pyppur 0.4.0

🪈 pyppur: Python Projection Pursuit Unsupervised Reduction

Overview

pyppur is a Python package that implements projection pursuit methods for dimensionality reduction. Unlike traditional methods such as PCA, pyppur focuses on finding interesting non-linear projections by minimizing either reconstruction loss or distance distortion.

Installation

pip install pyppur

Features

• Two optimization objectives:
  • Distance Distortion: Preserves pairwise distances between data points
  • Reconstruction: Minimizes reconstruction error using ridge functions
• Multiple initialization strategies (PCA-based and random)
• Full scikit-learn compatible API
• Supports standardization and custom weighting

Usage

Basic Example

import numpy as np
from pyppur import ProjectionPursuit, Objective
from sklearn.datasets import load_digits

# Load data
digits = load_digits()
X = digits.data
y = digits.target

# Projection pursuit with distance distortion
pp_dist = ProjectionPursuit(
    n_components=2,
    objective=Objective.DISTANCE_DISTORTION,
    alpha=1.5,  # Steepness of the ridge function
    n_init=3,   # Number of random initializations
    verbose=True
)

# Fit and transform
X_transformed = pp_dist.fit_transform(X)

# Projection pursuit with reconstruction loss (tied weights)
pp_recon_tied = ProjectionPursuit(
    n_components=2,
    objective=Objective.RECONSTRUCTION,
    alpha=1.0,
    tied_weights=True
)

# Projection pursuit with reconstruction loss (free decoder)
pp_recon_free = ProjectionPursuit(
    n_components=2,
    objective=Objective.RECONSTRUCTION,
    alpha=1.0,
    tied_weights=False,
    l2_reg=0.01
)

# Fit and transform
X_transformed_recon_tied = pp_recon_tied.fit_transform(X)
X_transformed_recon_free = pp_recon_free.fit_transform(X)

# Evaluate the methods
dist_metrics = pp_dist.evaluate(X, y)
recon_tied_metrics = pp_recon_tied.evaluate(X, y)
recon_free_metrics = pp_recon_free.evaluate(X, y)

print("Distance distortion method:")
print(f"  Trustworthiness: {dist_metrics['trustworthiness']:.4f}")
print(f"  Silhouette: {dist_metrics['silhouette']:.4f}")
print(f"  Distance distortion: {dist_metrics['distance_distortion']:.4f}")
print(f"  Reconstruction error: {dist_metrics['reconstruction_error']:.4f}")

print("\nReconstruction method (tied weights):")
print(f"  Trustworthiness: {recon_tied_metrics['trustworthiness']:.4f}")
print(f"  Silhouette: {recon_tied_metrics['silhouette']:.4f}")
print(f"  Distance distortion: {recon_tied_metrics['distance_distortion']:.4f}")
print(f"  Reconstruction error: {recon_tied_metrics['reconstruction_error']:.4f}")

print("\nReconstruction method (free decoder):")
print(f"  Trustworthiness: {recon_free_metrics['trustworthiness']:.4f}")
print(f"  Silhouette: {recon_free_metrics['silhouette']:.4f}")
print(f"  Distance distortion: {recon_free_metrics['distance_distortion']:.4f}")
print(f"  Reconstruction error: {recon_free_metrics['reconstruction_error']:.4f}")

API Reference

The main class in pyppur is ProjectionPursuit, which provides the following methods:

• fit(X): Fit the model to data
• transform(X): Apply dimensionality reduction to new data
• fit_transform(X): Fit the model and transform data
• reconstruct(X): Reconstruct data from projections
• reconstruction_error(X): Compute reconstruction error
• distance_distortion(X): Compute distance distortion
• compute_trustworthiness(X, n_neighbors): Measure how well local structure is preserved
• compute_silhouette(X, labels): Measure how well clusters are separated
• evaluate(X, labels, n_neighbors): Compute all evaluation metrics at once

Theory

Projection pursuit finds interesting low-dimensional projections of multivariate data. When used for dimensionality reduction, it aims to optimize an "interestingness" index which can be:

1. Distance Distortion: Minimizes the difference between pairwise distances in original and projected spaces (optionally with nonlinearity)
2. Reconstruction Error: Minimizes the error when reconstructing the data using ridge functions

Mathematical Formulations

Tied-Weights Ridge Autoencoder (Default)

Z = g(X A^T)
X̂ = Z A

Free Decoder Ridge Autoencoder (Available with tied_weights=False)

Z = g(X A^T)  
X̂ = Z B

Where:

• X is the input data matrix (n_samples × n_features)
• A are the encoder projection directions (n_components × n_features)
• B are the decoder weights (n_components × n_features, when untied)
• g(z) = tanh(α * z) is the ridge function with steepness parameter α
• Z is the projected data (n_samples × n_components)
• X̂ is the reconstructed data

Distance Distortion Options

• With nonlinearity: Compares distances between original space and g(X A^T)
• Without nonlinearity: Compares distances between original space and linear projections X A^T

Requirements

• Python 3.10+
• NumPy (>=1.20.0)
• SciPy (>=1.7.0)
• scikit-learn (>=1.0.0)
• matplotlib (>=3.3.0)

License

MIT

Citation

If you use pyppur in your research, please cite it as:

@software{pyppur,
  author = {Gaurav Sood},
  title = {pyppur: Python Projection Pursuit Unsupervised Reduction},
  url = {https://github.com/gojiplus/pyppur},
  version = {0.2.0},
  year = {2025},
}

🔗 Adjacent Repositories

• gojiplus/get-weather-data — Get weather data for a list of zip codes for a range of dates
• gojiplus/text-as-data — Pipeline for Analyzing Text Data: Acquire, Preprocess, Analyze
• gojiplus/calibre — Advanced Calibration Models
• gojiplus/skiplist_join
• gojiplus/rmcp — R MCP Server