View Report

Classifying Stars, Galaxies, and Quasars: An End-to-End ML Project

Project Summary

Built a production-ready machine learning pipeline to automatically classify 30,000 celestial objects from astronomical survey data, achieving 97% accuracy through data preprocessing, feature engineering, and model optimization.

Tech Stack: Python, scikit-learn, pandas, NumPy, matplotlib
Skills: Data Cleaning, Feature Engineering, Model Selection, Cross-Validation, Performance Optimization

Business Problem

Astronomical surveys collect millions of observations that need classification. Manual review is impractical at scale. This project demonstrates how I built an automated classifier while handling real-world data challenges: class imbalance, duplicate records, outliers, and high dimensionality.

What I Did

1. Data Quality & Cleaning

Challenges in the Dataset:

• 2,569 duplicate entries (8.5% of dataset)
• Severe class imbalance: 59% galaxies, 22% stars, 19% quasars
• Irrelevant features with constant values
• Extreme outliers in critical features

My Approach:

• Removed duplicates using unique identifiers
• Selected appropriate evaluation metrics (Balanced Accuracy, Weighted F1) to handle imbalance
• Dropped 6 irrelevant features after exploratory analysis
• Applied domain-appropriate transformations to skewed distributions

Key Decision: Applied log transformation to redshift feature after discovering it was severely right-skewed. This single transformation significantly improved model performance by normalizing the most discriminative feature in the dataset.

2. Feature Engineering & Selection

Discovered Strong Correlations:

• Photometric features (u, g, r, i, z bands) showed 0.86-0.97 correlations
• Redshift emerged as the dominant predictor (0.756 mutual information score)

Applied Two Feature Selection Methods:

Filter Method (Mutual Information): Fast, model-agnostic ranking identified top 3 features: redshift, plate, z-band

Wrapper Method (Sequential Forward Selection): Tested feature subsets specific to each classifier, revealing that different models benefit from different feature combinations

Impact of Feature Selection: Reduced features from 10 to 3 with only 2% accuracy loss, improving training time for production deployment and experimentation.

3. Model Development & Evaluation

Systematic Comparison of 6 Classifiers:

Evaluation Strategy:

• 5-fold cross-validation for accurate performance estimates
• Multiple metrics to handle class imbalance
• Learning curve analysis
• ROC curve analysis for per-class performance insights

Findings: Linear and RBF SVMs achieved identical 97% accuracy, but Linear SVM trains 20% faster—making it the optimal choice for production.

4. Model Interpretation & Insights

ROC Analysis Revealed:

• Near-perfect separation (>99% AUC) between stars and other objects
• Main classification challenge: distinguishing galaxies from quasars at intermediate redshift values
• This insight points to specific areas for future improvement

Achievements

97% classification accuracy on imbalanced dataset
67% dimensionality reduction while maintaining performance
Identified optimal model through model comparisons Production-ready pipeline with proper cross-validation
Documented trade-offs between accuracy, speed, and complexity

Technical Highlights

Data Preprocessing:

• Handled missing data and duplicates
• Applied appropriate scaling (StandardScaler for normal distributions, MinMaxScaler for bounded features)
• Domain-specific transformations (log-scaling for redshift)

Model Selection:

• Compared 6 different algorithms with consistent evaluation framework
• Used stratified k-fold cross-validation to ensure reliable results
• Analyzed learning curves to optimize training set size (saved compute resources)

Feature Engineering:

• Both filter and wrapper methods for comprehensive feature analysis
• Discovered redshift as key discriminative feature
• Identified multi-collinearity in photometric features

Evaluation:

• Multiple metrics appropriate for imbalanced classification
• One-vs-one ROC analysis for multi-class insights
• Confusion matrix analysis to identify specific error patterns

What I Learned

Skills:

• Advanced feature engineering techniques for imbalanced datasets
• Trade-offs between model complexity, accuracy, and training time
• How to select appropriate evaluation metrics based on business context
• Importance of systematic model comparison with proper cross-validation

Practical Insights:

• Sometimes simple transformations (like log-scaling) have dramatic impact
• Feature selection isn't just about accuracy—it's about production feasibility
• Understanding where models fail is as valuable as overall performance metrics
• Default hyperparameters often perform surprisingly well with proper preprocessing

Future Improvements

If continuing this project, I would:

1. Engineer composite features combining photometric measurements to better separate galaxies from quasars
2. Implement ensemble methods to combine predictions from multiple models
3. Deploy model as REST API for real-time classification
4. Add automated retraining pipeline to handle data drift
5. Create interactive dashboard for non-technical stakeholders