A Dual Augmented Two-tower Model for Online Large-scale Recommendation

Sep 17, 2026

Uploaded by Dipper

Many modern recommender systems rely on two-tower architectures to perform large-scale retrieval by learning query and item representations from content features. However, the two-tower design lacks direct information interaction between towers, and imbalanced item categories can degrade performance on underrepresented classes. This paper proposes the Dual Augmented Two-tower Model (DAT), integrating an Adaptive-Mimic Mechanism to build augmented vectors and a Category Alignment Loss to better align uneven category representations. Offline experiments and online A/B tests validate improved recommendation quality.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Abstract

1 Introduction

A Dual Augmented Two-tower Model for Online Large-scale Recommendation Yantao Yu, Weipeng Wang, Zhoutian Feng, Daiyue Xue Meituan Beijing, China {yuyantao,wangweipeng02,fengzhoutian,[xuedaiyue}@meituan.com](xuedaiyue}@meituan.com) ABSTRACT Many modern recommender systems have a very large corpus, anda common industrial recipe for handling large-scale retrieval is to learn query and item representations from their content features with the two-tower model. However, the model suffers from lack of information interaction between the two towers. Besides, imbalanced category data also hinders the model performance. In this paper, we propose a novel model named Dual Augmented Two-tower Model (DAT), which integrates a novel Adaptive-Mimic Mechanism (AMM) and a Category Alignment Loss (CAL) . Our AMM customizes an augmented vector for each query and item to mitigate the lack of information interaction. Moreover, our CAL can further improve performance by aligning item representation of uneven categories. Offline experiments on large-scale datasets are conducted to show the superior performance of DAT. Moreover, online A/B testings confirm that DAT can lead to improved recommendation quality for industrial applications. CCS CONCEPTS • Information systems → Recommender systems; Information retrieval. KEYWORDS Recommender Systems, Information Retrieval, Neural Networks ACM Reference Format: Yantao Yu, Weipeng Wang, Zhoutian Feng, Daiyue Xue. 2021. A Dual Augmented Two-tower Model for Online Large-scale Recommendation. In Proceedings of DLP-KDD 2021. ACM, New York, NY, USA, 4 pages. [https:](https:)//[doi.org/10.1145/1122445.1122456](doi.org/10.1145/1122445.1122456) 1 INTRODUCTION Recommender systems are indispensable in filtering out items that users are interested in from huge amounts of items. One of the most crucial challenges in large-scale recommendation is to score millions or billions of items in real-time accurately. A common practice is to design the recommender as a two-phase architecture where a retrieval model first retrieves a small fraction of related items given user’s query from a large corpus, and a ranking model Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions [from permissions@acm.org](from permissions@acm.org). DLP-KDD 2021, August 15, 2021, Singapore © 2021 Association for Computing Machinery. ACM ISBN 978-1-4503-XXXX-X/18/06. . . $15.00 [https://doi.org/10.1145/1122445.1122456](https://doi.org/10.1145/1122445.1122456) ranks the retrieved items based on clicks or user-ratings [2] . Obviously, the quality of candidates retrieved in the retrieval stage plays a critical role in the whole system. In this work, we focus on the retrieval problem with an aim to improve the retrieval system’s performance for personalized recommendation with millions of queries and items. A scalable retrieval model usually learns query and item representations first and then uses a cosine similarity between the query and item representations to obtain recommendations tailored for the query. However, in industrial-scale applications, the corpus of items can be enormously large and the training data collected from users’ feedback is likely to be very sparse for most queries and items, which may lead to inaccurate model predictions for long-tail users and items. To tackle the above challenges, the two-tower model [4], with towers referring to encoders based on deep neural networks (DNNs), are usually applied. Despi

FAQs

What limitation of the two-tower model does the paper address?

The model lacks information interaction between the query tower and item tower, because the item representation must be pre-computed independently for online retrieval.

How does DAT improve interaction between the two towers?

DAT introduces an Adaptive-Mimic Mechanism that customizes augmented vectors for queries and items, updated based on the other tower’s representation for positive-labeled samples.

How does DAT handle the problem of imbalanced item categories?

DAT adds a Category Alignment Loss during training to align item representations across uneven categories, improving performance for smaller categories.

Sep 17, 2026

Uploaded by Dipper