Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FairSync: Ensuring Amortized Group Exposure in Distributed Recommendation Retrieval

Chen Xu, Jun Xu, Yiming Ding, Xiao Zhang, Qi Qi

TL;DR

The paper tackles the problem of ensuring amortized group max-min fairness in the retrieval stage of distributed recommender systems, a prerequisite for reliable downstream exposure in stage-2. It introduces FairSync, a dual-space distributed optimization framework in which a central vector $\boldsymbol{\mu}$ aggregates historical fairness signals and guides per-server retrieval through dual-space queries, updated periodically by gradient-based optimization. The approach builds on a distributed dense retrieval architecture and derives a low-dimensional dual formulation $W^{Dual}$, enabling scalable, online, and distributive enforcement of group exposure constraints $e_g \ge m_g$ over horizon $T$. Empirical results on two public datasets show that FairSync achieves the target exposure levels (ESP) while preserving high retrieval accuracy (Recall, NDCG, HR) across multiple base models, outperforming baseline fairness methods that either fail the exposure constraint or degrade accuracy. The work demonstrates the practical viability of incorporating amortized fairness directly into the retrieval stage and offers a scalable path for fair and efficient distributed recommender systems.

Abstract

In pursuit of fairness and balanced development, recommender systems (RS) often prioritize group fairness, ensuring that specific groups maintain a minimum level of exposure over a given period. For example, RS platforms aim to ensure adequate exposure for new providers or specific categories of items according to their needs. Modern industry RS usually adopts a two-stage pipeline: stage-1 (retrieval stage) retrieves hundreds of candidates from millions of items distributed across various servers, and stage-2 (ranking stage) focuses on presenting a small-size but accurate selection from items chosen in stage-1. Existing efforts for ensuring amortized group exposures focus on stage-2, however, stage-1 is also critical for the task. Without a high-quality set of candidates, the stage-2 ranker cannot ensure the required exposure of groups. Previous fairness-aware works designed for stage-2 typically require accessing and traversing all items. In stage-1, however, millions of items are distributively stored in servers, making it infeasible to traverse all of them. How to ensure group exposures in the distributed retrieval process is a challenging question. To address this issue, we introduce a model named FairSync, which transforms the problem into a constrained distributed optimization problem. Specifically, FairSync resolves the issue by moving it to the dual space, where a central node aggregates historical fairness data into a vector and distributes it to all servers. To trade off the efficiency and accuracy, the gradient descent technique is used to periodically update the parameter of the dual vector. The experiment results on two public recommender retrieval datasets showcased that FairSync outperformed all the baselines, achieving the desired minimum level of exposures while maintaining a high level of retrieval accuracy.

FairSync: Ensuring Amortized Group Exposure in Distributed Recommendation Retrieval

TL;DR

The paper tackles the problem of ensuring amortized group max-min fairness in the retrieval stage of distributed recommender systems, a prerequisite for reliable downstream exposure in stage-2. It introduces FairSync, a dual-space distributed optimization framework in which a central vector aggregates historical fairness signals and guides per-server retrieval through dual-space queries, updated periodically by gradient-based optimization. The approach builds on a distributed dense retrieval architecture and derives a low-dimensional dual formulation , enabling scalable, online, and distributive enforcement of group exposure constraints over horizon . Empirical results on two public datasets show that FairSync achieves the target exposure levels (ESP) while preserving high retrieval accuracy (Recall, NDCG, HR) across multiple base models, outperforming baseline fairness methods that either fail the exposure constraint or degrade accuracy. The work demonstrates the practical viability of incorporating amortized fairness directly into the retrieval stage and offers a scalable path for fair and efficient distributed recommender systems.

Abstract

In pursuit of fairness and balanced development, recommender systems (RS) often prioritize group fairness, ensuring that specific groups maintain a minimum level of exposure over a given period. For example, RS platforms aim to ensure adequate exposure for new providers or specific categories of items according to their needs. Modern industry RS usually adopts a two-stage pipeline: stage-1 (retrieval stage) retrieves hundreds of candidates from millions of items distributed across various servers, and stage-2 (ranking stage) focuses on presenting a small-size but accurate selection from items chosen in stage-1. Existing efforts for ensuring amortized group exposures focus on stage-2, however, stage-1 is also critical for the task. Without a high-quality set of candidates, the stage-2 ranker cannot ensure the required exposure of groups. Previous fairness-aware works designed for stage-2 typically require accessing and traversing all items. In stage-1, however, millions of items are distributively stored in servers, making it infeasible to traverse all of them. How to ensure group exposures in the distributed retrieval process is a challenging question. To address this issue, we introduce a model named FairSync, which transforms the problem into a constrained distributed optimization problem. Specifically, FairSync resolves the issue by moving it to the dual space, where a central node aggregates historical fairness data into a vector and distributes it to all servers. To trade off the efficiency and accuracy, the gradient descent technique is used to periodically update the parameter of the dual vector. The experiment results on two public recommender retrieval datasets showcased that FairSync outperformed all the baselines, achieving the desired minimum level of exposures while maintaining a high level of retrieval accuracy.
Paper Structure (28 sections, 2 theorems, 8 equations, 6 figures, 4 tables, 2 algorithms)

This paper contains 28 sections, 2 theorems, 8 equations, 6 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Problem Formulation
  4. Our approach
  5. Distributed Dense Retrieval Architecture
  6. Dual Space of Retrieval
  7. FairSync Algorithm
  8. Inference phase
  9. Online learning phase
  10. Discussion
  11. Experiment
  12. Experimental settings
  13. Datasets
  14. Evaluation
  15. Baselines and Base Models
  16. ...and 13 more sections

Key Result

Theorem 1

The dual problem objective $W^{\text{Dual}}$ of Equation eq:retrieval can be write as where we can have a dual variable $\boldsymbol{\mu} \in \mathbb{R}^{|\mathcal{G}|}$, $\mathbf{A}\in\mathbb{R}^{|\mathcal{I}|\times|\mathcal{G}|}$ is the item-group adjacent matrix, and $A_{ig} = 1$ indicates item $i\in \mathcal{I}_g$, and 0 otherwise. Moreover, the dual problem is a strong-dual prob

Figures (6)

  • Figure 1: (a) The two-stage pipelines of recommender system, including retrieval (stage-1) and ranking (stage-2). (b) Simulations depicting the changes for the minimum exposures across two stages.
  • Figure 2: FairSync Framework. Sub-figure (a) illustrates an example to show the intuitive example of how FairSync works. Sub-figure (b) illustrates that the online retrieval process when user $u_t$ arrives, while sub-figure (c) depicts the offline item embeddings in the dual space construction process.
  • Figure 3: The accuracy curve (Recall, NDCG, and HR) of FairSync (ours) and the best baseline Uncalibrated under different minimum exposure threshold $m_g$. The experiments were conducted based on the best retrieval base model ComiRec-DR.
  • Figure 4: The three sub-figures in the first row illustrate the t-SNE visualization item embeddings and user embeddings of Comirec-DR and our model FairSync using Comirec-DR as the base model under different time $t$. The three sub-figures in the final row depict the category exposures under different time steps $t$. The experiment was conducted on the Amazon-book dataset with retrieval number $K=50$.
  • Figure 5: Inference time per user w.r.t. w.r.t. batch size $B$.
  • ...and 1 more figures

Theorems & Definitions (4)

  • Theorem 1
  • Remark 1: Distributed solution in dual space
  • Remark 2: Small computational costs
  • Lemma 1