FairMixRep: Self-supervised Robust Representation Learning for Heterogeneous Data with Fairness constraints

Chakraborty, Souradip; Verma, Ekansh; Sahoo, Saswata; Datta, Jyotishka

doi:10.1109/icdmw51313.2020.00069

Cited by 1 publication

(1 citation statement)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Earlier works in fair representation learning intended to obfuscate any information about sensitive attributes to approximately satisfy demographic parity (Zemel et al, 2013) while a wealth of more recent works focus on using adversarial methods or feature disentanglement in latent spaces of VAEs (Locatello et al, 2019a;Kingma & Welling, 2013;Gretton et al, 2006;Louizos et al, 2015;Amini et al, 2019;Alemi et al, 2018;Burgess et al, 2018;Chen et al, 2018b;Kim & Mnih, 2018;Esmaeili et al, 2019;Song et al, 2019;Gitiaux & Rangwala, 2021;Rodríguez-Gálvez et al, 2020;Sarhan et al, 2020;Paul & Burlina, 2021;Chakraborty et al, 2020). In this setting, the literature has focused on optimizing on approximations of the mutual information between representations and sensitive attributes: maximum mean discrepancy (Gretton et al, 2006) for deterministic or variational (Li et al, 2014;Louizos et al, 2015) autoencoders (VAEs); cross-entropy of an adversarial network that predicts sensitive attributes from the representations (Edwards & Storkey, 2015;Xie et al, 2017;Beutel et al, 2017;Madras et al, 2018;Xu et al, 2018); balanced error rate on both target loss and adversary loss ; Weak-Conditional InfoNCE for conditional contrastive learning (Tsai et al, 2021).…”

Section: Introductionmentioning

confidence: 99%

Is Fairness Only Metric Deep? Evaluating and Addressing Subgroup Gaps in Deep Metric Learning

Dullerud¹,

Roth²,

Hamidieh³

et al. 2022

Preprint

View full text Add to dashboard Cite

Deep metric learning (DML) enables learning with less supervision through its emphasis on the similarity structure of representations. There has been much work on improving generalization of DML in settings like zero-shot retrieval, but little is known about its implications for fairness. In this paper, we are the first to evaluate state-of-the-art DML methods trained on imbalanced data, and to show the negative impact these representations have on minority subgroup performance when used for downstream tasks. In this work, we first define fairness in DML through an analysis of three properties of the representation space -interclass alignment, intra-class alignment, and uniformity -and propose finDML, the f airness in non-balanced DML benchmark to characterize representation fairness. Utilizing finDML, we find bias in DML representations to propagate to common downstream classification tasks. Surprisingly, this bias is propagated even when training data in the downstream task is re-balanced. To address this problem, we present Partial Attribute De-correlation (PARADE) to de-correlate feature representations from sensitive attributes and reduce performance gaps between subgroups in both embedding space and downstream metrics.

show abstract