Reproduce “Two-stage dynamic creative optimization under sparse ambiguous samples for e-commerce 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

Guandong Li, Xian Yang

Researchers from Suning's AI Center developed a two-stage dynamic creative optimization framework to improve ad performance by addressing sparse and ambiguous data in e-commerce advertising with automatically synthesized creatives. The system, which integrates AutoCO, Transformer-based reranking, and a bandit model, demonstrated a 10% increase in click-through rate during online A/B testing.

Key insights

  • Decoupling DCO into a two-stage framework provides an effective balance between managing vast creative spaces and performing fine-grained, real-time selection.
  • Knowledge distillation, using a Transformer-based rerank model guided by a teacher rank model, successfully mitigates the impact of sparse and ambiguous training data by learning relative creative orders.
  • Extending AutoCO with a diverse set of interaction operators and a one-shot search algorithm enables more complex and nuanced modeling of relationships between various creative elements.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.