TRAGIC (Time seRies Analysis using Graph attentIon networks Classifier)

A deep learning framework that uses Graph Attention Networks (GAT) for time series classification. This project implements a novel approach to transform time series data into graph structures, enabling geometric deep learning techniques for analysis and classification.

Core Architecture

Graph Representation

• Time series are converted into graphs where:
  • Each timestep becomes a node
  • Edges connect neighboring timesteps within a window
  • Node features include the time series value and normalized position
  • Self-loops are added to allow nodes to maintain their own information

Main Model Components

The GNNTimeSeriesClassifier architecture includes:

• Input projection layer
• Two GAT layers with attention mechanisms
• Layer normalization for stability
• Global mean pooling for graph-level representations
• Final classification layers

Key Features

Robust Training

• K-fold cross validation for reliable evaluation
• Early stopping to prevent overfitting
• Class weighting for imbalanced datasets
• Per-fold data normalization to prevent leakage

Visualization Capabilities

• Attention weights visualization
• Saliency maps
• ROC curves with confidence intervals
• Confusion matrices
• Time series examples

Interpretability Tools

• Attention visualization for important timesteps
• Saliency analysis for feature influence
• Attention weight distributions
• Average attention profiles across samples

Comprehensive Evaluation

• Metrics:
  • Accuracy
  • Balanced accuracy
  • Matthews Correlation Coefficient (MCC)
  • AUC-ROC (for binary classification)
  • Per-class metrics
  • Bootstrap confidence intervals

Data Processing

• Support for UCR/UEA time series datasets
• Graph structure conversion
• Proper train/validation/test splitting
• Leakage-free data normalization

Output & Logging

Saves detailed results including:

• Metrics with confidence intervals
• Visualizations
• Per-fold performance
• Classification reports
• Summary statistics (CSV/JSON)

Project Structure

• *.ipynb: Jupyter notebooks containing experiments
• plots/: Generated visualizations
• results/: Experimental results and metrics

Requirements

See requirements.txt for detailed package dependencies.

Usage

1. Clone the repository

git clone https://github.com/salilp42/TRAGIC.git
cd TRAGIC

2. Install dependencies

pip install -r requirements.txt

3. Run experiments through Jupyter notebooks

License

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