Reproduce “Beyond Semantics: Learning a Behavior Augmented Relevance Model with Self-supervised Learning” 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
Zeyuan Chen, Wei Chen, Jia Xu, Zhongyi Liu, Wei Zhang
This paper from Ant Group introduces BARL-ASe, a behavior-augmented relevance model that integrates dual behavior neighbors with multi-level co-attention and self-supervised learning for search. The model achieved the highest AUC (0.9078) on an Alipay dataset and, in A/B testing, improved click-through rate by 0.72% while reducing latency by 10ms for Alipay's mini-apps search.
Key insights
- Pure semantic matching often falls short in accurately discerning user search intent due to short, ambiguous queries; user behavior data provides crucial disambiguation.
- Dual behavior neighbors (query-to-item and item-to-query interactions) provide a richer context than single-sided or purely graph-based behavioral approaches.
- Self-supervised learning, specifically contrastive and mutual learning between semantic and behavior views, effectively mitigates noise in real-world user behavior data and strengthens representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.