🌟 Pulsar Star Classification using Machine Learning

An end-to-end machine learning pipeline to identify pulsar stars from radio telescope data, achieving 92% recall and minimizing false negatives to maximize astronomical discoveries.

---

📋 Table of Contents

• Overview
• Problem Statement
• Dataset
• Methodology
• Results
• Key Learnings
• Installation
• Usage
• Project Structure
• Technologies
• Future Improvements
• Contact

---

🔭 Overview

Pulsars are highly magnetized rotating neutron stars that emit beams of electromagnetic radiation. With only ~3,000 pulsars discovered out of an estimated 100,000 in our galaxy, automated detection is crucial for accelerating astronomical research.

This project builds a machine learning classifier to identify pulsar candidates from radio telescope data, reducing manual review workload by 98% while maintaining 92% detection accuracy.

🎯 Key Achievements

• ✅ 92% Recall - Successfully detects 284 out of 308 pulsars
• ✅ 24 False Negatives - Lowest miss rate across all tested models
• ✅ 96.68% Accuracy - Strong overall performance
• ✅ Handles Severe Class Imbalance - 91:9 negative-to-positive ratio

---

🎯 Problem Statement

Why This Matters

Pulsars provide insights into:

• Extreme physics (neutron stars, gravitational waves)
• Tests of general relativity
• Deep space navigation systems
• Understanding stellar evolution

The Challenge

1. Severe Class Imbalance: Only 9.16% of candidates are actual pulsars
2. High-Stakes Decisions: Missing a pulsar = lost scientific discovery
3. Asymmetric Cost: False negatives (missed discoveries) are worse than false positives (false alarms)

Solution Approach

Build a machine learning classifier that prioritizes recall (maximizing pulsar detection) while maintaining acceptable precision (minimizing wasted telescope follow-up time).

---

📊 Dataset

Source: HTRU2 Dataset from UCI Machine Learning Repository

Specifications

• Size: 17,898 observations
• Features: 8 continuous variables
• Target: Binary (0 = Not Pulsar, 1 = Pulsar)
• Class Distribution:
  • Not Pulsar: 16,259 (90.84%)
  • Pulsar: 1,639 (9.16%)
  • Imbalance Ratio: ~10:1

Features

Integrated Profile Statistics:

• Mean, Standard Deviation, Kurtosis, Skewness of integrated pulse profile

DM-SNR Curve Statistics:

• Mean, Standard Deviation, Kurtosis, Skewness of DM-SNR curve

---

🔬 Methodology

Phase 1: Exploratory Data Analysis

• Statistical analysis and distribution checks
• Correlation analysis and feature relationships
• Outlier detection and class imbalance assessment
• Visualization of feature separability

Key Finding: Kurtosis features capture pulse shape characteristics critical for classification.

Phase 2: Data Preprocessing

Outlier Handling

Decision: Retained outliers
Rationale: Pulsars are extreme objects; their "outlier" values represent real astronomical signals, not measurement errors.

Feature Scaling

Method: RobustScaler
Rationale: Uses median and IQR, unaffected by outliers (unlike StandardScaler which uses mean/std).

# Proper train-test split to prevent data leakage
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Fit scaler on training data only
scaler = RobustScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)  # Use training parameters

Class Imbalance Handling

Approaches Tested:

1. SMOTE (Synthetic Minority Over-sampling)

   • Created synthetic minority samples
   • Balanced training set to 50:50 ratio
   • Result: Good baseline performance

2. Class Weights (Final Choice) ✅

   • Used XGBoost's scale_pos_weight parameter
   • Trained on original imbalanced data with weighted loss
   • Result: Best performance (24 false negatives)

scale_pos_weight = len(y_train[y_train==0]) / len(y_train[y_train==1])
# ≈ 9.76 - tells XGBoost to treat minority class as 9.76x more important

Phase 3: Model Development

Systematically evaluated multiple algorithms:

1. Logistic Regression (Baseline)

• Purpose: Establish minimum performance benchmark
• Result: 96.76% accuracy, 91% recall, 29 FN
• Conclusion: Strong baseline, room for improvement

2. Random Forest

• Purpose: Capture non-linear relationships and feature interactions
• Result: 97.79% accuracy (highest), 90% recall, 30 FN
• Notable: Revealed Kurtosis_Integrated as most important feature (33%)

3. XGBoost (Final Model) ⭐

• Purpose: State-of-the-art gradient boosting with class imbalance handling
• Result: 96.70% accuracy, 92% recall (highest), 24 FN (lowest)
• Decision: Selected for minimizing false negatives

Hyperparameter Tuning Experiment

Used RandomizedSearchCV to optimize XGBoost:

• Unexpected Result: Tuning increased false negatives (27→31)
• Decision: Retained simpler model with better FN performance
• Learning: Default parameters often well-designed; blind tuning can hurt target metrics

---

📈 Results

Final Model Performance

Model: XGBoost with Class Weights
════════════════════════════════════════

Accuracy:  96.70%
Recall:    92.21% ⭐ (Highest)
Precision: 74.93%
F1-Score:  82.75%

Confusion Matrix:
                 Predicted
              Not Pulsar  |  Pulsar
Actual ──────────────────┼──────────
Not Pulsar    3178  ✅   |   94  ❌
Pulsar          24  ❌   |  284  ✅

False Negatives: 24 (7.79% of pulsars missed)
False Positives: 95 (acceptable for discovery science)

Model Comparison

Model	Accuracy	Recall	Precision	False Negatives
Logistic Regression	96.76%	91%	76%	29
Random Forest	97.79%	90%	85%	30
XGBoost (Final)	96.70%	92%	75%	24 ⭐

Business Impact

Without ML:

• Manual review of 17,898 candidates
• Time-consuming and error-prone
• Potential missed pulsars due to fatigue

With ML System:

• Flags 379 high-probability candidates (98% reduction)
• 284 true pulsars + 94 false alarms
• Only 24 pulsars missed (7.79%)
• Result: 98% workload reduction, 92% detection rate

---

💡 Key Learnings

Technical Insights

1. Class Imbalance: scale_pos_weight outperformed SMOTE for XGBoost
2. Feature Engineering: Model-based importance differed from visual EDA
3. Scaling: RobustScaler essential for astronomical data with outliers
4. Data Leakage: Always split before preprocessing
5. Evaluation: Accuracy misleading for imbalanced data; recall critical

Methodological Insights

1. Hyperparameter Tuning: Doesn't always improve target metrics
2. Multiple Approaches: SMOTE vs class weights - context determines best choice
3. Domain Knowledge: False negatives worse than false positives in discovery science
4. Model Selection: Best model ≠ highest accuracy; depends on business objective

---

🚀 Installation

Prerequisites

Python 3.8+

Setup

1. Clone the repository

git clone https://github.com/[YOUR_USERNAME]/pulsar-star-classification.git
cd pulsar-star-classification

2. Install dependencies

pip install -r requirements.txt

Or install manually:

pip install numpy pandas scikit-learn xgboost imbalanced-learn matplotlib seaborn

3. Download dataset

# Automatically downloads HTRU2 dataset
python download_data.py

---

💻 Usage

Training the Model

# Run complete pipeline
python train_model.py

# Or step-by-step in Jupyter
jupyter notebook notebooks/pulsar_classification.ipynb

Making Predictions

import pickle
import numpy as np

# Load model and scaler
with open('models/xgboost_final.pkl', 'rb') as f:
    model = pickle.load(f)
    
with open('models/scaler.pkl', 'rb') as f:
    scaler = pickle.load(f)

# Prepare new data
new_data = np.array([[140.5, 55.7, -0.23, -0.70, 3.2, 19.1, 8.0, 74.2]])
scaled_data = scaler.transform(new_data)

# Predict
prediction = model.predict(scaled_data)
probability = model.predict_proba(scaled_data)

print(f"Prediction: {'Pulsar' if prediction[0] == 1 else 'Not Pulsar'}")
print(f"Confidence: {probability[0][prediction[0]]:.2%}")

---

📁 Project Structure

pulsar-star-classification/
│
├── data/
│   ├── raw/                    # Original HTRU2 dataset
│   └── processed/              # Preprocessed data
│
├── notebooks/
│   ├── 01_EDA.ipynb           # Exploratory Data Analysis
│   ├── 02_preprocessing.ipynb # Data preprocessing
│   ├── 03_modeling.ipynb      # Model training & comparison
│   └── 04_evaluation.ipynb    # Final evaluation & analysis
│
├── src/
│   ├── data_preprocessing.py  # Preprocessing functions
│   ├── model_training.py      # Model training utilities
│   ├── evaluation.py          # Evaluation metrics
│   └── utils.py               # Helper functions
│
├── models/
│   ├── xgboost_final.pkl      # Final trained model
│   └── scaler.pkl             # Fitted RobustScaler
│
├── visualizations/
│   ├── eda_plots/             # EDA visualizations
│   ├── model_comparison/      # Model performance plots
│   └── confusion_matrices/    # Confusion matrices
│
├── requirements.txt           # Python dependencies
├── train_model.py            # Training script
├── download_data.py          # Dataset download script
├── README.md                 # This file
└── LICENSE                   # MIT License

---

🛠️ Technologies

Core Libraries:

• NumPy - Numerical computing
• Pandas - Data manipulation
• Scikit-learn - ML algorithms, preprocessing, evaluation
• XGBoost - Gradient boosting implementation
• imbalanced-learn - SMOTE and imbalance handling

Visualization:

• Matplotlib - Basic plotting
• Seaborn - Statistical visualizations

Development:

• Jupyter - Interactive development
• Pycharm - Local IDE
• Git/GitHub - Version control

---

🔮 Future Improvements

Short-term

• Ensemble voting (combine Logistic Regression + RF + XGBoost)
• SHAP values for model interpretability
• Streamlit web app for easy predictions

Long-term

• Deep learning approach (1D CNN on raw signal data)
• Anomaly detection framework (Isolation Forest)
• Real-time prediction API
• Integration with telescope data pipelines

---

📚 References

• R. J. Lyon et al. (2016). "Fifty Years of Pulsar Candidate Selection: From simple filters to a new principled real-time classification approach." Monthly Notices of the Royal Astronomical Society.
• Dataset: UCI Machine Learning Repository - HTRU2
• XGBoost Documentation: https://xgboost.readthedocs.io/

---

📧 Contact

[Ansul Suryawanshi]
📧 Email: [ansul2612@gmail.com]
💼 LinkedIn: Ansul Suryawanshi 🐙 GitHub: Ansul-S

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🌟 Acknowledgments

• Dataset: HTRU2 from UCI Machine Learning Repository
• Inspiration: Real-world astronomical discovery challenges
• Learning: Hands-on project-based learning approach

---

⭐ If you found this project helpful, please consider giving it a star!

---

Built with ❤️ and lots of efforts