Ching Law, Jingping Shao
A two-stage generative framework, GenCTR, enhances Click-Through Rate (CTR) prediction by pre-training a conditional generative model on user behavior sequences for next-item prediction. This method, deployed at JD.com, improved online CTR by up to 1.88% and Revenue Per Mille by up to 2.91% during A/B testing.
Key insights
- Generative pre-training on a next-item prediction task effectively learns rich, contextual user and item representations that significantly enhance downstream discriminative CTR prediction.
- Conditioning the generative model's predictions on item side information and using category-conditioned negative sampling refines item representations, allowing for finer distinctions within relevant categories.
- A two-stage training approach successfully reconciles the disparate data requirements of generative sequence modeling and discriminative click prediction, creating a synergistic framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.