GrATSI: GRaph-based Attributions for Time Series explainability

GrATSI introduces a framework for graph-based attributions in time series models, bridging saliency methods to graphs recovery. By using the post-hoc explainers to generate a structured graph, it enables understanding and evaluation of feature interaction and relevance in temporal modelling.

---

📦 Installation

Clone and set up the environment:

git clone https://github.com/<your-username>/grats-xai.git
cd grats-xai

# Create conda environment
conda create -n grats python=3.9 -y
conda activate grats

# Install requirements
pip install -r requirements.txt

Project Structure

├── configs/               # YAML configs for data generation & experiments
│   └── data_gen.yaml
├── src/
│   ├── datasets/          # Synthetic DBN generator
│   │   └── synthetic_dbn.py
│   ├── models/            # Simple baselines (e.g. LSTM)
│   │   └── simple_lstm.py
│   ├── explainers/        # Explainability methods (IG, TimeRISE, etc.)
│   └── evaluation/        # Metrics (infidelity, comprehensiveness, etc.)
├── runs/                  # Auto-saved experiments (ignored via .gitignore)
└── README.md

🚀 Usage

All supported in the config

Quick check (debugging)

python scripts/pipeline.py --config configs/pipeline_quick.yaml

Complete Run

python scripts/pipeline.py --config configs/pipeline.yaml

Compile and aggregate the runs to visualize

python scripts/agg_results.py

1. Generate synthetic data

Outputs are saved under:

runs/dbn_n{params}/
  ├── train.pkl
  ├── val.pkl
  └── plots/

2. Train a model

3. Run explanability

Choose an explainer:

# Integrated Gradients
# TimeRISE

// TODO: Integrated Hessians DeepLIFT Timex++ etc.

4. Asymmetric Perturbation Response to estimate W_hat (p copies for the lag)

5. Attribution Refinement using graph

📊 Example Output