Evaluating Fairness in Self-supervised and Supervised Models for Sequential Data

TL;DR

This paper investigates how self-supervised learning (SSL) compares to supervised learning in terms of fairness and performance on sequential data, highlighting SSL's potential to improve fairness with minimal performance loss.

Contribution

It systematically compares SSL and supervised models on fairness and performance, demonstrating SSL's ability to significantly enhance fairness in human-centric applications.

Findings

SSL achieves up to 27% increase in fairness

SSL maintains performance comparable to supervised models

SSL is effective in data-scarce, high-stakes domains like healthcare

Abstract

Self-supervised learning (SSL) has become the de facto training paradigm of large models where pre-training is followed by supervised fine-tuning using domain-specific data and labels. Hypothesizing that SSL models would learn more generic, hence less biased, representations, this study explores the impact of pre-training and fine-tuning strategies on fairness (i.e., performing equally on different demographic breakdowns). Motivated by human-centric applications on real-world timeseries data, we interpret inductive biases on the model, layer, and metric levels by systematically comparing SSL models to their supervised counterparts. Our findings demonstrate that SSL has the capacity to achieve performance on par with supervised methods while significantly enhancing fairness--exhibiting up to a 27% increase in fairness with a mere 1% loss in performance through self-supervision. Ultimately, this work underscores SSL's potential in human-centric computing, particularly high-stakes, data-scarce application domains like healthcare.