Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Few-Shot Fairness: Unveiling LLM's Potential for Fairness-Aware Classification

Garima Chhikara, Anurag Sharma, Kripabandhu Ghosh, Abhijnan Chakraborty

TL;DR

The paper addresses whether large language models can produce fairness-aware classifications when guided by prompt-based in-context learning. It introduces a fairness framework with abstract (π_A) and detailed (π_D) rules, and evaluates zero-shot and few-shot settings using Retrieval-Augmented Generation to select demonstrations on the UCI Adult dataset across GPT-4, LLaMA-2, and Gemini. GPT-4 consistently achieves the best balance of accuracy and fairness across multiple definitions, though metrics like Disparate Impact, TPR, and FPR reveal persisting group biases. Generic Fairness prompts perform comparably to abstract definitions, highlighting the potential of simple, broad fairness signals to improve outputs, while also underscoring the need for broader datasets and models to reduce remaining biases before deployment.

Abstract

Employing Large Language Models (LLM) in various downstream applications such as classification is crucial, especially for smaller companies lacking the expertise and resources required for fine-tuning a model. Fairness in LLMs helps ensure inclusivity, equal representation based on factors such as race, gender and promotes responsible AI deployment. As the use of LLMs has become increasingly prevalent, it is essential to assess whether LLMs can generate fair outcomes when subjected to considerations of fairness. In this study, we introduce a framework outlining fairness regulations aligned with various fairness definitions, with each definition being modulated by varying degrees of abstraction. We explore the configuration for in-context learning and the procedure for selecting in-context demonstrations using RAG, while incorporating fairness rules into the process. Experiments conducted with different LLMs indicate that GPT-4 delivers superior results in terms of both accuracy and fairness compared to other models. This work is one of the early attempts to achieve fairness in prediction tasks by utilizing LLMs through in-context learning.

Few-Shot Fairness: Unveiling LLM's Potential for Fairness-Aware Classification

TL;DR

The paper addresses whether large language models can produce fairness-aware classifications when guided by prompt-based in-context learning. It introduces a fairness framework with abstract (π_A) and detailed (π_D) rules, and evaluates zero-shot and few-shot settings using Retrieval-Augmented Generation to select demonstrations on the UCI Adult dataset across GPT-4, LLaMA-2, and Gemini. GPT-4 consistently achieves the best balance of accuracy and fairness across multiple definitions, though metrics like Disparate Impact, TPR, and FPR reveal persisting group biases. Generic Fairness prompts perform comparably to abstract definitions, highlighting the potential of simple, broad fairness signals to improve outputs, while also underscoring the need for broader datasets and models to reduce remaining biases before deployment.

Abstract

Employing Large Language Models (LLM) in various downstream applications such as classification is crucial, especially for smaller companies lacking the expertise and resources required for fine-tuning a model. Fairness in LLMs helps ensure inclusivity, equal representation based on factors such as race, gender and promotes responsible AI deployment. As the use of LLMs has become increasingly prevalent, it is essential to assess whether LLMs can generate fair outcomes when subjected to considerations of fairness. In this study, we introduce a framework outlining fairness regulations aligned with various fairness definitions, with each definition being modulated by varying degrees of abstraction. We explore the configuration for in-context learning and the procedure for selecting in-context demonstrations using RAG, while incorporating fairness rules into the process. Experiments conducted with different LLMs indicate that GPT-4 delivers superior results in terms of both accuracy and fairness compared to other models. This work is one of the early attempts to achieve fairness in prediction tasks by utilizing LLMs through in-context learning.
Paper Structure (41 sections, 11 equations, 3 figures, 9 tables)

This paper contains 41 sections, 11 equations, 3 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Fairness in LLMs
  4. In-context Learning
  5. Experimental Setup
  6. Dataset
  7. Large Language Models
  8. Fairness Definition
  9. Definitions based on Predicted Outcome
  10. Definitions based on Predicted and Actual Outcome
  11. Definitions based on Similarity
  12. Fairness Metrics
  13. Disparate Impact (DI)
  14. True Positive Rate (TPR)
  15. False Positive Rate (FPR)
  16. ...and 26 more sections

Figures (3)

  • Figure 1: An example showcasing a scenario where a user inquires GPT-4 about the acceptance of their university admission application. Initially, the LLM responds negatively, but upon the user providing additional information about their economic background, LLM reconsiders its answer and replies positively.
  • Figure 2: This example shows a part of the conversation with GPT-4 about the Stop and Frisk Policy (the complete conversation can be found in the Appendix Fig. \ref{['fig:stopfriskdetailed']}). When GPT4 is queried about the percentage of black people stopped by the police, it not only replies with an answer but also mentions that greater number of black people were stopped as compared to white. When queried about fairness, the model adheres to the concept of Proportional Representation, also known as Statistical Parity 10.1145/2783258.2783311, asserting that if black people constitute 23% of the population, they should comprise only 23% of the stops in the entire population.
  • Figure 3: Conversation with GPT-4 about Stop and Frisk Policy.