Confronting LLMs with Traditional ML: Rethinking the Fairness of Large Language Models in Tabular Classifications (2310.14607v2)

Abstract: Recent literature has suggested the potential of using LLMs to make classifications for tabular tasks. However, LLMs have been shown to exhibit harmful social biases that reflect the stereotypes and inequalities present in society. To this end, as well as the widespread use of tabular data in many high-stake applications, it is important to explore the following questions: what sources of information do LLMs draw upon when making classifications for tabular tasks; whether and to what extent are LLM classifications for tabular data influenced by social biases and stereotypes; and what are the consequential implications for fairness? Through a series of experiments, we delve into these questions and show that LLMs tend to inherit social biases from their training data which significantly impact their fairness in tabular classification tasks. Furthermore, our investigations show that in the context of bias mitigation, though in-context learning and finetuning have a moderate effect, the fairness metric gap between different subgroups is still larger than that in traditional machine learning models, such as Random Forest and shallow Neural Networks. This observation emphasizes that the social biases are inherent within the LLMs themselves and inherited from their pretraining corpus, not only from the downstream task datasets. Besides, we demonstrate that label-flipping of in-context examples can significantly reduce biases, further highlighting the presence of inherent bias within LLMs.

Investigating Fairness in LLMs for Tabular Predictions

The paper "Investigating the Fairness of LLMs for Predictions on Tabular Data" by Yanchen Liu et al. evaluates the fairness of LLMs when employed for tabular prediction tasks. Given the prevalent use of tabular data in various critical applications and the recognized presence of inherent social biases within LLMs, this work aims to understand the extent and implications of such biases in tabular prediction contexts.

The exploration is structured around three primary inquiries: the sources of information LLMs rely upon when making tabular predictions, the influence of social biases and stereotypes on these predictions, and potential fairness implications. Through a series of comprehensive experiments employing GPT-3.5 in zero-shot, few-shot, and fine-tuning modes, alongside Random Forest (RF) and shallow Neural Network (NN) models, this research systematically examines these inquiries.

Key Findings

1. Zero-Shot Fairness Analysis: The results reveal significant fairness concerns when LLMs are used for tabular tasks in a zero-shot setting. LLMs exhibit a greater fairness metric gap between protected and non-protected groups compared to traditional models such as RF and NN, highlighting the potential risk of employing LLMs for tabular predictions without further mitigation.
2. In-Context Learning: Incorporation of few-shot examples demonstrates a moderate reduction of bias, yet fails to eliminate them completely. This illustrates that the LLMs' inherent biases, inherited from pre-training data, supersede those arising directly from task datasets.
3. Label-Flipping Experiment: By flipping labels within the few-shot examples, the paper observes a noteworthy reduction of bias, albeit at the cost of predictive performance. This substantiates the hypothesis that LLMs' social biases are primarily rooted in the pre-training phase rather than simply the downstream task dataset or the inherent feature structure.
4. Fine-Tuning and Resampling: Fine-tuning LLMs on entire training datasets partially alleviates these biases, with varying effectiveness. However, traditional resampling methods (oversampling and undersampling) do not significantly mitigate bias in LLMs as they do with traditional machine learning models.

Implications and Future Directions

The findings underscore the complexities associated with leveraging LLMs for tabular data predictions, particularly in scenarios demanding high fairness standards. These insights suggest that depending solely on existing LLMs without considering appropriate bias mitigation strategies might exacerbate societal inequalities, especially in high-stakes applications such as finance and criminal justice, where tabular data is prevalent.

Future research could explore more nuanced and domain-specific tweaking of LLMs, possibly involving hybrid models that combine the strengths of LLMs with fairness-aware traditional ML techniques. It also opens avenues for developing novel pre-training methodologies or architectural modifications aimed specifically at reducing inherent biases.

The paper draws attention to the necessity for fairness-centric approaches when deploying LLMs, advocating for continued exploration into effective mitigation strategies that transcend current straightforward methods like in-context learning and data resampling.

In conclusion, this work provides a detailed analysis of the fairness of LLMs for tabular data tasks, highlighting significant challenges and providing a foundation for future research initiatives targeting bias reduction in AI systems. It contributes to the broader discourse on ethical AI, emphasizing the importance of equitable algorithmic performance across diverse application domains.