RankUp: Towards High-rank Representations for Large Scale Advertising Recommender Systems

TL;DR

RankUp introduces a novel architecture to improve representation capacity in large-scale recommender systems, addressing representation collapse and achieving significant GMV improvements in production.

Contribution

The paper proposes RankUp, a new architecture with permutation splitting, multi-embedding, and token integration to enhance expressive capacity in recommender systems.

Findings

RankUp improves GMV by up to 4.81% in production.

Representation capacity scales better with depth using RankUp.

RankMixer's effective rank oscillates and degrades in deep layers, motivating RankUp.

Abstract

The scaling laws for recommender systems have been increasingly validated, where MetaFormer-based architectures consistently benefit from increased model depth, hidden dimensionality, and user behavior sequence length. However, whether representation capacity scales proportionally with parameter growth remains unexplored. Prior studies on RankMixer reveal that the effective rank of token representations exhibits a damped oscillatory trajectory across layers, failing to increase consistently with depth and even degrading in deeper layers. Motivated by this observation, we propose RankUp, an architecture designed to mitigate representation collapse and enhance expressive capacity through randomized permutation splitting over sparse features, a multi-embedding paradigm, global token integration and crossed pretrained embedding tokens. RankUp has been fully deployed in large-scale production across Weixin Video Accounts, Official Accounts and Moments, yielding GMV improvements of 3.41%, 4.81% and 2.12%, respectively.