clustertk

ClusterTK

A comprehensive Python toolkit for cluster analysis with full pipeline support.

ClusterTK provides a complete, sklearn-style pipeline for clustering: from raw data preprocessing to cluster interpretation and export. Perfect for data analysts who want powerful clustering without writing hundreds of lines of code.

Features

• 🔄 Complete Pipeline - One-line solution from raw data to insights
• 📊 Multiple Algorithms - K-Means, GMM, Hierarchical, DBSCAN, HDBSCAN
• 🎯 Auto-Optimization - Automatic optimal cluster number selection
• 🧮 Smart Dimensionality Reduction - PCA/UMAP/None with algorithm-specific auto-mode
• 🎯 Feature Selection - Find optimal feature subsets for better clustering (NEW in v0.16.0!)
• 🎨 Rich Visualization - Beautiful plots (optional dependency)
• 📁 Export & Reports - CSV, JSON, HTML reports with embedded plots
• 💾 Save/Load - Persist and reload fitted pipelines
• 🔍 Interpretation - Profiling, naming, and feature importance analysis

Quick Start

Installation

# Core functionality
pip install clustertk

# With visualization
pip install clustertk[viz]

Basic Usage

import pandas as pd
from clustertk import ClusterAnalysisPipeline

# Load data
df = pd.read_csv('your_data.csv')

# Create and fit pipeline
pipeline = ClusterAnalysisPipeline(
    dim_reduction='auto',      # Smart selection (PCA/UMAP/None based on algorithm)
    handle_missing='median',
    correlation_threshold=0.85,
    n_clusters=None,           # Auto-detect optimal number
    verbose=True
)

pipeline.fit(df, feature_columns=['feature1', 'feature2', 'feature3'])

# Get results
labels = pipeline.labels_
profiles = pipeline.cluster_profiles_
metrics = pipeline.metrics_

print(f"Found {pipeline.n_clusters_} clusters")
print(f"Silhouette score: {metrics['silhouette']:.3f}")

# Export
pipeline.export_results('results.csv')
pipeline.export_report('report.html')

# Visualize (requires clustertk[viz])
pipeline.plot_clusters_2d()
pipeline.plot_cluster_heatmap()

Documentation

• Installation Guide - Detailed installation instructions
• Quick Start - Get started in 5 minutes
• User Guide - Complete component documentation
  • Preprocessing
  • Feature Selection
  • Clustering
  • Evaluation
  • Interpretation - Profiles, naming, feature importance
  • Visualization
  • Export
• Examples - Real-world use cases
• FAQ - Common questions

Pipeline Workflow

Raw Data → Preprocessing → Feature Selection → Dimensionality Reduction
→ Clustering → Evaluation → Interpretation → Export

Each step is configurable through pipeline parameters or can be run independently.

Key Capabilities

Preprocessing

• Missing value handling (median/mean/drop)
• Univariate outlier handling (winsorize/robust/clip/remove)
• Multivariate outlier detection (IsolationForest/LOF/EllipticEnvelope)
• Automatic scaling (robust/standard/minmax)
• Skewness transformation

Dimensionality Reduction (v0.15.0+)

• Auto-mode - Smart selection based on algorithm + data
• PCA - Linear, preserves global structure (best for K-Means/GMM)
• UMAP - Non-linear, preserves local density (best for HDBSCAN/DBSCAN)
• None - Work in original feature space (low-dimensional data)

Algorithm	Features	Auto Selection
K-Means/GMM	<50	None
K-Means/GMM	≥50	PCA
HDBSCAN/DBSCAN	<30	None
HDBSCAN/DBSCAN	≥30	UMAP

Clustering Algorithms

• K-Means - Fast, spherical clusters
• GMM - Probabilistic, elliptical clusters
• Hierarchical - Dendrograms, hierarchical structure
• DBSCAN - Density-based, arbitrary shapes
• HDBSCAN - Advanced density-based, varying densities (v0.8.0+)

Evaluation & Interpretation

• Silhouette score, Calinski-Harabasz, Davies-Bouldin metrics
• Automatic optimal k selection
• Cluster profiling and automatic naming
• Feature importance analysis (v0.9.0+)
  • Permutation importance
  • Feature contribution (variance ratio)
  • SHAP values (optional)

Export & Reports

• CSV export (data + labels)
• JSON export (metadata + profiles)
• HTML reports with embedded visualizations
• Pipeline serialization (save/load)

Examples

HDBSCAN with UMAP (v0.15.0+)

# Perfect for high-dimensional density-based clustering
pipeline = ClusterAnalysisPipeline(
    dim_reduction='umap',         # Preserves local density
    umap_n_components=10,          # NOT 2! For clustering, not viz
    clustering_algorithm='hdbscan'
)

pipeline.fit(high_dim_data)

# UMAP preserves density → HDBSCAN finds real clusters!
print(f"Found {pipeline.n_clusters_} clusters")
print(f"Noise ratio: {pipeline.cluster_profiles_.noise_ratio_:.1%}")

Why UMAP for HDBSCAN?

• PCA destroys local density → HDBSCAN finds only noise
• UMAP preserves local structure → HDBSCAN works correctly
• Auto-mode selects UMAP for HDBSCAN/DBSCAN when features >30

Important: Use n_components=10-20 for clustering, NOT 2-3 (visualization only)!

Feature Selection for Better Clustering (v0.16.0+)

# Problem: You have 30 features, but not all are useful for clustering
# More features ≠ better clustering (curse of dimensionality)

# Step 1: Fit on all features
pipeline = ClusterAnalysisPipeline(dim_reduction='pca')
pipeline.fit(df)  # 30 features → Silhouette: 0.42

# Step 2: Find which features matter most
importance = pipeline.get_pca_feature_importance()
print(importance.head(10))  # Top 10 features by PCA loadings

# Step 3: Try refitting with top 10 features
comparison = pipeline.refit_with_top_features(
    n_features=10,
    importance_method='permutation',  # Best for clustering quality
    compare_metrics=True,
    update_pipeline=False  # Just compare, don't update yet
)

# Step 4: If metrics improved, update pipeline
if comparison['metrics_improved']:
    print(f"Improvement: {comparison['weighted_improvement']:+.1%}")
    pipeline.refit_with_top_features(n_features=10, update_pipeline=True)
    # New silhouette: 0.58 (+38% improvement!)

Why Feature Selection?

• Irrelevant features dilute clustering signal (noise)
• PCA can't fix bad features, only compress them
• 10 good features > 30 mixed features

Three Importance Methods:

• 'permutation' - Best for clustering quality (default)
• 'contribution' - Variance ratio analysis
• 'pca' - PCA loadings (only if dim_reduction='pca')

Feature Importance Analysis

# Understand which features drive your clustering
results = pipeline.analyze_feature_importance(method='all')

# View permutation importance
print(results['permutation'].head())

# View feature contribution (variance ratio)
print(results['contribution'].head())

# Use top features for focused analysis
top_features = results['permutation'].head(5)['feature'].tolist()

Algorithm Comparison

# Compare multiple algorithms automatically
results = pipeline.compare_algorithms(
    X=df,
    feature_columns=['feature1', 'feature2', 'feature3'],
    algorithms=['kmeans', 'gmm', 'hierarchical', 'dbscan'],
    n_clusters_range=(2, 8)
)

print(results['comparison'])  # DataFrame with metrics
print(f"Best algorithm: {results['best_algorithm']}")

# Visualize comparison
pipeline.plot_algorithm_comparison(results)

Customer Segmentation

pipeline = ClusterAnalysisPipeline(
    n_clusters=None,  # Auto-detect
    auto_name_clusters=True
)

pipeline.fit(customers_df,
            feature_columns=['age', 'income', 'purchases'],
            category_mapping={
                'demographics': ['age', 'income'],
                'behavior': ['purchases']
            })

pipeline.export_report('customer_segments.html')

Anomaly Detection

pipeline = ClusterAnalysisPipeline(
    clustering_algorithm='dbscan'
)

pipeline.fit(transactions_df)
anomalies = transactions_df[pipeline.labels_ == -1]

More examples: docs/examples.md

Requirements

• Python 3.8+
• numpy >= 1.20.0
• pandas >= 1.3.0
• scikit-learn >= 1.0.0
• scipy >= 1.7.0
• joblib >= 1.0.0

Optional (for visualization):

• matplotlib >= 3.4.0
• seaborn >= 0.11.0

Contributing

Contributions are welcome! Please check:

• GitHub Issues - Report bugs
• GitHub Discussions - Questions

License

MIT License - see LICENSE file for details.

Citation

If you use ClusterTK in your research, please cite:

@software{clustertk2024,
  author = {Veselov, Aleksey},
  title = {ClusterTK: A Comprehensive Python Toolkit for Cluster Analysis},
  year = {2024},
  url = {https://github.com/alexeiveselov92/clustertk}
}

Links

• PyPI: https://pypi.org/project/clustertk/
• GitHub: https://github.com/alexeiveselov92/clustertk
• Documentation: docs/
• Author: Aleksey Veselov ([e-mail])

Made with ❤️ for the data science community