🌐 Temporal Graph Forecasting

Forecast temporal trends on graphs with TGN, TGCN, DCRNN and LSTM baselines

---

Forecast temporal trends on graphs using PyTorch Geometric Temporal (TGN, TGCN, DCRNN) and sequence baselines (LSTM). This project showcases end-to-end pipelines for building temporal graph datasets, training models, and evaluating RMSE on edge trends.

🌟 Highlights

• Temporal GNNs: TGN memory-based predictor, TGCN, DCRNN
• Baselines: LSTM for sequence modeling

📁 Repository structure

notebooks/
  tutorials/
    01-PyGTemporal-Chickenpox-DCRNN-TGCN-TGN.ipynb
  baselines/
    01-LSTM-EdgeTrend-Baseline.ipynb
    02-RMSE-LSTM-Baseline-Eval.ipynb
  tgn/
    01-TGN-Trend-Predictor-EndToEnd.ipynb
    02-RMSE-TGN-Edge-Score-Predictor.ipynb
    04-TGN-Trend-Long-Forecast.ipynb
    05-TGN-Trend-Predictor-Variable-Window.ipynb
    06-Generate-TGN-Embeddings-Corvic-Feature-Views.ipynb
    07-Generate-TGN-Embeddings-Corvic-Feature-Views-Refactored.ipynb
    08-Write-TGN-Embeddings-to-GCloud.ipynb
  experiments/
    02-Decoupled-MLP-LSTM-with-TGN-Embeddings.ipynb
    03-Decoupled-MLP-LSTM-LongForecast-TGN-Embeddings.ipynb
    04-Decoupled-LSTM-GCloud-TGN-Embeddings.ipynb
    05-Decoupled-LSTM-LongForecast-GCloud-TGN-Embeddings.ipynb
    06-MLP-LSTM-Variable-Length.ipynb
    07-RMSE-Decoupled-MLP-LSTM-TGN-Embeddings.ipynb

🚀 Quickstart

1) Environment

make setup

2) Install PyG wheels (if needed)

PyTorch Geometric extensions depend on your Torch and CUDA/CPU version. If pip install from requirements.txt fails, install PyG wheels using the official index, then re-run make setup:

import torch
v = torch.__version__.split('+')[0]
cu = 'cpu' if torch.version.cuda is None else 'cu' + torch.version.cuda.replace('.', '')
print(f"Use: https://data.pyg.org/whl/torch-{v}+{cu}.html")

Then install the matching wheels (example):

pip install torch-scatter -f https://data.pyg.org/whl/torch-2.6.0+cpu.html
pip install torch-sparse -f https://data.pyg.org/whl/torch-2.6.0+cpu.html
pip install torch-cluster -f https://data.pyg.org/whl/torch-2.6.0+cpu.html
pip install torch-spline-conv -f https://data.pyg.org/whl/torch-2.6.0+cpu.html
pip install torch-geometric torch-geometric-temporal

3) Run notebooks

Open any notebook in notebooks/. Each has a short description and can be executed end-to-end.

🗂️ Data

Some notebooks reference GCS paths (gs://...) for scientific trend datasets. Replace with your accessible path, or mount GCS using gcsfuse/gsutil, or copy locally to data/.

📓 Notebooks guide

• Tutorials: Chickenpox dataset walk-through of DCRNN/TGCN/TGN
• TGN: End-to-end trend predictor, RMSE evaluators, long-horizon/variable-window studies, and embedding generation
• Baselines: LSTM edge-trend baselines + RMSE evaluation
• Experiments: Decoupled MLP+LSTM variants leveraging TGN embeddings

🧑‍💻 Development

• Install hooks: make hooks
• Lint all files: make lint
• Auto-format: make format