FEED: Fairness-Enhanced Meta-Learning for Domain Generalization

TL;DR

This paper introduces Fairness-Enhanced Meta-Learning for Domain Generalization (FEED), a novel framework that disentangles latent data representations into content, style, and sensitive vectors and embeds a fairnessaware invariance criterion within the meta-learning process, ensuring consistently fair learned parameters across domains with varied characteristics.

Abstract

Generalizing to out-of-distribution data while being aware of model fairness is a significant and challenging problem in meta-learning. The goal of this problem is to find a set of fairness-aware invariant parameters of classifier that is trained using data drawn from a family of related training domains with distribution shift on non-sensitive features as well as different levels of dependence between model predictions and sensitive features so that the classifier can achieve good generalization performance on unknown but distinct test domains. To tackle this challenge, existing state-of-the-art methods either address the domain generalization problem but completely ignore learning with fairness or solely specify shifted domains with various fairness levels. This paper introduces an approach to fairness-aware meta-learning that significantly enhances domain generalization capabilities. Our framework, Fairness-Enhanced Meta-Learning for Domain Generalization (FEED), disentangles latent data representations into content, style, and sensitive vectors. This disentanglement facilitates the robust generalization of machine learning models across diverse domains while adhering to fairness constraints. Unlike traditional methods that focus primarily on domain invariance or sensitivity to shifts, our model integrates a fairness-aware invariance criterion directly into the meta-learning process. This integration ensures that the learned parameters uphold fairness consistently, even when domain characteristics vary widely. We validate our approach through extensive experiments across multiple benchmarks, demonstrating not only superior performance in maintaining high accuracy and fairness but also significant improvements over existing state-of-the-art methods in domain generalization tasks.

Figures (4)

• Figure 1: Illustration of fairness-aware domain generalization problems using the ccMNIST digit dataset. The domains correspond to different digit colors (red/green/blue). Each image has a black or white background color as the sensitive label. For simplicity, digits 2 and 9 are used as toy examples to demonstrate the setting. Each domain is associated with various group fairness levels estimated using the demographic parity metric.
• Figure 2: Causal interpretation of fairness-aware domain generalization tasks. We assume that the raw features ($\mathbf{x}$) and class label ($y$) of each example are generated by the latent content factor ($\mathbf{c}$), style factor ($\mathbf{s}$), and sensitive factor ($\mathbf{a}$). The sensitive factor ($\mathbf{a}$) is dependent on the sensitive attribute ($z$) of this example and may or may not be dependent on the domain. The style factor $\mathbf{s}$ depends on the domain, but the content factor $\mathbf{c}$ is independent of the domain $e$. Each domain label is unobserved.
• Figure 3: (Top) An overview of our framework. The red lines and the blue lines correspond to outer loop and inner loop respectively. (Bottom) The transformation model $T$. It generates an augmented example having the same content factor as the input example but has different style and sensitive factors sampled from their associated distributions that encode a new synthetic domain.
• Figure 4: Ablation study on four datasets. Results are plotted as averages across all domains.

Theorems & Definitions (1)

• Definition 1: Fairness-aware $T$-Invariance