Beyond the Seeds: Fairness Testing via Counterfactual Analysis of Non-Seed Instances
Bibliographic record
Abstract
As machine learning software increasingly shapes crucial decisions in our daily lives, ensuring the fairness of these decisions is paramount. Individual fairness guarantees non-discrimination based on protected attributes, such as race or gender. Discriminatory instances reveal individual discrimination included in machine learning software. Existing methods for detecting individual discrimination often rely on initial “seed” instances, which are data points selected from the dataset that have more likelihood of exhibiting discrimination. These seed instances are then used as the basis for generating more discriminatory instances. While effective, this approach may inadvertently overlook discrimination embedded within seemingly fair non-seed instances. To overcome this limitation, this paper proposes FairBS, an approach that utilizes non-seed instances to generate discriminatory instances using counterfactual analysis. FairBS first constructs an explainer based on dataset input features and a model under test. It then generates an input instance and checks it for discrimination. If the input instance is non-discriminatory, FairBS uses the explainer to create counterfactual examples of that instance, by causing minimal perturbation to its feature values, which then produce other instances with opposite predictions. Extensive experiments on five datasets and five machine learning models demonstrate that our proposed approach outperforms state-of-the-art methods in both efficiency and effectiveness across all datasets and models. Our approach generates an average of$\times 13.44$more discriminatory instances at$\times 14.51$faster speed compared to existing seed-based methods. These findings indicate that FairBS expands the boundaries of fairness testing beyond the discriminatory seed instances, providing a powerful tool that can be used by software engineers to better ensure fairness in machine learning software.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.077 | 0.266 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.003 | 0.002 |
| Science and technology studies | 0.002 | 0.006 |
| Scholarly communication | 0.004 | 0.007 |
| Open science | 0.005 | 0.005 |
| Research integrity | 0.003 | 0.005 |
| Insufficient payload (model declined to judge) | 0.002 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".