🌌 Astrophysical Objects Classification

📜 Overview

This project applies machine learning (ML) techniques to classify astrophysical objects using observational data from the Large Synoptic Survey Telescope (LSST). By leveraging time-series and static numerical features, we aim to develop an accurate model for classifying celestial objects, such as comets and asteroids, based on their spatial positions, light curves, and motion.

🎯 Problem Explanation

The LSST dataset consists of astronomical time-series data collected through optical sky scanning. The telescope tracks the brightness and motion of celestial objects to support cosmic mapping and asteroid threat assessments.

Challenges in Classification

1. Complex Feature Representation:

   • The dataset contains both static (e.g., position, redshift) and time-series (e.g., brightness over time) attributes.
   • Integrating these features for classification is non-trivial.

2. Imbalanced Classes:

   • Certain astrophysical objects are overrepresented, leading to biased model predictions.

3. Selection of the Best ML Model:

   • Decision trees, boosting methods, and neural networks have different strengths for time-series and tabular data.
   • We evaluate multiple models to determine the best approach.

📌 Dataset: LSST Photometric Time Series (PLAsTiCC Competition)
🔢 Observations: 4,925 total (2,459 train | 2,466 test)
📊 Classes: 16 astrophysical object types

📊 Data Preprocessing & Feature Engineering

1️⃣ Dataset Overview

The dataset includes two types of data:

Feature Type	Description
Static Features	Spatial attributes (RA, DEC, Galactic coordinates)
Time-Series Features	Brightness variations over time

2️⃣ Handling Class Imbalance

• Applied class weighting to ensure fair representation.
• Downsampling techniques adjusted sample distribution.

3️⃣ Feature Engineering

• Rolling time-window transformations on flux values.
• Feature selection using correlation analysis.
• Normalization & scaling applied for neural network compatibility.

🔍 Exploratory Data Analysis (EDA)

1️⃣ Meta Data (Static Features)

• Right Ascension (RA) vs. Galactic Longitude (GL) shows an inverse parabolic trend.
• High variance in brightness motion data across objects.

📊 EDA Visualization:

• Feature correlation heatmaps
• Class distribution histograms
• Light curve plots of six sample objects

🤖 Machine Learning Models

1️⃣ Random Forest (RF) 🌲

✔️ Strengths:

• Handles high-dimensional data well.
• Provides high accuracy & interpretability.

📊 Results:

• Train Accuracy: 97% | Log Loss: 1.05
• Downsampled Accuracy: 91% | Log Loss: 3.17

2️⃣ Gradient Boosting (GB) 🚀

✔️ Strengths:

• Reduces bias via sequential boosting.
• More effective in handling complex relationships.

📊 Results:

• Train Accuracy: 68% | Log Loss: 11.64
• Downsampled Accuracy: 65% | Log Loss: 12.53

3️⃣ Adaptive Boosting (AB) 🏋️

✔️ Strengths:

• Assigns weights to observations, improving classification of hard-to-predict samples.

📊 Results:

• Train Accuracy: 46% | Log Loss: 19.54
• Downsampled Accuracy: 23% | Log Loss: 7.78

4️⃣ Neural Networks (NNs) 🧠

✔️ Strengths:

• Captures complex temporal dependencies in time-series data.
• Uses LSTM layers for time-dependent pattern recognition.

📊 Results:

• Train Accuracy: 46%
• Weighted Accuracy: 59%
• Log Loss: 9.6

🏆 Best Performing Model

📌 Random Forest (RF) is the most effective model:

• Highest accuracy (97%) on full train data.
• Performs well with both static & time-series features.
• Faster training time compared to Boosting & Neural Networks.

🚀 Key Observations:

• Feature selection improved model performance by removing redundant features.
• Downsampling maintained class distribution while improving computational efficiency.
• Neural Networks struggled with time complexity and performance tuning.

📢 Key Findings & Recommendations

✅ Key Insights:

• Light curve patterns are highly predictive of object classification.
• Feature selection significantly improves classification accuracy.
• Class imbalance correction is essential for fair model evaluation.

🔧 Recommended Future Work:

• 📊 Test alternative deep learning architectures (CNNs for feature extraction).
• ⚡ Optimize feature engineering for time-series processing.
• 🔬 Use SMOTE for class balancing instead of downsampling.

🚀 Technologies Used

🛠 ML Frameworks:

• Python (Scikit-learn, TensorFlow, Keras)
• Pandas, NumPy for preprocessing
• Matplotlib, Seaborn for visualization

📡 Dataset: LSST Photometric Time Series Data (PLAsTiCC)