Closing the Gap in the Trade-off between Fair Representations and Accuracy

TL;DR

The paper addresses representation-level fairness in NLP encodings by analyzing how protected-group information is encoded along principal components and how this affects downstream classification. It compares two simple encoding strategies—vector averaging and vector extrema—on Hindi and English datasets using SVM classifiers, revealing a trade-off between fairness and accuracy. To mitigate bias without sacrificing performance, the authors propose a convex combination of encodings, finding that a weight near $\lambda \approx 0.97$ can reduce reconstruction disparities while preserving high accuracy on the target tasks. This approach offers a practical, replication-friendly pathway to balance fairness and performance in NLP representations, with potential applicability beyond binary classification. Future work includes more principled optimization of the combination parameter and extending the method to other NLP tasks.

Abstract

The rapid developments of various machine learning models and their deployments in several applications has led to discussions around the importance of looking beyond the accuracies of these models. Fairness of such models is one such aspect that is deservedly gaining more attention. In this work, we analyse the natural language representations of documents and sentences (i.e., encodings) for any embedding-level bias that could potentially also affect the fairness of the downstream tasks that rely on them. We identify bias in these encodings either towards or against different sub-groups based on the difference in their reconstruction errors along various subsets of principal components. We explore and recommend ways to mitigate such bias in the encodings while also maintaining a decent accuracy in classification models that use them.

Figures (7)

• Figure 1: PCA reconstruction error for MTC-gen and HLDC datasets at each number of dimensions used for PCA reconstruction with each of vector extrema and vector averaging techniques
• Figure 2: PCA Reconstruction Error within groups (i.e., random sub-groups of same group) and across groups
• Figure 3: Accuracy of SVMs while using different number of dimensions (5d, 10d, 15d, 30d) at various values of $\lambda$ to form a convex combination of vector average and vector extrema encodings
• Figure 4: PCA reconstruction difference versus classification accuracy for various $\lambda$ convex combination of vector average and vector extrema, at various dimensions of PCA reconstruction
• Figure 5: PCA reconstruction errors for each ethnicity group on the MTC-eth data