Self-Organizing Maps (SOM)

Python Example Using MiniSom

We’ll use the MiniSom package for a basic SOM implementation.

Install Library

pip install minisom

Sample Code

from minisom import MiniSom
from sklearn.datasets import load_iris
from sklearn.preprocessing import MinMaxScaler
import matplotlib.pyplot as plt
import numpy as np

# Load and normalize data
iris = load_iris()
X = iris.data
scaler = MinMaxScaler()
X_scaled = scaler.fit_transform(X)

# Initialize SOM: 7x7 grid, input_len = 4 (features)
som = MiniSom(x=7, y=7, input_len=4, sigma=1.0, learning_rate=0.5)
som.random_weights_init(X_scaled)
som.train_random(X_scaled, 100)

# Visualize SOM distance map (U-Matrix)
plt.figure(figsize=(7, 7))
plt.pcolor(som.distance_map().T, cmap='coolwarm')  # distance map as heatmap
plt.colorbar()
plt.title("SOM - Distance Map (U-Matrix)")
plt.show()

Output-

Output Explanation

• Brighter cells indicate greater distance (edges between clusters)
• Darker cells represent more similarity (cluster centers)

Applications of SOM

• Customer segmentation
• Anomaly detection
• Document or text clustering
• Gene expression pattern analysis
• Image compression

Advantages

• Intuitive data visualization
• Preserves data topology
• Works well with unlabeled, high-dimensional data
• Identifies patterns/clusters without supervision

Limitations

• SOMs require careful tuning of map size and parameters
• Training is relatively slow for large datasets
• Interpretation can be less precise than traditional clustering models

Tips for Beginners

• Always normalize data before training a SOM
• Use the U-Matrix to visually interpret cluster boundaries
• Start with small grid sizes and scale as needed

Tips for Professionals

• Use SOM as a preprocessing step for classification or anomaly detection
• Combine with other clustering techniques like DBSCAN for hybrid modeling
• Customize color-mapped heatmaps for deeper insight
• Evaluate map quality using quantization and topographic errors

Summary

Self-Organizing Maps (SOM) are a powerful tool for unsupervised learning and visualization of complex data. They help you discover hidden structures and relationships, especially in high-dimensional datasets, making them a valuable tool for exploratory data analysis and clustering tasks.