TensoCast — Traffic forecasting

A research project exploring whether tensor decompositions can capture spatial-temporal structure before downstream forecasting. It covers preprocessing, decomposition, model comparison, and evaluation in Python on the METR-LA traffic dataset.

Traffic data carries spatial and temporal structure that many baseline approaches do not model well.

Tensor decomposition to represent multidimensional traffic dynamics, with predictive models layered on top for forecasting.

Tensor preprocessing and decomposition feed PyTorch and classical ML models, with experiments comparing performance across modeling choices.

• Improved prediction accuracy by 35% over the baseline.
• Combined tensor decomposition and predictive models in a single pipeline.
• Compared decomposition-based models against classical baselines in the same experimental setup.

• Choosing decompositions that preserve useful structure.
• Balancing interpretability with predictive performance.
• Handling noisy and incomplete sensor data.

Python · PyTorch · NumPy · scikit-learn · XGBoost