RankTower: A Synergistic Framework for Enhancing Two-Tower Pre-Ranking Model

TL;DR

RankTower introduces a new neural network architecture for pre-ranking in large-scale systems, balancing efficiency and accuracy by capturing user-item interactions and optimizing training objectives, leading to superior performance.

Contribution

The paper presents RankTower, a novel architecture that improves pre-ranking models by efficiently modeling interactions and employing hybrid training objectives for better cascade system integration.

Findings

RankTower outperforms existing pre-ranking models on public datasets.

It effectively balances efficiency and accuracy in large-scale ranking systems.

The hybrid training approach enhances the model's ranking capability.

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.