Reproduce “Towards Personalized Bundle Creative Generation with Contrastive Non-Autoregressive Decoding” 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
Penghui Wei, Shaoguo Liu, Xuanhua Yang, Liang Wang, Bo Zheng
Researchers at Alibaba Group developed Conna, a model for personalized bundle creative generation that combines items, slogans, and templates, achieving a 2.8x speedup over autoregressive baselines while improving creative quality and diversity. This method leverages contrastive learning and a set-based objective to generate coherent multi-object creatives efficiently for e-commerce advertising.
Key insights
- Non-autoregressive decoding significantly enhances generation efficiency and eliminates the problem of order dependency inherent in sequential generation for multi-object creatives.
- Employing a set-based learning objective, enabled by the Hungarian algorithm, effectively handles the unordered nature of objects within a creative, leading to more accurate model training.
- Integrating a contrastive learning objective with positive and negative creative examples allows the model to learn and optimize for the holistic quality of the entire creative, beyond individual component relevance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.