Adaptive Fair Representation Learning for Personalized Fairness in Recommendations via Information Alignment

TL;DR

This work tackles personalized fairness in recommendations by addressing the explosion of attribute combinations and the fairness–accuracy trade-off. It introduces Adaptive Fair Representation Learning (AFRL), a single, inference-adaptive model that generates a personalized fair embedding $^*$ from an unfair embedding $$ and a user-specific fairness vector $ s_u$. AFRL achieves this through two core components: Information Alignment Module (IAlignM), which constructs attribute-specific embeddings $_i$ and a debiased collaborative embedding $_0$ using a bilevel information bottleneck and adversarial training, and Information Aggregation Module (IAggM), which fuses these embeddings according to $ s_u$ to produce $^*$ for downstream prediction. Theoretical results prove convergence and informativeness guarantees, while experiments on ML-1M and Taobao show that AFRL improves fairness with minimal loss in accuracy, outperforming state-of-the-art personalized fairness methods. Overall, AFRL offers a scalable, adaptable approach to personalized fairness in recommender systems with strong empirical and theoretical support.

Abstract

Personalized fairness in recommendations has been attracting increasing attention from researchers. The existing works often treat a fairness requirement, represented as a collection of sensitive attributes, as a hyper-parameter, and pursue extreme fairness by completely removing information of sensitive attributes from the learned fair embedding, which suffer from two challenges: huge training cost incurred by the explosion of attribute combinations, and the suboptimal trade-off between fairness and accuracy. In this paper, we propose a novel Adaptive Fair Representation Learning (AFRL) model, which achieves a real personalized fairness due to its advantage of training only one model to adaptively serve different fairness requirements during inference phase. Particularly, AFRL treats fairness requirements as inputs and can learn an attribute-specific embedding for each attribute from the unfair user embedding, which endows AFRL with the adaptability during inference phase to determine the non-sensitive attributes under the guidance of the user's unique fairness requirement. To achieve a better trade-off between fairness and accuracy in recommendations, AFRL conducts a novel Information Alignment to exactly preserve discriminative information of non-sensitive attributes and incorporate a debiased collaborative embedding into the fair embedding to capture attribute-independent collaborative signals, without loss of fairness. Finally, the extensive experiments conducted on real datasets together with the sound theoretical analysis demonstrate the superiority of AFRL.

Figures (6)

• Figure 1: The overview of AFRL. AFRL contains Information Alignment Module (IAlignM) and Information Aggregation Module (IAggM). IAlignM generates $M$ attribute-specific embeddings $\{\mathbf{z}_{u,i}\}$ ($1 \le i \le M$) and a debiased collaborative embedding $\mathbf{z}_{u,0}$ from the user embedding $\mathbf{u}$, using attribute encoders $\{E_{i}\}$ ($1 \le i \le M$) and a debiased collaborative signal encoder $F$, respectively. IAggM aggregates these embeddings with respect to the personalized fairness requirement $\mathbf{s}_u$ to produce the final fair user embedding $\mathbf{u}^{*}$ for downstream recommendation.
• Figure 2: Illustration of the attribute-specific embedding learning for Information Alignment, where (a) and (b) are the cases of information misalignment, while (c) is the optimal case of information alignment.
• Figure 3: Pareto front curves under different fairness requirements with SASRec kang2018self as base model. In each curve, the points from left to right correspond to $\boldsymbol{\lambda}\in\{100,10,1,0.1,0.01,0.001\}$, respectively.
• Figure 4: Pareto front curves under different fairness requirements with BERT4Rec sun2019bert4rec as base model. In each curve, the points from left to right correspond to $\boldsymbol{\lambda}\in\{100,10,1,0.1,0.01,0.001\}$, respectively.
• Figure 5: Contributions to accuracy by $\mathbf{z}_{u, 0}$ and $\mathbf{z}_{u, i}$.

Theorems & Definitions (1)

• Definition 1: Counterfactually fair recommendation li2021towards