Reproduce “ESANS: Effective and Semantic-Aware Negative Sampling for Large-Scale Retrieval Systems” 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
Haibo Xing, Kanefumi Matsuyama, Hao Deng, Jinxin Hu, Yu Zhang, Xiaoyi Zeng
ESANS introduces an effective negative sampling framework for large-scale retrieval systems, combining multimodal semantic-aware clustering with a dense interpolation strategy to generate high-quality, semantically consistent negative samples. The approach improves item retrieval accuracy in offline experiments and demonstrates significant gains in advertising revenue, click-through rate, and gross merchandise volume during online A/B testing on Alibaba's e-commerce platform.
Key insights
- Integrating multimodal semantic understanding via hierarchical clustering effectively reduces false negatives and provides a structured space for selecting diverse and challenging negative samples.
- Dense interpolation in the low-dimensional embedding space offers a computationally efficient method to augment the quantity and diversity of negative samples, enhancing the learning signal without significant overhead.
- Controllably generating hard negatives through interpolation (tuning the \u03bb parameter) helps define a more precise decision boundary for the model, leading to improved discriminative power.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.