MGP-Imputer: Missing Value Imputation with Deep Gaussian Processes

A PyTorch-based implementation of Missing Gaussian Processes (MGP) for missing value imputation, wrapped in a user-friendly scikit-learn compatible API.

This package allows you to seamlessly integrate Deep Gaussian Process models into your data preprocessing pipelines for robust and uncertainty-aware imputation. It is based on the paper "Gaussian processes for missing value imputation".

Chained Deep Gaussian Processes for Missing Value Imputation

Features

• Scikit-learn Compatible: Use fit, predict, and fit_transform methods just like any other scikit-learn transformer.
• Two Imputation Strategies:
  • chained (Default): Builds a separate GP layer for each feature with missing values, modeling dependencies in a chained fashion (MGP).
  • holistic: Builds a single, multi-output Deep GP to model all features simultaneously.
• Probabilistic Imputation: Returns both the imputed values and the standard deviation, giving you a measure of uncertainty for each imputed value.
• GPU Accelerated: Leverages PyTorch to run on CUDA devices for significant speedups.

Installation

You can install mgp-imputer directly from PyPI:

pip install mgp-imputer

Quick Start

Here's how to use MGPImputer to fill in missing values (np.nan) in your dataset.

import numpy as np
import pandas as pd
from mgp import MGPImputer

# 1. Create a synthetic dataset with 20% missing values
np.random.seed(42)
n_samples, n_features = 200, 5
X_true = np.random.rand(n_samples, n_features) * 10
X_missing = X_true.copy()
missing_mask = np.random.rand(n_samples, n_features) < 0.2
X_missing[missing_mask] = np.nan

print(f"Created a dataset with {np.sum(missing_mask)} missing values.")

# 2. Initialize the MGPImputer
# Strategies can be 'chained' (default) or 'holistic'
imputer = MGPImputer(
    imputation_strategy='chained',
    n_inducing_points=100,
    n_iterations=1000, # Use more iterations for real data
    learning_rate=0.01,
    batch_size=64,
    verbose=True,
    seed=42
)

# 3. Fit on the data and transform it to get imputed values
# The imputer returns the imputed data and the standard deviation of the predictions
X_imputed, X_std = imputer.fit_transform(X_missing)

# 4. Evaluate the imputation quality
rmse = np.sqrt(np.mean((X_imputed[missing_mask] - X_true[missing_mask])**2))
print(f"\nImputation complete.")
print(f"RMSE on missing values: {rmse:.4f}")

# The result is a complete numpy array
print("\nImputed data shape:", X_imputed.shape)
print("Number of NaNs in imputed data:", np.isnan(X_imputed).sum())

Configuration Options

You can customize the behavior of MGPImputer by passing parameters during initialization. Here are the available options:

Citation

If you use this work in your research, please cite the original paper:

Jafrasteh, B., Hernández-Lobato, D., Lubián-López, S. P., & Benavente-Fernández, I. (2023). Gaussian processes for missing value imputation. Knowledge-Based Systems, 273, 110603. Missing GPs

License

This project is licensed under the MIT License.