OversampleQA

A diagnostic toolkit to validate, audit, and benchmark oversampling methods for imbalanced classification.

What It Does

• Validates oversampling quality with a hidden-majority error rate (binary and multiclass).
• Offers a broad set of distance metrics (Hassanat, Euclidean, Mahalanobis, etc.).
• Includes optimized and memory-efficient distance matrix computation.
• Supports benchmarking across datasets and oversamplers with exportable results.
• Provides a rich CLI with profiles, templates, shell completion, and diagnostics.
• Extensible via a plugin system for custom metrics and validators.

Concepts

OversampleQA validates synthetic samples by hiding a portion of the majority class and asking whether generated points look more like the hidden majority or the real minority. Each synthetic sample is scored by its nearest-neighbor distance to both groups using a chosen metric. If a synthetic sample is closer to the hidden majority than to the minority, it is counted as an error. The resulting error rate is a direct signal of how often oversampling produces majority-like artifacts. Lower error rates suggest better minority fidelity, but the absolute value depends on the dataset, metric, and hidden ratio. For multiclass data, the same idea generalizes to a confusion-style error matrix across classes.

Install

Python 3.10+ is required. Install the latest release from PyPI:

pip install oversampleqa

For the optional performance helpers:

pip install "oversampleqa[performance]"

For development or unreleased changes, install from source:

git clone https://github.com/diogoribeiro7/OversampleQA.git
cd OversampleQA
poetry install

To depend on the current repository state from another project:

pip install git+https://github.com/diogoribeiro7/OversampleQA.git

Quick Start (Python)

python - <<'PY'
from sklearn.datasets import make_classification
from imblearn.over_sampling import SMOTE
from oversampleqa import validate_oversampling

X, y = make_classification(
    n_samples=1000,
    n_features=20,
    n_informative=10,
    n_redundant=10,
    n_clusters_per_class=1,
    weights=[0.9, 0.1],
    random_state=42,
)

error_rate = validate_oversampling(
    X=X,
    y=y,
    minority_label=1,
    oversampler=SMOTE(random_state=42),
    hidden_ratio=0.1,
    metric="hassanat",
)

print(f"SMOTE validation error rate: {error_rate:.3f}")
PY

CLI

oversampleqa --help

oversampleqa validate data.csv \
  --target target \
  --minority-label 1 \
  --oversampler SMOTE \
  --metric hassanat \
  --hidden-ratio 0.1 \
  --export json \
  --output runs

oversampleqa profiles
oversampleqa template --template production -o oversampleqa.yaml
oversampleqa benchmark --output benchmark_results
oversampleqa doctor

Legacy minimal CLI (if you prefer a smaller surface):

oversampleqa-validate --help

Configuration

The enhanced CLI loads configuration from ~/.oversampleqa/config.yaml by default. You can override it with --config and select profiles with --profile.

Examples And Docs

• Code samples live in examples/ and tutorials/.
• Sphinx documentation sources are in docs/.

Build docs:

make docs

Development

# One-liner
make setup

# Or run onboarding helper
poetry run python scripts/onboard.py

# Manual steps
poetry install
poetry run pre-commit install
poetry run pre-commit install --hook-type commit-msg

Quality Checks

# Lint and typecheck -- both pass clean
make lint typecheck
# Run tests with coverage
make coverage
# Security audit
make security
# Build docs the way CI does, with warnings as errors
make docs
# Full pre-commit suite
poetry run pre-commit run --all-files

Linting and formatting are handled by ruff alone; the enforced rule set lives in pyproject.toml. These commands work on Windows and Linux alike.

Citation

If you use OversampleQA in academic work, please cite it. Machine-readable metadata lives in CITATION.cff, which GitHub renders as a "Cite this repository" button.

@software{ribeiro_oversampleqa,
  author  = {Ribeiro, Diogo},
  title   = {{OversampleQA: a diagnostic toolkit to validate, audit,
             and benchmark oversampling methods}},
  version = {0.5.1},
  year    = {2026},
  doi     = {10.5281/zenodo.21940361},
  url     = {https://doi.org/10.5281/zenodo.21940361}
}

The DOI above is the concept DOI: it always resolves to the newest archived version. To cite an exact archived release, use the version DOI listed in CITATION.cff after Zenodo has minted it. See Citing OversampleQA for the release and DOI sequence.

License

MIT. See LICENSE.