Unsupervised Machine Learning in Practice: Complete Technical Workflow for a Star Classification Project

This article presents a complete unsupervised machine learning project based on star data, covering processes such as data preparation, exploratory analysis, dimensionality reduction, anomaly detection, clustering analysis, and visualization evaluation. The project uses techniques like PCA/MDS dimensionality reduction, Isolation Forest anomaly detection, K-means/hierarchical clustering/OPTICS clustering, and Grid Search hyperparameter optimization to demonstrate the entire process of unsupervised learning from data to insights. Its methodology can be transferred to tasks like customer segmentation and document clustering.

Most real-world data has no labels. Unsupervised learning can discover the inherent structure of data and solve problems without clear answers, such as customer profiling and anomaly detection. This project takes star classification as a case study, based on the stars.csv dataset (containing physical features like star temperature and luminosity), and explores the natural grouping rules of stars without relying on predefined category labels.

Data preprocessing steps include missing value handling (deletion/filling), feature scaling (Z-score/Min-Max normalization), and feature engineering (log transformation, interaction features). Exploratory analysis uses descriptive statistics (mean, median), box plots, histograms, and scatter matrices to understand feature distributions and relationships.

• Dimensionality Reduction: Use PCA (linear, maximizes variance) and MDS (non-linear, preserves relative distances) to address the curse of dimensionality;
• Anomaly Detection: Adopt Isolation Forest (random partitioning, identifies anomalies via path length) to discover special celestial bodies;
• Clustering: K-means (preset K, select K using elbow method/silhouette coefficient), Hierarchical Clustering (agglomerative, supports different linkage criteria), OPTICS (density-based, identifies clusters of arbitrary shapes).

• Hyperparameter Optimization: Use Grid Search to exhaustively search parameter combinations, and evaluate clustering quality using metrics like silhouette coefficient and Calinski-Harabasz index;
• Result Visualization: Scatter plots after dimensionality reduction (with cluster labels), feature distribution plots (box plots/violin plots), clustering heatmaps (feature-cluster mean matrix);
• Interpretation: Combine astrophysical knowledge to map clusters to known types like main-sequence stars and giants, or discover new categories.

The project refines a transferable workflow: Data understanding and preparation → Dimensionality reduction exploration → Anomaly handling → Multi-algorithm attempts → Hyperparameter optimization → Result evaluation and interpretation → Visualization. This workflow applies to various unsupervised tasks, with the core being a human-machine collaboration model where algorithms discover structure and humans assign meaning.

• Limitations: No labels to verify clustering correctness, subjective algorithm selection, information loss in high-dimensional visualization;
• Improvement Directions: Semi-supervised learning (combining a small number of labels), deep clustering (autoencoders), ensemble clustering (consensus of multiple algorithms), interactive exploration tools.