TensoCast

Traffic forecasting using tensor decomposition and machine learning on METR-LA dataset.

1) Install

Requires Python 3.9+.

pip install -r requirements.txt

2) Data

Place the dataset at:

• data/processed/METR-LA.h5

3) Run

Run as module:

Baselines only:

python -m src.main --models ha,naive --quick

Quick LSTM (1 epoch):

python -m src.main --models lstm --quick

Quick XGBoost:

python -m src.main --models xgb --quick

Full run (all models, default settings):

python -m src.main --models ha,naive,lstm,xgb

Use CP decomposition too (optional, slower):

python -m src.main --models ha,naive,lstm --use-decomposition --rank 50

CLI options

• --data-path (str): Path to H5 file. Default: data/processed/METR-LA.h5
• --window-size (int): History window length. Default: 12
• --horizon (int): Forecast steps. Default: 3
• --models (str): Comma-separated subset of ha,naive,lstm,xgb
• --quick (flag): Fast run (limits samples, LSTM 1 epoch, fewer XGB trees)
• --limit-samples (int): Manually cap samples used before split
• --use-decomposition (flag): Apply CP decomposition to normalized data
• --rank (int): CP rank if --use-decomposition is set

Outputs

• results/metrics.csv: Summary metrics per model (mae, rmse, mape, smape)
• results/figures/metrics_comparison.png: Bar charts across models
• results/figures/lstm_predictions.png: Example plot (if LSTM selected)

Project layout

• src/main.py: Orchestrates the pipeline and CLI
• src/data_utils.py: Load H5, normalize, sliding windows
• src/decomposition.py: CP/Tucker helpers (optional)
• src/baselines.py: Historical Average, Naive persistence
• src/models/lstm_model.py: PyTorch LSTM forecaster
• src/models/xgb_model.py: XGBoost forecaster (per-horizon, multi-output)
• src/metrics.py: mae, rmse, mape, smape
• src/visualization.py: Plots and results table

Notes

• Run from repo root as module: python -m src.main
• Plots saved to results/figures/