🚀 NSSC 2025: Data Analytics Event

🧩 Team Information

• Team ID: T-0823769
• Team Name: ASHSUM
• Team Members:
  1. Sumit Pandey : 25-222056
  2. Ashirwad Sinha : 25-412666

---

🧠 Project Overview

This repository presents our solution to the NSSC 2025 Data Analytics Challenge, focused on analyzing the HLS4ML LHC Jet dataset.

Dataset: Drive Link

Our project is built around two major objectives:

1. Jet Classification
   We classify particle jets into one of five categories using:

   • Deep Learning Approach: A fine-tuned ResNet-18 model trained on jet image data.
   • Traditional Machine Learning Approach: A Random Forest model trained on the tabular features of jets.

2. Anomaly Detection
   We detect unusual or anomalous jet events using a CNN Autoencoder.

   • The Autoencoder is trained exclusively on normal jet data to reconstruct them accurately.
   • Significant reconstruction error signals potential anomalies.

---

🧱 Project Structure

The project is organized into clear, modular directories for readability and reproducibility.

NSSC_DataAnalytics_Team_ASHSUM143/
├── 📂 Notebook/
│   ├── 1_Data_Preprocessing.ipynb
│   ├── 2_CNN_Model.ipynb
│   ├── 3_Tabular_Models.ipynb
│   ├── 4_Model_Comparison.ipynb
│   └── 5_Anomaly_Detection.ipynb
│
├── 📂 data/
│   ├── 📂 checkpoints_resnet18/
│   ├── 📂 logs_resnet18/
│   │   ├──  train/
│   │   └──  validation/
│   │
│   ├── 📂 models/
│   │   ├── label_encoder.joblib
│   │   ├── pca.joblib
│   │   ├── resnet18_jet_classifier_best_model.keras
│   │   ├── rf_baseline.joblib
│   │   ├── rf_pca.joblib
│   │   └── scaler.joblib
│   │
│   ├── 📂 results/
│   │   ├── 📂 plots/
│   │   │   ├── anomaly_error_distribution.png
│   │   │   ├── cnn_accuracy_loss_curves.png
│   │   │   ├── cnn_confusion_matrix.png
│   │   │   ├── cnn_misclassified_jets.png
│   │   │   ├── confusion_matrix_pca.png
│   │   │   ├── confusion_matrix_raw.png
│   │   │   ├── model_comparison.png
│   │   │   ├── pca_scree_plot.png
│   │   │   └── top_5_anomalies.png
│   │   │
│   │   ├── anomaly_scores.csv
│   │   ├── cnn_summary.csv
│   │   ├── sample_images.npy
│   │   ├── tabular_feature_summary.csv
│   │   └── tabular_summary.csv
│   │
│   ├── 📂 train_data/       (Contains 36 .h5 files)
│   └── 📂 validation_data/  (Contains 27 .h5 files)
|
|── 📁 Report/
|   ├── NSSC_Report_T-0823769.docx
|
├── README.md
└── requirements.txt

---

⚙️ Key Components

🔹 1. Data Preprocessing

• Handled missing values, normalization, and reshaping for image-based models.
• Ensured class balance using stratified sampling techniques.

🔹 2. Exploratory Data Analysis (EDA)

• Visualized jet distributions, feature importance, and correlations.
• Compared tabular and image-based feature representations.

🔹 3. Classification Models

🧠 ResNet-18 (Image-based)

• Fine-tuned on jet images.
• Implemented using PyTorch.
• Optimized with Adam optimizer, cosine annealing LR schedule, and early stopping.

🔹 Random Forest (Tabular)

• Used for interpretable classification on structured features.
• Tuned with GridSearchCV for best performance.

🔹 4. Anomaly Detection

• Built a CNN Autoencoder to reconstruct normal jet events.
• Reconstruction error used as the anomaly score.
• Evaluated using ROC AUC and precision-recall metrics.

🔹 5. Model Comparison

• Compared metrics such as Accuracy, F1-Score, ROC-AUC, and Inference Time.
• Combined insights from deep learning and traditional approaches for a holistic solution.

---

📊 Results Summary

Model	Acc	F1	ROC AUC	Notes
RF	0.8225	0.8198	0.9012	Best overall
PCA RF	0.7939	0.7901	0.8834	Fast and efficient
CNN	0.7437	0.7392	0.8645	Learns jet structure

---

Technologies Used

• Python 3.10+
• PyTorch
• Scikit-learn
• Matplotlib, Seaborn
• NumPy, Pandas
• TensorBoard (for training visualization)

---

🧪 Reproducibility

To reproduce our results:

# Clone the repository
git clone https://github.com/5umitpandey/IIT_Kh_NSSC_Data_Analytics_Hackathon.git
cd IIT_Kh_NSSC_Data_Analytics_Hackathon

# Install dependencies
pip install -r requirements.txt

# Run notebooks
jupyter notebook Notebook/

PPT: PPT Live Link

Photos: