Bias Mitigation in Large Language Models for Tabular Data Classification

Abstract Large Language Models (LLMs) perform well in tabular prediction tasks with limited data, using their ability to understand instructions and learn from examples. However, their reliance on training data can perpetuate social biases, leading to unfair outcomes and disproportionately impacting underprivileged groups. Addressing these biases is critical as LLMs see wider adoption in tabular data tasks. Traditional bias mitigation strategies in machine learning, such as balancing datasets or applying fairness constraints, are less effective with LLMs. Our research explores whether bias in LLMs for tabular data classification can be mitigated. Through extensive experiments, we found that using LLMs in a zero-shot setting introduces bias, and in-context learning slightly reduces these disparities. Meanwhile, fine-tuning and retrieval augmented generation show limited effectiveness in bias mitigation. We introduced three instruction-based prompting strategies to enhance fairness: Fair Prompting , Generalised Prompting , and Descriptive Prompting . The results show that combining descriptive prompting with in-context learning, particularly the Equal Samples Across Demographics approach, consistently narrowed fairness gaps across demographic subgroups, and yielded accuracy gains ranging from 3.27% to 15.05% across multiple datasets, underscoring its potential as a promising strategy in the ongoing effort to mitigate bias in LLMs.