Reproduce “RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Model” 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
YaChen Yan, Liubo Li
Credit Karma researchers developed RankTower, a framework that enhances two-tower pre-ranking models by integrating sophisticated interaction networks and a hybrid training objective with full-stage sampling. This framework consistently achieved superior Recall@K and NDCG@K across Alimama, Taobao, and KuaiRand datasets compared to state-of-the-art baselines.
Key insights
- Complex, bi-directional user-item interactions can be efficiently modeled within a decoupled pre-ranking architecture through gated networks and cross-attention mechanisms.
- Full-stage sampling, which exposes the model to diverse item relevance levels from impression, candidate, and random sets, significantly enhances generalization and robustness by mitigating sample selection bias.
- A hybrid loss function, strategically combining knowledge distillation and various listwise ranking objectives, is crucial for achieving both consistency with downstream ranking models and precise item ordering.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.