pyecsago 0.2.1

Python implementation of the evolutionary clustering ECSAGO - Evolutionary Clustering with Self Adaptive Genetic Operators

pyecsago

Python implementation of ECSAGO (Evolutionary Clustering with Self-Adaptive Genetic Operators), a robust evolutionary clustering algorithm that automatically discovers the number of clusters in data while being resistant to noise.

Features

• Automatic cluster detection — no need to specify k in advance
• Noise-robust density-based fitness function (RBF kernel weights)
• Self-adaptive operators via HAEA (Hybrid Adaptive Evolutionary Algorithm)
• Deterministic Crowding for niche maintenance
• MDE refinement (Maximal Density Estimator) for prototype center/spread convergence
• GPU acceleration — optional CUDA support via CuPy

Installation

CPU only:

pip install pyecsago

With CUDA support:

pip install pyecsago[cuda]

From source

git clone https://github.com/pwnaoj/pyecsago
cd pyecsago
pip install .

Quick start

import numpy as np
from pyecsago import ECSAGO

# Sample data: 3 Gaussian clusters
data = np.vstack([
    np.random.randn(100, 2) + [0, 0],
    np.random.randn(100, 2) + [5, 5],
    np.random.randn(100, 2) + [10, 0],
])

config = {
    "population_size": 100,
    "weight_threshold": 0.3,
    "max_generations": 30,
    "iterations": 10,
    "extraction_type": {2: 0.25},   # PROPORTION_MAX with 25% threshold
    "k": 13.8,
    "use_cuda": False,
}

ecsago = ECSAGO(config)
results = ecsago.run(data)

prototypes = results["refined_prototypes"]
labels = results["cluster_assignments"]

print(f"Clusters found: {len(prototypes)}")

Step-by-step usage

For finer control you can invoke each stage independently:

ecsago = ECSAGO(config)

# 1. Load data
ecsago.context.set_data(data)

# 2. Evolve the population
ecsago.evolve()

# 3. Extract prototypes
prototypes = ecsago.extract_prototypes(
    extraction_type={2: 0.25},
    k=13.8,
)

# 4. Refine with MDE
refined = ecsago.refine_prototypes(
    prototypes=prototypes,
    iterations=10,
    k=13.8,
)

Configuration

All parameters are passed as a dictionary to ECSAGO(config).

Parameter	Type	Description
population_size	int	Number of individuals in the evolutionary population
weight_threshold	float	Threshold for weight binarization (0–1)
max_generations	int	Maximum number of evolutionary generations
iterations	int	Number of MDE refinement iterations
extraction_type	dict	Extraction method — key is the type (0–4), value is the threshold
k	float	Chi-squared factor for minimum inter-prototype distance
use_cuda	bool	Enable CUDA/GPU acceleration via CuPy

Extraction types

Key	Method	Threshold meaning
0	ABSOLUTE_VALUE	Absolute fitness threshold (auto-calculated)
1	PROPORTION_AVG	Proportion of average fitness
2	PROPORTION_MAX	Proportion of maximum fitness
3	PROPORTION_MEDIAN	Proportion of median fitness
4	MINIMUM_DENSITY	Based on minimum density

Output

ecsago.run(data) returns a dictionary:

Key	Description
final_population	Full evolved population
prototypes	Extracted prototypes (before refinement)
refined_prototypes	Refined prototypes (after MDE)
cluster_assignments	Cluster label for each data point

Each prototype is an ECSAGOIndividual with attributes genome (center), sigma2 (spread), and fitness.

Architecture

pyecsago/
├── core/               # Abstract base classes and exceptions
├── implementations/
│   └── ecsago/         # ECSAGO algorithm, context, individual, population
├── strategies/
│   ├── evolution/      # Evolution strategy (HAEA + Deterministic Crowding)
│   ├── extraction/     # Prototype extraction (fitness, niche, composite)
│   ├── fitness/        # Fitness calculation (CPU and CUDA)
│   ├── niching/        # Deterministic Crowding
│   ├── operators/      # Genetic operators (mutation, crossover, HAEA)
│   └── refinement/     # MDE refinement (CPU and CUDA)
└── utils/              # Data type strategies, compatibility, utilities

License

pyecsago was created by Joan Sebastian Tamayo Rivera. It is licensed under the terms of the MIT license.

References

• León, E. "Scalable and Adaptive Evolutionary Clustering for Noisy and Dynamic Data"
• León, E., Nasraoui, O., & Gómez, J. "ECSAGO: Evolutionary Clustering with Self-Adaptive Genetic Operators"
• Gómez, J. "Self Adaptation of Operator Rates for Multimodal Optimization"
• Tamayo, J. "GPU/CUDA-Based Parallelization of the ECSAGO Evolutionary Algorithm"