mlforget 0.1.0

Machine unlearning methods for energy consumption prediction

Machine Unlearning Library

This library provides modular implementations of machine unlearning techniques for time-series data. It is designed for use in research, evaluation, and deployment scenarios where data privacy and the right to be forgotten are important.

Features

• Support for Exact Retraining, SISA, and Knowledge Distillation as unlearning strategies
• LSTM-based model architecture for building energy consumption prediction
• Modular preprocessing, training, evaluation, and unlearning components
• Support for measuring carbon emissions via CodeCarbon
• Designed as a reusable, extendable Python package

---

Project Structure

├── data_utils.py         # Data loading, preprocessing, and forget set selection
├── model_utils.py        # LSTM model definition
├── training.py           # Training loops and model evaluation
├── evaluation.py         # Metrics: RMSE, MAE, R²
├── unlearning.py         # Implementations of unlearning methods
├── workflow.py           # High-level run_* functions for each unlearning method
├── __init__.py           # Public API

---

Getting Started

Installation

Clone the repository and make sure dependencies are installed:

pip install torch pandas scikit-learn

---

Example Usage

from Machine_Unlearning_Tool import (
    load_dataset, preprocess_data, run_sisa_unlearning
)

# Load and preprocess data
df = load_dataset("data/train.csv")
config = { ... }  # preprocessing configuration
df = preprocess_data(df, config)

# Convert to tensors and run SISA
X = ...
y = ...
results = run_sisa_unlearning(X, y, df, input_cols, target_col, id_col, forget_ids, device)

---

Public API Overview

load_dataset(filepath: str) -> pd.DataFrame

Loads a dataset from CSV.

preprocess_data(df: pd.DataFrame, config: dict) -> pd.DataFrame

Preprocesses the dataset using normalization and categorical encoding.

train_model(model, ...)

Trains a PyTorch model on provided data.

run_sisa_unlearning(...)

Performs SISA unlearning: slices and shards the dataset, retrains on retained data.

run_exact_retraining(...)

Performs full retraining on retained data.

run_knowledge_distillation(...)

Trains a student model using teacher model outputs while excluding forget set.

evaluate_on_loader(model, loader, loss_fn, device)

Computes RMSE, MAE, and R² metrics on a test or validation loader.

---

Output

Each run_* method returns:

• Trained model or model dictionary
• Pre- and post-unlearning performance metrics
• (Optional) CodeCarbon energy emission data

---

Notes

• Forget set is defined using user-specified id_column and list of IDs
• Can be extended with new models or additional unlearning methods
• Suitable for research on GDPR-compliant model behavior

---

License

No License specified.