TensoCast – Traffic Forecasting

Problem

Traffic prediction is crucial for smart city planning, but classical methods struggle with the complex spatial-temporal patterns in traffic data.

Solution

Proposed a novel method combining CANDECOMP/PARAFAC (CP) and Tucker tensor decomposition with LSTM networks, achieving 35% improvement over classical baselines.

Architecture

Uses PyTorch for deep learning, NumPy for tensor operations, and XGBoost for final prediction refinement, evaluated on the METR-LA dataset.

Tech Stack

Key Challenges

• Designing efficient tensor decomposition algorithms
• Balancing model complexity with interpretability
• Handling missing data in traffic sensors