Adaptive fairness improvement based on causality analysis

Mengdi Zhang, Jun Sun

Introduction

As neural networks are increasingly deployed in high-stakes domains such as credit scoring and criminal sentencing, ensuring algorithmic fairness has become a critical priority. Current fairness-improving techniques are generally categorized into pre-processing (data modification), in-processing (model architecture or loss function changes), and post-processing (prediction adjustment). However, selecting the appropriate method remains a significant challenge. Empirical evidence suggests that applying the wrong mitigation strategy can be a "coin toss," potentially worsening bias or severely degrading model performance.

The Problem: The Risk of Misapplied Fairness

There is no universal solution for algorithmic bias. A study of existing tools revealed that in approximately 20% of cases, applying a fairness method actually increased discrimination. Furthermore, the impact on accuracy can be devastating. For instance, applying the "Reject Option" (RO) method to the COMPAS dataset resulted in a 35% drop in accuracy, as illustrated in the performance fluctuations across various benchmarks.

Accuracy changes across different datasets and fairness mitigation methods.

Currently, practitioners often rely on trial-and-error to find a suitable method, a process that is both computationally expensive and risky. A method that improves fairness for one attribute, such as gender, may simultaneously exacerbate bias for another, such as race.

Methodology: A Causal Framework for Diagnosis

To address these challenges, this paper introduces an Adaptive Fairness Improvement framework. Instead of arbitrary selection, the framework performs a "causal medical checkup" to diagnose the origin of unfairness—whether it stems from the raw training data or the internal representations of the neural network.

Causality Analysis

The framework treats the neural network as a Structured Causal Model (SCM). By utilizing a technique called Average Interventional Expectation, the system intervenes on specific input attributes and hidden neurons to quantify their "causal responsibility" for the observed bias.

Causal effect analysis across input attributes and hidden layers, identifying the responsibility of different components for model bias.

Selection Logic

The diagnostic results guide the selection of the mitigation strategy:

  1. Post-processing: Recommended if the causal responsibility for bias is low across both inputs and neurons (the least invasive option).
  2. Pre-processing: Recommended if bias is highly concentrated in specific input attributes, as determined by the Coefficient of Variation (CV).
  3. In-processing: Recommended if the bias is primarily concentrated within the hidden neurons, suggesting the model itself is amplifying the unfairness.

Experimental Results

The framework was evaluated using standard benchmarks: Adult Income, German Credit, Bank Marketing, and COMPAS. Key findings include:

  • Optimal Selection: The adaptive approach identified the most effective fairness method 100% of the time for group fairness across all tested scenarios.
  • Accuracy Preservation: By avoiding unnecessary or invasive corrections, the method maintained significantly higher accuracy compared to blindly applying popular techniques like META or Reject Option Classification.
  • Efficiency: The causality analysis adds an average overhead of only 5 minutes, offering a massive efficiency gain over traditional exhaustive search methods.

Conclusion and Future Work

While highly effective for feed-forward and convolutional architectures, the current framework does not yet support Recurrent Neural Networks (RNNs) due to their complex feedback loops. Future research will explore hybrid mitigation strategies—applying multiple types of interventions simultaneously when bias is distributed across both data and model layers—and extending the causal analysis to more complex temporal architectures.

Summary updated

BibTeX
@inproceedings{zhang2022adaptive,
  author = {Mengdi Zhang and Jun Sun},
  title = {{Adaptive fairness improvement based on causality analysis}},
  booktitle = {ESEC/SIGSOFT FSE},
  year = {2022},
  url = {https://openreview.net/forum?id=gEqJvHpS2C},
}
Figure