Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Post-Training Fairness Control: A Single-Train Framework for Dynamic Fairness in Recommendation

Weixin Chen, Li Chen, Yuhan Zhao

TL;DR

Cofair tackles the rigidity of fixed fairness during training by introducing a single-training framework that enables dynamic post-training fairness control in recommender systems. It uses a shared representation layer plus fairness-conditioned adapters to generate level-specific user embeddings, with a user-level regularization ensuring monotonic fairness improvements across levels. An adversarial fairness objective upper-bounds demographic parity, and the framework supports extensions to other fairness notions. Empirical results on Movielens-1M and Last.fm show Cofair achieves favorable fairness-accuracy trade-offs across multiple backbones with no need for retraining, while remaining efficient and adaptable for real-world deployments.

Abstract

Despite growing efforts to mitigate unfairness in recommender systems, existing fairness-aware methods typically fix the fairness requirement at training time and provide limited post-training flexibility. However, in real-world scenarios, diverse stakeholders may demand differing fairness requirements over time, so retraining for different fairness requirements becomes prohibitive. To address this limitation, we propose Cofair, a single-train framework that enables post-training fairness control in recommendation. Specifically, Cofair introduces a shared representation layer with fairness-conditioned adapter modules to produce user embeddings specialized for varied fairness levels, along with a user-level regularization term that guarantees user-wise monotonic fairness improvements across these levels. We theoretically establish that the adversarial objective of Cofair upper bounds demographic parity and the regularization term enforces progressive fairness at user level. Comprehensive experiments on multiple datasets and backbone models demonstrate that our framework provides dynamic fairness at different levels, delivering comparable or better fairness-accuracy curves than state-of-the-art baselines, without the need to retrain for each new fairness requirement. Our code is publicly available at https://github.com/weixinchen98/Cofair.

Post-Training Fairness Control: A Single-Train Framework for Dynamic Fairness in Recommendation

TL;DR

Cofair tackles the rigidity of fixed fairness during training by introducing a single-training framework that enables dynamic post-training fairness control in recommender systems. It uses a shared representation layer plus fairness-conditioned adapters to generate level-specific user embeddings, with a user-level regularization ensuring monotonic fairness improvements across levels. An adversarial fairness objective upper-bounds demographic parity, and the framework supports extensions to other fairness notions. Empirical results on Movielens-1M and Last.fm show Cofair achieves favorable fairness-accuracy trade-offs across multiple backbones with no need for retraining, while remaining efficient and adaptable for real-world deployments.

Abstract

Despite growing efforts to mitigate unfairness in recommender systems, existing fairness-aware methods typically fix the fairness requirement at training time and provide limited post-training flexibility. However, in real-world scenarios, diverse stakeholders may demand differing fairness requirements over time, so retraining for different fairness requirements becomes prohibitive. To address this limitation, we propose Cofair, a single-train framework that enables post-training fairness control in recommendation. Specifically, Cofair introduces a shared representation layer with fairness-conditioned adapter modules to produce user embeddings specialized for varied fairness levels, along with a user-level regularization term that guarantees user-wise monotonic fairness improvements across these levels. We theoretically establish that the adversarial objective of Cofair upper bounds demographic parity and the regularization term enforces progressive fairness at user level. Comprehensive experiments on multiple datasets and backbone models demonstrate that our framework provides dynamic fairness at different levels, delivering comparable or better fairness-accuracy curves than state-of-the-art baselines, without the need to retrain for each new fairness requirement. Our code is publicly available at https://github.com/weixinchen98/Cofair.
Paper Structure (36 sections, 2 theorems, 19 equations, 3 figures, 3 tables)

This paper contains 36 sections, 2 theorems, 19 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Recommendation Task
  4. Fairness Task
  5. Methodology
  6. Model Architecture
  7. Shared Representation Layer
  8. Fairness-Conditioned Adapters
  9. Output Layer
  10. Loss Functions
  11. Recommendation Loss
  12. Fairness Loss via Adversarial Network
  13. User-Level Regularization
  14. Adaptive Weighting of the Fairness Loss
  15. Overall Objective
  16. ...and 21 more sections

Key Result

Lemma 1

Consider any measurable function $G : \mathbf{e}_u^{(t)} \to [0,1]$ representing the predicted preference. Then, at fairness level $t$, the demographic parity difference $\Delta_{\text{DP}}^{(t)}$ is upper bounded by the optimal value of adversarial fairness objective:

Figures (3)

  • Figure 1: Fairness-accuracy curves, where points closer to the left (indicating greater fairness) and the top (indicating higher accuracy), are more Pareto efficient. Notably, the results for baselines are obtained after multiple training, while the results for our Cofair are obtained after single training and multiple forward pass by indicating different values of $t$.
  • Figure 2: Means and variances of DP ($\mu_{DP}$ and $\sigma^2_{DP}$) and NDCG ($\mu_{NDCG}$ and $\sigma^2_{NDCG}$) on BPR backbone and Movielens-1M dataset with different values of $\lambda_0$, $\eta$, and $\beta$, respectively.
  • Figure 3: Comparison of four fairness methods without and with our controllable fairness framework (indicated by “+ Cofair”), using BPR as backbone on MovieLens-1M.

Theorems & Definitions (4)

  • Lemma 1
  • proof
  • Theorem 1
  • proof