PyImbalReg

Pre-processing techniques for imbalanced datasets in regression

---

Dealing with imbalanced datasets for regression

Your trained regression model has a heteroskedasticity problem? It can't predict extreme values well? These pre-processing techniques can help when the issues are caused by an imbalanced target distribution.

• Random Oversampling (RO)
• Gaussian Noise and Undersampling (GN)
• Weighted Relevance-based Combination Strategy (WERCS)

---

How to use (2-minute read)

1. Pass your data as a pandas DataFrame to any of the techniques.
2. Define a relevance function that maps the target variable to [0, 1] (higher value = rarer samples).
3. Set a threshold to flag rare vs normal samples.
4. Set method-specific parameters (e.g. oversampling/undersampling ratios).
5. Call .get() to obtain the resampled dataset.

---

Installation

From PyPI (pip)

pip install PyImbalReg

From PyPI with uv

uv pip install PyImbalReg

From GitHub

pip install git+https://github.com/vd1371/PyImbalReg.git

Development (uv)

This project uses uv for fast, reliable dependency management.

1. Install uv (if needed):

   curl -LsSf https://astral.sh/uv/install.sh | sh
   # or: pip install uv

2. Clone and sync:

   git clone https://github.com/vd1371/PyImbalReg.git
   cd PyImbalReg
   uv sync

   This creates a virtual environment, installs the package in editable mode, and installs dev dependencies (e.g. pytest, seaborn).

3. Run tests:

   uv sync --extra dev
   uv run pytest tests/

4. Run examples (requires dev deps for seaborn/matplotlib):

   uv run python examples/ro.py
   uv run python examples/gn.py
   uv run python examples/gnhf.py
   uv run python examples/wercs.py

5. Lock dependencies (optional):

   uv lock

   Commit uv.lock for reproducible installs.

---

Example

import PyImbalReg as pir
from seaborn import load_dataset

data = load_dataset("dots")

ro = pir.RandomOversampling(
    df=data,
    rel_func="default",
    threshold=0.7,
    o_percentage=5,  # (o_percentage - 1) × n_rare_samples will be added
)
new_data = ro.get()

---

Requirements

• Python ≥ 3.8
• NumPy
• Pandas
• SciPy

(All are declared in pyproject.toml and installed automatically with the package.)

---

More examples

Runnable scripts (use uv run python examples/<name>.py after uv sync --extra dev):

• Random Oversampling
• Gaussian Noise
• GNHF (Gaussian Noise + Histogram Frequencies)
• WERCS

---

Contributing

Issues, new techniques, and pull requests are welcome.

---

Citation

If you use this repository, please cite:

Branco, P., Torgo, L. and Ribeiro, R.P., 2019.
Pre-processing approaches for imbalanced distributions in regression.
Neurocomputing, 343, pp.76–99.

---

License

© Vahid Asghari, 2020. Licensed under the GNU General Public License v3.0 (GPLv3).

Some parts of the README and code were inspired by smogn.