RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Model
YaChen Yan, Liubo Li
TL;DR
RankTower addresses the efficiency-accuracy gap in pre-ranking by introducing a three-component architecture that models bi-directional user-item interactions while enabling online efficiency through user-item decoupling. It integrates full-stage sampling with a hybrid loss blending distillation, fine-grained, and coarse-grained ranking signals to align pre-ranking with downstream cascade dynamics. Empirical results on three public datasets show RankTower consistently outperforming state-of-the-art pre-ranking models in Recall@K and NDCG@K, with notable gains attributed to its cross-attention mechanism and stage-aware training. The work offers a practical approach to improving pre-ranking in industrial systems and suggests future directions for end-to-end optimization of cascade ranking.
Abstract
In large-scale ranking systems, cascading architectures have been widely adopted to achieve a balance between efficiency and effectiveness. The pre-ranking module plays a vital role in selecting a subset of candidates for the subsequent ranking module. It is crucial for the pre-ranking model to maintain a balance between efficiency and accuracy to adhere to online latency constraints. In this paper, we propose a novel neural network architecture called RankTower, which is designed to efficiently capture user-item interactions while following the user-item decoupling paradigm to ensure online inference efficiency. The proposed approach employs a hybrid training objective that learns from samples obtained from the full stage of the cascade ranking system, optimizing different objectives for varying sample spaces. This strategy aims to enhance the pre-ranking model's ranking capability and improvement alignment with the existing cascade ranking system. Experimental results conducted on public datasets demonstrate that RankTower significantly outperforms state-of-the-art pre-ranking models.