[Paper Review] The Rich Get Richer: Disparate Impact of Semi-Supervised Learning

Zhaowei Zhu, Tianyi Luo+1 more|arXiv (Cornell University)|Oct 12, 2021

Domain Adaptation and Few-Shot Learning43 references8 citations

TL;DR

This paper reveals that semi-supervised learning (SSL) can exacerbate fairness disparities, where sub-populations with higher baseline accuracy ('rich') benefit more from SSL, while those with lower baseline accuracy ('poor') may see performance degradation. The authors propose the benefit ratio metric to quantify this disparate impact and demonstrate through theory and experiments that data balancing and increased labeled data can mitigate these inequities across image and text classification tasks.

ABSTRACT

Semi-supervised learning (SSL) has demonstrated its potential to improve the model accuracy for a variety of learning tasks when the high-quality supervised data is severely limited. Although it is often established that the average accuracy for the entire population of data is improved, it is unclear how SSL fares with different sub-populations. Understanding the above question has substantial fairness implications when different sub-populations are defined by the demographic groups that we aim to treat fairly. In this paper, we reveal the disparate impacts of deploying SSL: the sub-population who has a higher baseline accuracy without using SSL (the "rich" one) tends to benefit more from SSL; while the sub-population who suffers from a low baseline accuracy (the "poor" one) might even observe a performance drop after adding the SSL module. We theoretically and empirically establish the above observation for a broad family of SSL algorithms, which either explicitly or implicitly use an auxiliary "pseudo-label". Experiments on a set of image and text classification tasks confirm our claims. We introduce a new metric, Benefit Ratio, and promote the evaluation of the fairness of SSL (Equalized Benefit Ratio). We further discuss how the disparate impact can be mitigated. We hope our paper will alarm the potential pitfall of using SSL and encourage a multifaceted evaluation of future SSL algorithms.

Motivation & Objective

• To investigate whether semi-supervised learning (SSL) disproportionately benefits certain sub-populations over others.
• To understand the root causes of disparate impacts in SSL, especially when sub-populations are defined by class-level or demographic-level accuracy.
• To propose a new fairness evaluation metric, the benefit ratio, to measure normalized improvement in model accuracy across sub-populations after SSL.
• To evaluate and suggest practical mitigation strategies—such as data balancing and increased labeled data—for reducing unfairness in SSL.
• To promote a multifaceted evaluation of SSL algorithms beyond average accuracy, emphasizing fairness in model improvement.

Proposed method

• The authors develop a unified analytical framework that connects consistency regularization in SSL to learning with noisy labels, enabling theoretical analysis of generalization error.
• They derive an upper bound on the generalization error of SSL, showing that smaller-scale supervised datasets lead to heterogeneous errors, which drives disparate impacts.
• A new metric, benefit ratio, is introduced to quantify the normalized improvement in accuracy for each sub-population after applying SSL.
• The benefit ratio is used to detect the 'Matthew effect' in SSL: high baseline accuracy groups see larger gains (ratio >1), while low baseline groups may suffer (negative ratio).
• The authors empirically evaluate the benefit ratio on CIFAR-10, CIFAR-100, and the Jigsaw toxicity dataset using multiple SSL methods, including MixMatch.
• They apply two mitigation strategies: balancing labeled and unlabeled data, and collecting additional labeled data, and evaluate their impact on reducing benefit ratio variance and negative ratios.

Experimental results

Research questions

• RQ1 Does semi-supervised learning lead to unequal improvements across sub-populations defined by class or demographic groups?
• RQ2 Why do sub-populations with higher baseline accuracy benefit more from SSL, while others may see performance degradation?
• RQ3 Can the benefit ratio metric effectively reveal and quantify disparate impacts in SSL across different groups?
• RQ4 To what extent can data balancing and increased labeled data mitigate the unfairness in SSL improvements?
• RQ5 Is there a theoretical basis for the observed Matthew effect—'the rich get richer and the poor get poorer'—in SSL?

Key findings

• The 'rich get richer' phenomenon is widespread in SSL: sub-populations with high baseline accuracy consistently achieve higher benefit ratios, with some exceeding 1, indicating amplified gains.
• In the unbalanced setting on CIFAR-10, the dog class (low baseline accuracy) saw a negative benefit ratio of -9.57% under MixMatch, indicating performance degradation after SSL.
• On the Jigsaw dataset, the unbalanced setting produced 4 out of 8 race identities with negative benefit ratios, highlighting significant fairness issues in text classification.
• After balancing both labeled and unlabeled data on Jigsaw with 800 total labeled instances, the standard deviation of benefit ratios dropped to 7.64%, and only one group had a negative benefit ratio.
• The benefit ratio metric effectively captures disparate impacts and is grounded in theoretical generalization bounds, validating its use for fairness evaluation in SSL.
• Mitigation strategies such as data balancing and collecting more labeled data significantly reduce the standard deviation of benefit ratios and eliminate negative improvements in most sub-populations.