Research map / Probabilistic and scientific modeling / Molecules and proteins

Constructing Fair Latent Space for Intersection of Fairness and Explainability

AAAI 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Probabilistic embeddings and relational inference · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes an invertible neural network module attached to a frozen pretrained generative model (DiffAE) that builds a fair latent space separating labels and sensitive attributes, enabling counterfactual explanations of fairness.

Problem

Existing fairness explanation methods explain overall model disparity or output variations rather than internal model behavior, and fair representation learning typically requires training whole generative/VAE models, which is costly and reduces reconstruction quality.

Approach

Train a small Glow-style invertible network on the latent space of a frozen pretrained diffusion autoencoder, using information-bottleneck-inspired losses (covariance diagonalization, variance equalization, and a distance loss pulling together same-label/different-sensitive samples and pushing apart same-sensitive/different-label samples) to disentangle label and sensitive information into separat

Key results

• On CelebA (Y=attractive, S=male), the proposed module reduces EO from 33.4 (DiffAE) to 5.9 and DP from 51.2 to 25.5, while raising WGA from 52.9 to 70.4.
• Training time on CelebAHQ 256 is 2.6 A6000-hours versus 1700.7 for DiffAE, roughly 1/654 of the cost.
• Average accuracy drops with the module (e.g., 78.1 to 75.2 on CelebA attractive), which the authors attribute to dependence on the frozen generator's performance.

Measured effects reported in the paper

Surprising findings

• The label-dimension distance loss without the decomposition step yields little fairness gain, and the authors report that removing the label-information assumption makes the distance loss fail.

Limitations

• Average accuracy depends on the performance of the frozen pretrained generative model and can drop (e.g., CelebA attractive 78.1 to 75.2).
• Evaluated only on facial attribute datasets with binary, gender-related sensitive attributes.
• Gender classification of counterfactual images relies on CLIP zero-shot prompts as a proxy.

Experimental setup

Datasets: CelebA CelebA-HQ UTKFace

Benchmarks: CelebA CelebA-HQ UTKFace

Metrics: Equalized Odds (EO) Demographic Parity (DP) Worst-Group Accuracy (WGA) average accuracy

Baselines: DiffAE SimCLR SupCon FSCL

Builds on: Diffusion Autoencoder Glow information bottleneck FSCL noise-contrastive estimation

Closest papers by meaning

• Fairness without Demographics through Shared Latent Space-Based Debiasing (AAAI 2024 · similarity 0.87)
• Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis (NeurIPS 2025 · similarity 0.87)
• The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.86)
• Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.86)
• Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.85)
• Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.85)