Research map / Label-efficient and robust learning / Recognition and long-tail
GeoPro-Net: Learning Interpretable Spatiotemporal Prediction Models Through Statistically-Guided Geo-Prototyping
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GeoPro-Net, an intrinsically interpretable prototype-based deep model for spatiotemporal event forecasting (crime and traffic accidents) that uses statistical-test-based Geo-concept convolution, channel fusion, and geographic pooling.
Problem
Forecasting spatiotemporal events such as crimes and traffic accidents while providing intrinsic, human-understandable interpretations of multi-source spatiotemporal features under spatial heterogeneity.
Approach
Statistical significance tests (Poisson likelihood ratio, Kolmogorov-Smirnov) act as parameter-free convolutions that encode grid regions into binary Geo-concepts; these are fused across window sizes, pooled geographically (distance rings and directions), and matched to learned prototypes that are projected back to training samples for explanation.
Key results
- GeoPro-Net achieves the best value in most columns of the Chicago and NYC crime/accident tables, e.g., Chicago accident 0.53 vs 0.56 for the best prior baseline in the first metric.
- Replacing spatial geo-pooling with max pooling or removing pooling lowers performance (e.g., Chicago crime first metric 0.49 vs 0.51 and 0.67).
- GeoPro-Net is the only model in the comparison categorized as intrinsically interpretable with competitive accuracy against blackbox baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use two concept-convolution windows rather than one (ablation) | GeoPro-Net · spatiotemporal event forecasting datasets | prediction performance | Helped |
| Add spatial concept pooling instead of removing pooling (ablation) | GeoPro-Net · Chicago and New York City crime and accident | prediction performance | Helped |
| Replace max pooling with spatial concept pooling (ablation) | GeoPro-Net · Chicago and New York City crime and accident | prediction performance | Helped |
Surprising findings
- Removing the pooling layer causes a large performance drop, attributed to noisy information.
- On several metrics, Blackbox baselines such as HeteroConvLSTM and DCRNN outperform GeoPro-Net, so the accuracy advantage is not uniform.
Limitations
- Statistical test choices and window sizes must be specified by users rather than learned.
- Evaluated only on crime and traffic accident datasets from two US cities.
- Interpretation of ProtoPNet and ASTGCN is qualitative; no quantitative interpretability metric is reported.
Experimental setup
Datasets: Chicago crime Chicago traffic accident New York City crime New York City traffic accident
Benchmarks: Chicago crime Chicago accident New York City crime New York City accident
Metrics: error metric (reported x10^-3, unspecified in extract)
Baselines: ConvLSTM DCRNN Hetero-ConvLSTM DSTPP ViT GSNet AGL-STAN ASTGCN ProtoPNet
Builds on: ProtoPNet concept-based models Hetero-ConvLSTM
Closest papers by meaning
- Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach (NeurIPS 2025 · similarity 0.88)
- Drawing Informative Gradients from Sources: A One-stage Transfer Learning Framework for Cross-city Spatiotemporal Forecasting (AAAI 2025 · similarity 0.87)
- Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework (NeurIPS 2024 · similarity 0.87)
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.87)
- How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning (NeurIPS 2025 · similarity 0.86)
- Geolocation Representation from Large Language Models Are Generic Enhancers for Spatio-Temporal Learning (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.