Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Peer-induced Fairness: A Causal Approach for Algorithmic Fairness Auditing

Shiqi Fang, Zexun Chen, Jake Ansell

TL;DR

A novel framework, ``peer-induced fairness'', which combines the strengths of counterfactual fairness and peer comparison strategy, creating a reliable and robust tool for auditing algorithmic fairness and highlighting the framework's potential for broader applications in ensuring equitable AI-driven decision-making.

Abstract

With the European Union's Artificial Intelligence Act taking effect on 1 August 2024, high-risk AI applications must adhere to stringent transparency and fairness standards. This paper addresses a crucial question: how can we scientifically audit algorithmic fairness? Current methods typically remain at the basic detection stage of auditing, without accounting for more complex scenarios. We propose a novel framework, ``peer-induced fairness'', which combines the strengths of counterfactual fairness and peer comparison strategy, creating a reliable and robust tool for auditing algorithmic fairness. Our framework is universal, adaptable to various domains, and capable of handling different levels of data quality, including skewed distributions. Moreover, it can distinguish whether adverse decisions result from algorithmic discrimination or inherent limitations of the subjects, thereby enhancing transparency. This framework can serve as both a self-assessment tool for AI developers and an external assessment tool for auditors to ensure compliance with the EU AI Act. We demonstrate its utility in small and medium-sized enterprises access to finance, uncovering significant unfairness-41.51% of micro-firms face discrimination compared to non-micro firms. These findings highlight the framework's potential for broader applications in ensuring equitable AI-driven decision-making.

Peer-induced Fairness: A Causal Approach for Algorithmic Fairness Auditing

TL;DR

A novel framework, ``peer-induced fairness'', which combines the strengths of counterfactual fairness and peer comparison strategy, creating a reliable and robust tool for auditing algorithmic fairness and highlighting the framework's potential for broader applications in ensuring equitable AI-driven decision-making.

Abstract

With the European Union's Artificial Intelligence Act taking effect on 1 August 2024, high-risk AI applications must adhere to stringent transparency and fairness standards. This paper addresses a crucial question: how can we scientifically audit algorithmic fairness? Current methods typically remain at the basic detection stage of auditing, without accounting for more complex scenarios. We propose a novel framework, ``peer-induced fairness'', which combines the strengths of counterfactual fairness and peer comparison strategy, creating a reliable and robust tool for auditing algorithmic fairness. Our framework is universal, adaptable to various domains, and capable of handling different levels of data quality, including skewed distributions. Moreover, it can distinguish whether adverse decisions result from algorithmic discrimination or inherent limitations of the subjects, thereby enhancing transparency. This framework can serve as both a self-assessment tool for AI developers and an external assessment tool for auditors to ensure compliance with the EU AI Act. We demonstrate its utility in small and medium-sized enterprises access to finance, uncovering significant unfairness-41.51% of micro-firms face discrimination compared to non-micro firms. These findings highlight the framework's potential for broader applications in ensuring equitable AI-driven decision-making.
Paper Structure (18 sections, 2 theorems, 18 equations, 7 figures, 1 table, 1 algorithm)

This paper contains 18 sections, 2 theorems, 18 equations, 7 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Counterfactual fairness and its representation
  3. Peer-induced fairness with causal method
  4. Discrimination from peer comparisons
  5. Fairness through peer observations
  6. Peer definition and identification
  7. Peer-induced fairness
  8. Peer-induced fairness auditing
  9. Overall auditing workflow
  10. Experiment
  11. Results
  12. Algorithmic fairness auditing
  13. Data scarcity and imbalance
  14. Explainable fairness discovery
  15. Concluding remarks and discussion
  16. ...and 3 more sections

Key Result

Theorem 1

Consider an individual $A = (s_{-}, \bm{x}_0)$ and assuming there are a group of individuals $\mathcal{B} = \{B_1, B_2, \cdots$ } from unprotected group, where $B_j = (s_{+}, \bm{x}_j)$. An individual $C \in \mathcal{B}$ is identified as a $\delta$-peer of $A$ if:

Figures (7)

  • Figure 1: SWIGs for Graphical Causal Models (GCM). The nodes with black border represent random variables, while red ones indicate fixed values of random variables, representing experimental interventions. Arrows depict causal relationships between variables. (a): The SWIG $\mathcal{G}(s_{-}, \bm{x})$ represents the actual scenario for an individual with features $(s_{-}, \bm{x})$. (b): The SWIG $\mathcal{\tilde{G}}(s_{+}, \bm{x})$ illustrates the counterfactual scenario, assuming the individual's protected feature changes from $s_{-}$ to $s_{+}$, while their other features $\bm{x}$ remain the same. (c): The SWIG $\mathcal{G}(s_{+}, \bm{x}')$ represents the actual scenario for an individual with features $(s_{+}, \bm{x}')$. The actual SWIG $\mathcal{G}(s_{-}, \bm{x})$ corresponds to the conditional distribution $\hat{Y}_{s_{-}} | s_{-}, \bm{x}$. Conversely, in the counterfactual SWIG $\mathcal{\tilde{G}}(s_{+}, \bm{x})$ refers to $\hat{Y}_{s_{+}} | s_{-}, \bm{x}$, denoting the outcome distribution had the individual been featured with $s_{+}$, given that the actual features are $(s_{-}, \bm{x})$. Thus the directed link from $s_{+}$ to $\bm{X}(s_{-})$ is not the fact (shown in green colour). Note: $\mathcal{\tilde{G}}(s_{+},\bm{x}) \neq \mathcal{G}(s_{+},\bm{x}')$ because $\mathcal{\tilde{G}}(s_{+},\bm{x})$ is counterfactual scenario with actual features $(s_{-},\bm{x})$ and $\mathcal{G}(s_{+},\bm{x}')$ is the fact with features $(s_{+},\bm{x}')$.
  • Figure 2: The overall auditing workflow. "Compute $\mathbb{P}(\hat{Y} = 1|s_, x)$" step requires a given prediction model, "Compute $IC$" step requires a given fitting model, "Find peers in B using $IC$, denote $\mathcal{C}(a)$" step requires a given $\delta$. Although the fitting model is usually the same as the prediction model, distinct choices are also allowed.
  • Figure 3: Comparative analysis of loan approval likelihood for micro-firms against peers. The black dashed 45-degree line, denoting $Y=X$, symbolises perfect fairness. Red and orange data points represent micro-firms with approval likelihoods significantly lower or higher, respectively than the average of their peers. Blue points denote no significant difference.
  • Figure 4: Comparative analysis of loan approval likelihood for micro-firms under each algorithmic treatment category against peers. (a)-(c): Extremely discriminated (ED) micro-firms. (d)-(f): Fairly treated (FT) micro-firms. (g)-(i): Extremely privileged (EP) micro-firms. The coloured data points in the first column of each row represent a comparison among peers within each category. The x-axis shows the approval likelihood for micro-firms, while the y-axis displays the average approval likelihood of the peers. The second column compares the approval likelihood between these micro-firms (i.e., coloured points in (a), (d), (g)) and their peers at the group level. The third column provides the comparison, at the individual level, between the selected micro-firm (i.e., coloured triangle in (a), (d), (g)) and its peers.
  • Figure 5: Rejection rates of micro-firms across algorithmic treatment categories and their peers. The algorithmic treatment categories include extremely discriminated (ED), fairly treated (FT), and extremely privileged (EP). Each category includes multiple micro-firms with a single rejection rate, shown as histograms, while the rejection rate of peers of each micro-firm in this category is represented in the black line with error bars to indicate variability.
  • ...and 2 more figures

Theorems & Definitions (8)

  • Definition 1: Counterfactual fairness
  • Definition 2
  • Definition 3: $\delta$-peer
  • Theorem 1: $\delta$-peer identification
  • Definition 4: $(\delta, f)$-peer-induced fairness
  • Proposition 1
  • proof
  • proof