Pulsar Classification for Class Prediction

Project Objective

The goal of this project is to build a robust classification model capable of predicting whether a space observation instance is a Pulsar (1) or not a Pulsar (0).

Pulsars are rare astronomical objects, and identifying them requires high-precision modeling to handle the significant class imbalance in the dataset.

Challenge:

High imbalance due to rare nature of Pulsar stars

Dataset Overview

The dataset contains 17,898 entries with features derived from the Integrated Profile and the DM-SNR Curve of the observations.

Features:

• Mean_Integrated: Mean of the integrated profile.
• SD: Standard deviation of the integrated profile.
• EK: Excess kurtosis of the integrated profile.
• Skewness: Asymmetry of the probability distribution.
• Mean_DMSNR_Curve: Mean of the DM-SNR curve.
• SD_DMSNR_Curve: Standard deviation of the DM-SNR curve.
• EK_DMSNR_Curve: Excess kurtosis of the DM-SNR curve.
• Skewness_DMSNR_Curve: Skewness of the DM-SNR curve.
• Class: Target variable (0: Not Pulsar, 1: Pulsar).

Note on DM-SNR Curve: Radio waves emitted from pulsars reach Earth through space filled with free electrons. Higher frequencies slow down less than lower frequencies (dispersion), creating a curve used for identification.

Data Analysis & Preprocessing

1. Data Understanding

• Imbalance Handling: Class 0 has 16,259 entries while Class 1 only has 1,640.
• Correlation Insights: Features like EK and Skewness show high positive correlation with the Target Class, while Mean_Integrated shows high negative correlation.

2. Data Cleaning

• Missing Values: Handled using Median Imputation for features like SD, Skewness, and SD_DMSNR_Curve.
• Duplicates: Identified and removed to ensure data integrity.

3. Feature Engineering & Scaling

• Feature Selection: Dropped highly redundant features (Mean_Integrated, Skewness, Mean_DMSNR_Curve, EK_DMSNR_Curve) to reduce multicollinearity.
• Standardization: Applied StandardScaler to normalize feature distributions.

Model Building

I utilized a Random Forest Classifier for this task.

Why Random Forest?

• Ensemble Power: It operates by constructing multiple decision trees and outputting the average prediction, which helps mitigate overfitting.
• Imbalance Resilience: It is highly effective at handling imbalanced datasets where one class (non-pulsars) significantly outweighs the other.

Model Parameters:

• n_estimators: 500
• random_state: 42
• test_size: 20%

Evaluation Results

The model achieved high performance across all key metrics, proving that it can identify rare Pulsars without sacrificing overall accuracy.

• MetricScoreAccuracy --> 98.35%
• Precision (Class 1) --> 92%
• Recall (Class 1) --> 90%
• F1-Score (Class 1) --> 91%

Confusion Matrix:

• True Negatives (Non-Pulsar): 3223
• True Positives (Pulsar): 298
• False Positives: 25
• False Negatives: 34

---

Tech Stack

• Language: Python
• Libraries: Pandas, NumPy, Scikit-learn, Seaborn, Matplotlib