Source-linked AI summary
Adaptive Fairness Improvement Based on Causality Analysis
Mengdi Zhang, Jun Sun
TL;DR
Choosing a fairness-improvement method is difficult because the wrong choice can reduce accuracy or worsen fairness. The paper uses causality analysis of responsible attributes and neurons to select a method, reporting optimal group-fairness choices in all evaluated cases and individual-discrimination choices in most.
Problem
Fairness methods can harm accuracy or fairness, and the effort required to test them all makes selecting the right method difficult.
Source: §1 · p. 1Method
The approach uses causal-effect distributions across input attributes and hidden neurons to choose a fairness-improvement method.
Source: §1 · pp. 1–2Results
The approach selected the optimal method for group-fairness improvement in all evaluated cases and for reducing individual discrimination in most cases.
Source: §1 · p. 2Takeaways & Limitations
The evaluation reports the lightweight approach as effective in choosing an optimal fairness-improvement method.
Source: §8 · p. 11The approach may be affected by distributional shifts; its stability was evaluated under slight shifts on Adult Income.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Given a discriminating neural network, the problem of fairness improvement is to systematically reduce discrimination without significantly scarifies its performance (i.e., accuracy). Multiple categories of fairness improving methods have been proposed for neural networks, including pre-processing, in-processing and post-processing. Our empirical study however shows that these methods are not always effective (e.g., they may improve fairness by paying the price of huge accuracy drop) or even not helpful (e.g., they may even worsen both fairness and accuracy). In this work, we propose an approach which adaptively chooses the fairness improving method based on causality analysis. That is, we choose the method based on how the neurons and attributes responsible for unfairness are distributed among the input attributes and the hidden neurons. Our experimental evaluation shows that our approach is effective (i.e., always identify the best fairness improving method) and efficient (i.e., with an average time overhead of 5 minutes).
1 INTRODUCTION
Fairness-improvement methods can behave differently across models and protected attributes, and the wrong choice may damage accuracy or worsen fairness. The paper selects methods through causality analysis of how unfairness is distributed across input attributes and hidden neurons.
1 INTRODUCTION
- Existing neural-network fairness-improvement methods fall into preprocessing, in-processing, and post-processing categories.They respectively target training data, the model or training process, and prediction results.Source: §1 · p. 1
- Applying the wrong fairness method can cause substantial accuracy loss or worsen fairness, and testing every method is infeasible.A method may also improve fairness for one protected attribute while worsening it for another.Source: §1 · p. 1
- The proposed approach analyzes whether unfairness is concentrated in input attributes or hidden neurons, then selects a fairness-improvement method accordingly.It uses the probability of high causal effects and the Coefficient of Variation to characterize causal-effect distributions.Source: §1 · pp. 1–2
- The approach selects an optimal method for group-fairness improvement in all evaluated cases and for reducing individual discrimination in most cases.The evaluation considered nine methods across four benchmark datasets.Source: §1 · p. 2
2 BACKGROUND
The background introduces group and individual discrimination as fairness concerns and surveys representative methods for improving neural-network fairness. It describes group-level metrics and individual-level label changes, alongside preprocessing, in-processing, and post-processing approaches.
2 BACKGROUND
- The background reviews fairness concepts and existing methods for improving fairness in neural networks.Source: §2 · p. 2
2.1 Fairness Definitions
- Group fairness compares outcomes across protected groups; SPD measures favorable-outcome probability differences for one binary attribute, while GDS extends the comparison to multiple attributes.GDS takes the maximum difference between probabilities for protected-attribute valuations.Source: §2.1 · p. 2
- Individual discrimination occurs when changing protected attributes alone changes a person's predicted label; CDS measures the fraction of inputs that are discriminatory instances.Source: §2.1 · p. 2
2.2 Fairness Improving Methods
- Preprocessing methods target bias in training data: Reweighing changes sample weights, while Disparate Impact Remover modifies non-protected attributes.Source: §2.2 · p. 3
- In-processing methods alter models or training, using fairness constraints, adversarial debiasing, prejudice-removal regularization, or cost-sensitive classification.Source: §2.2 · p. 3
- Post-processing methods alter predictions, including Equalized Odds and Calibrated Equalized Odds label changes and Reject Option decisions near uncertain boundaries.Source: §2.2 · p. 3
3 AN EMPIRICAL STUDY
The study compared fairness-improving methods across benchmark models and protected attributes, finding that group-fairness effects and accuracy costs vary by method and setting, and that group fairness does not reliably track individual discrimination.
3.1 Experimental Setup
- The comparison covered four benchmark datasets, 10 model–attribute combinations, and nine fairness-improving methods implemented using AIF360.Source: §3.1 · p. 3
3.2 Evaluation Results
- Table 3's best methods differed by model–attribute combination; RW improved group fairness across all combinations, unlike DIR on Credit-gender and COMPAS-race.Source: Table 3 · p. 4, §3.2 · p. 4
- 18 of 90 cases (20%) worsened group fairness after a fairness-improving method, mostly with in- or post-processing.Source: §3.2 · p. 4, Figure 1 · p. 5
- Some methods substantially reduced accuracy: on COMPAS, META, PR, or RO each caused drops over 40%, while RW generally had mild impact.Source: Figure 2 · p. 5, §3.2 · p. 5
- Group-fairness gains did not reliably reduce individual discrimination: RW worsened it in 8 combinations, and RO worsened it across all combinations.Source: Figure 3 · p. 6, §3.2 · p. 6
4 AN ADAPTIVE APPROACH
The adaptive approach analyzes how causal responsibility for unfairness is distributed across input attributes and hidden neurons, then selects a fairness-improving method accordingly. It estimates causal effects on fairness and uses their proportions and distribution to guide that choice.
4 AN ADAPTIVE APPROACH
- The adaptive approach estimates each attribute’s and neuron’s causal responsibility for unfairness, then chooses a fairness-improving method based on where responsibility lies and how concentrated it is.The approach characterizes causal-effect distributions using the proportion of effects above the fairness baseline and their coefficient of variation.Source: §4 · p. 6, §4.2 · pp. 7–8
4.1 Causality Analysis
- The analysis treats the neural network as a structured causal model and uses ACE, approximated by average interventional expectation, to assess how attributes and neurons affect fairness.Categorical input features are intervened on at every training-set value; neuron interventions use values sampled across the observed range.Source: §4.1 · p. 7
4.2 Adaptive Fairness Improvement
- When fewer than 10% of both attributes and neurons exceed the fairness baseline, the approach selects post-processing; otherwise, it selects pre-processing if CV_f > CV_n and in-processing otherwise.It prefers RW over DIR for pre-processing and chooses in-processing or post-processing methods for fairness improvement with the least accuracy cost.Source: §4.2 · p. 8
- For Adult Income gender fairness, 3 attributes (25%) and 33 neurons (26.6%) exceeded the 0.249 baseline, and CV_n 0.152 > CV_f 0.041 led to in-processing.Figure 4 plots attribute effects as black dots and neuron effects in layer-specific colors against the baseline.Source: Figure 4 · p. 8, §4.2 · p. 8
5 IMPLEMENTATION AND EVALUATION
The adaptive approach selects fairness-improvement methods from causal-effect distributions across attributes and neurons. It improves group fairness in all 10 cases and reduces individual discrimination in 9 of 10 while keeping accuracy loss low.
5 IMPLEMENTATION AND EVALUATION
- Causal-effect responsibility distributions varied across model-attribute combinations, and the group-fairness analysis used them to select processing categories.For COMPAS gender+race, higher CV among causal neurons led to an in-processing choice; several other group-fairness combinations led to pre-processing.Source: §5 · p. 9, Table 4 · p. 9
- No post-processing method was selected because every tested network had sufficiently many responsible neurons or attributes.Post-processing is selected only when both proportions are low, since it can significantly harm model performance.Source: §5 · p. 9
- For individual discrimination, Algorithm 3 selected in-processing except for Credit-gender and Bank-age; some COMPAS attributes had 0% high-causal-effect attributes.The paper interprets those zero proportions as meaning no input attribute is responsible for the individual discrimination.Source: Table 5 · p. 9, §5 · p. 9
- 7/10 cases had the best group-fairness improvement; when accuracy loss was included, the approach performed best in every case.For COMPAS, post-processing often reduced group-fairness metrics while accuracy fell by more than 30%.Source: §5 · p. 9, Figure 5 · p. 10
- 8/10 cases had the best individual-discrimination reduction minus accuracy loss, except Credit-gender and Bank-age, where the approach selected RW.RW left CDS nearly unchanged for Credit-gender and worsened it by around 2% for Bank-age.Source: §5 · pp. 9–10
- 10/10 cases achieved improved group fairness and 9/10 reduced individual discrimination, with low accuracy loss.The paper reports these as results of the adaptive approach.Source: §5 · p. 10
- Causality analysis took less than 10 minutes in every case, as additional time before applying the selected fairness method.Source: §5 · p. 10
6 THREATS TO VALIDITY
The paper identifies limitations in supported model structures, fairness metrics, causal-effect measurement, and robustness to distribution shifts. Its shift-stability evaluation covers slight shifts on Adult Income.
Limited model structures
- The current implementation supports feed-forward networks for tabular data and convolutional networks for images; RNN support would require extending causality analysis for feedback loops.Source: §6 · p. 10
Limited fairness metrics
- Evaluation uses SPD and GDS for group fairness and CDS for individual fairness, while the authors suggest possible applicability to related positive-classification-rate metrics.Source: §6 · p. 10
Causal effect measurement
- ACE is the causal-effect measure used; alternatives such as integrated gradients and gradients of causal effect may be sensitive and induce effects from other input features.Source: §6 · pp. 10–11
Distributional shift in the data
- The approach might be affected by distribution shifts, though tests on Adult Income with random splits and perturbed data found robustness to the evaluated slight shifts.Source: §6 · p. 11
7 RELATED WORK
Related work covers fairness testing and verification alongside empirical studies of fairness mitigation. The cited work includes discrimination measurement, comparisons of mitigation methods, and analyses of fairness-performance trade-offs.
Fairness Testing and Verification
- Fairness-testing work measures discrimination using multiple scores, protected attributes, generated samples, or input significance, while verification work checks fairness measures under data distributions.Examples include THEMIS, which measures group and causal discrimination for race and gender separately [21].Source: §7 · p. 11
Empirical Studies of Fairness
- Empirical studies compare mitigation effectiveness and performance trade-offs across models, techniques, attributes, and fairness settings.Examples include evaluations across eight real-world models and seven mitigation techniques [8], and tests of rich subgroup fairness versus accuracy [32].Source: §7 · p. 11
8 CONCLUSION
The paper evaluates existing fairness-improvement methods and proposes adaptive method selection based on causal responsibility across attributes and neurons. Its evaluation reports that the approach effectively selects an optimal improvement method.
8 CONCLUSION
- The empirical study evaluates 9 fairness-improvement methods across 4 real-world datasets, 90 model-attribute combinations, and 3 fairness metrics.Source: §8 · p. 11
- Existing methods are not always effective for group fairness and are often ineffective at reducing individual discrimination; the study also examines accuracy costs.Source: §8 · p. 11
- The proposed lightweight approach selects a fairness-improvement method adaptively from the distribution of responsible attributes and neurons, and the evaluation finds it effective at choosing an optimal method.Source: §8 · p. 11