No description
  • Python 50.5%
  • C++ 36.5%
  • Shell 13%
Find a file
2026-10-04 20:41:07 +09:00
Baselines Anonymous artifact submission 2026-06-30 12:00:00 +00:00
Experiments Implement EnSensCount IDI completion experiments for RQ3 2026-10-04 20:41:07 +09:00
RegionFT Anonymous artifact submission 2026-06-30 12:00:00 +00:00
Results Anonymous artifact submission 2026-06-30 12:00:00 +00:00
TreeFT Separate RQ1 seed accounting from generation budget 2026-10-04 19:09:50 +09:00
.gitignore Configure TreeFT Conda environment and model dependencies 2026-10-01 23:27:30 +09:00
analyze_results.py Anonymous artifact submission 2026-06-30 12:00:00 +00:00
environment.yml Configure TreeFT Conda environment and model dependencies 2026-10-01 23:27:30 +09:00
README.md Configure TreeFT Conda environment and model dependencies 2026-10-01 23:27:30 +09:00
requirements.txt Implement sharded RQ1 experiments and result analysis 2026-10-04 18:37:58 +09:00
run_RQ1_baseline_comparison.py Implement sharded RQ1 experiments and result analysis 2026-10-04 18:37:58 +09:00
run_RQ2_ablation.py Implement shared-seed RQ2 ablation experiments 2026-10-04 19:37:16 +09:00
run_RQ2_RQ3_ablation.py Use tree ensemble models in experiment configurations 2026-10-02 00:15:28 +09:00
run_RQ3_ensenscount.py Implement EnSensCount IDI completion experiments for RQ3 2026-10-04 20:41:07 +09:00
runner.py Separate RQ1 seed accounting from generation budget 2026-10-04 19:09:50 +09:00

RegionFT

This is the anonymized code artifact for ICSE 2027 submission #1447, “Region-Guided Fairness Testing via Adaptive Partitioning.” RegionFT tests the individual fairness of black-box classifiers by partitioning their input space and guiding test generation toward regions with higher estimated discrimination risk.

Beyond implementing RegionFT, this artifact integrates representative black-box individual fairness testing methods through a common runner.py interface, enabling researchers to reproduce and compare them more easily.

Setup

The project uses Python 3.9. Conda is recommended:

conda env create --prefix ./env_treeft --file environment.yml
conda activate ./env_treeft

A standard virtual environment can also be used:

python3.9 -m venv env_treeft
source env_treeft/bin/activate
pip install -r requirements.txt

Run the following commands from the repository root.

Run a Single Experiment

runner.py runs one combination of dataset, protected attribute, classifier, and testing method. It supports the proposed RegionFT method and the representative methods AFT, ExpGA, GRFT, LIMI, SG, Themis, VBT, and VBT-X. To view all options, run:

python runner.py --help

The following command runs RegionFT for 60 seconds on the Adult/GBDT classifier with sex as the protected attribute:

python runner.py \
  --dataset Adult \
  --protected sex \
  --model GBDT \
  --method regionft \
  --runtime 60 \
  --show-log

When --output is omitted, runner.py selects the next unused example directory:

Results/runner_examples/run_001/regionft/
├── log/*.json
├── disc/*.npy
└── partition/*.csv

The output contains:

  • log/*.json: configuration, metrics, and timing.
  • disc/*.npy: discovered IDI pairs.
  • partition/*.csv: learned regions and their statistics.

A completed example is retained in Results/runner_examples/run_001.

Experiment Scripts

The following scripts run individual configurations, execute the main RQ experiments, and analyze their results.

Script Purpose Output
runner.py Run one experiment configuration. Results/runner_examples/run_001, run_002, ...
run_RQ1_baseline_comparison.py Run the RQ1 comparison of RegionFT and five baselines. Results/RQ1_baseline_comparison/
run_RQ2_RQ3_ablation.py Run the RQ2/RQ3 partition and sampling ablations. Results/RQ2_RQ3_ablation/
analyze_results.py Generate readable RQ CSVs from per-run results. Results/ by default

See Experiments/README.md for batch execution.

Experiment Results

The logs and readable CSVs used in the evaluation are retained under Results/. Logs from the main RQ1–RQ3 experiments are stored in Results/main_experiments/, while the corresponding per-run metrics and summary tables are stored directly under Results/.

See Results/README.md for the file organization and metric definitions.

Project Layout

The repository is organized into the following top-level directories.

Directory Contents
RegionFT/ Implementation of the proposed RegionFT method.
Baselines/ Baseline implementations and their sources.
Experiments/ Datasets, trained classifiers, and shared experiment code.
Results/ Retained logs, readable CSVs, and a complete runner example.