Situation Testing-Based Discrimination Discovery: A Causal Inference Approach

Lu Zhang, Yongkai Wu, Xintao Wu

Introduction

Discrimination in automated decision-making is often obscured by complex variables and "proxy" attributes. To identify bias at the individual level, legal and social science experts frequently employ situation testing. This involves comparing the outcome for a specific individual against a "counterfactual" peer who is identical in all relevant aspects except for a protected attribute (e.g., gender or race). Researchers from the University of Arkansas have proposed a novel framework, CBN-DD (Causal Bayesian Network - Discrimination Discovery), which leverages causal inference to refine this process, making it both more accurate and significantly more efficient than traditional statistical methods.

The Challenge of Digital Situation Testing

In data science, situation testing is typically replicated by finding "similar" pairs in historical datasets using algorithms like K-Nearest Neighbor (KNN-DD). However, these traditional approaches suffer from two primary limitations:

  1. The Noise Problem: They often calculate similarity using all available attributes. If a dataset includes irrelevant information—such as height or weight for a software engineering role—these "noise" variables distort the similarity score.
  2. The Weighting Problem: They treat all attribute differences with equal importance. In reality, a difference in professional certification should carry more weight in a similarity calculation than a difference in residential zip code.

Methodology: Causal Bayesian Networks

The CBN-DD approach addresses these flaws by integrating Causal Bayesian Networks (CBNs) into the discovery process. The method relies on two core causal insights:

1. Causal Filtering

Instead of considering every attribute, the algorithm identifies the set of attributes ($Q$) that are direct causal parents of the decision ($E$). By focusing only on these causal drivers, the system ignores irrelevant noise.

Figure 1: A causal graph illustrating how certain attributes like 'height' and 'weight' are influenced by 'gender' but do not causally impact 'admission', whereas 'major' and 'test score' do.

2. Causal Weighting

The distance function between individuals is weighted by the Causal Effect (CE). This metric quantifies how much a change in a specific attribute (e.g., improving a grade from 'B' to 'A') actually alters the probability of a positive outcome. This ensures that "similarity" is defined by the factors that truly influence the decision-making process.

Experimental Evaluation

The researchers evaluated CBN-DD using the Dutch Census of 2001 (60,421 records). To rigorously test the algorithm's accuracy, they created a "clean" baseline and manually injected discrimination into 100 specific records. The goal was to see how effectively the algorithm could rediscover these known instances of bias.

Key Results

The study compared CBN-DD against the previous state-of-the-art, KNN-DD, across several metrics:

  • Accuracy: CBN-DD demonstrated superior True Positive (TP) rates. As shown in the results table, at $K=50$, CBN-DD correctly identified 85.3% of discriminated individuals, whereas KNN-DD's performance was significantly lower and less consistent.
  • Efficiency: The most striking improvement was in computational speed. CBN-DD averaged 0.3 seconds per test, compared to 20.3 seconds for KNN-DD—a 60x speedup. This efficiency stems from "causal profile" grouping; since the distance function only considers direct causes, many individuals share identical profiles, allowing for batch processing.
Figure 4: Performance comparison between CBN-DD and KNN-DD, showing True Positive (TP) and True Negative (TN) rates across different values of K.

Conclusion and Future Directions

By moving from purely statistical correlation to a causal framework, CBN-DD aligns digital discrimination discovery with legal interpretations of fairness. While the method currently depends on the accuracy of the underlying causal graph and focuses on single protected attributes, it represents a major step forward in building scalable, interpretable tools for auditing algorithmic bias. Future work will explore intersectional discrimination involving multiple protected attributes simultaneously.

Summary updated

Abstract

Discrimination discovery is to unveil discrimination against a specific individual by analyzing the historical dataset. In this paper, we develop a general technique to capture discrimination based on the legally grounded situation testing methodology. For any individual, we find pairs of tuples from the dataset with similar characteristics apart from belonging or not to the protected-by-law group and assign them in two groups. The individual is considered as discriminated if significant di?erence is observed between the decisions from the two groups. To find similar tuples, we make use of the Causal Bayesian Networks and the associated causal inference as a guideline. The causal structure of the dataset and the causal effect of each attribute on the decision are used to facilitate the similarity measurement. Through empirical assessments on a real dataset, our approach shows good efficacy both in accuracy and efficiency.
BibTeX
@inproceedings{zhang2016situation,
  author = {Lu Zhang and Yongkai Wu and Xintao Wu},
  title = {{Situation Testing-Based Discrimination Discovery: A Causal Inference Approach}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2016},
}
Figure