jcorpac ML Unsupervised Clustering
Unsupervised learning finds hidden structures in data without explicit labels.
Core Techniques
- K-Means: Simple and effective for spherical clusters. Use the "Elbow Method" to find the optimal
k. - DBSCAN: Density-based clustering that can handle irregular shapes and noise.
- PCA (Principal Component Analysis): Reducing feature space while retaining variance. Essential for visualization and avoiding the "curse of dimensionality."
Evaluation Metrics
Since there are no labels, we use:
- Silhouette Score: Measuring cluster separation and cohesion.
- Inertia: Sum of squared distances to cluster centers.
Best Practices
- Feature Scaling: Essential for distance-based algorithms like K-Means.
- Interpretablity: Plot your results (e.g., using t-SNE or UMAP for high-dimensional data).