Fairness-aware Class Imbalanced Learning

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2045–2051 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2045 Fairness-aware Class Imbalanced Learning Shivashankar Subramanian♣ Afshin Rahimi♠ Timothy Baldwin♣ Trevor Cohn♣ Lea Frermann♣ ♣School of Computing and Information Systems, University of Melbourne ♠School of Information Technology and Electrical Engineering, University of Queensland shivashankarrs@gmail.com a.rahimi@uq.edu.au {tbaldwin,t.cohn,lfrermann}@unimelb.edu.au Abstract Class imbalance is a common challenge in many NLP tasks, and has clear connections to bias, in that bias in training data often leads to higher accuracy for majority groups at the expense of minority groups. However there has traditionally been a disconnect between re- search on class-imbalanced learning and mit- igating bias, and only recently have the two been looked at through a common lens. In this work we evaluate long-tail learning meth- ods for tweet sentiment and occupation clas- sification, and extend a margin-loss based ap- proach with methods to enforce fairness. We empirically show through controlled experi- ments that the proposed approaches help miti- gate both class imbalance and demographic bi- ases.1 1 Introduction Class imbalance is common in many NLP tasks, in- cluding machine reading comprehension (Li et al., 2020), authorship attribution (Caragea et al., 2019), toxic language detection (Breitfeller et al., 2019), and text classification (Tian et al., 2020). skewed class distribution hurts the performanceA of deep learning models (Buda et al., 2018), and approaches such as instance weighting (Lin et al., 2017; Cui et al., 2019; Li et al., 2020), data aug- mentation (Juuti et al., 2020; Wei and Zou, 2019), and weighted max-margin (Cao et al., 2019) are commonly used to alleviate the problem. Bias in data often also manifests as skewed dis- tributions, especially when considered in combi- nation with class labels. This is often referred to as “stereotyping” whereby one or more private at- tributes are associated more frequently with cer- tain target labels, for instance more men being employed as surgeons than women. Prior work has identified several classes of bias, including 1Code available at: https://github.com/ shivashankarrs/classimb_fairness C Figure 1: Example of a two-class problem where grey and blue points denote majority and minority classes, respectively, and circles and squares denote two sub-groups. Imbalanced learning methods such as LDAM (Cao et al., 2019) maximise the (soft-)margin for minority classes and do not consider sub-groups within each class. bias towards demographic groups based on gen- der, disability, race or religion (Caliskan et al., 2017; May et al., 2019; Garimella et al., 2019; Nangia et al., 2020), and bias towards individu- als (Prabhakaran et al., 2019). Methods to mitigate these biases include data augmentation (Badjatiya et al., 2019), adversarial learning (Li et al., 2018), instance weighting based on group membership (Kamiran and Calders, 2011), regularization (Wick et al., 2019; Kennedy et al., 2020), and explicit sub- space removal (Bolukbasi et al., 2016; Ravfogel et al., 2020). This paper draws a connection between class- imbalanced learning and stereotyping bias. Most work has focused on class-imbalanced learning and bias mitigation as separate problems, but …

Content truncated to 500 words