Graph-based Deep Learning for Spatiotemporal Forecasting

Geometric and topological deep learning for node-level forecasting in spatiotemporal graphs.

This long-running project applies graph-based deep learning to forecast spatiotemporal signals, with emphasis on crime data and sparse hotspot dynamics.

Main Contributions

Applied geometric deep learning and state-of-the-art forecasting models on graph nodes.
Reached over 90% overall accuracy in broad scenarios.
Identified persistent performance gaps for sparse local hotspots (below 40% in difficult settings).

Ongoing Directions

Hypergraph-based formulations for richer relational structure.
Multi-resolution and U-Net-like architectures for local detail recovery.
Error analysis focusing on high-variance and low-density regions.

Status

In progress (July 2023 - Present).