Research map / Video / Video segmentation and tracking
Hawkes-Enhanced Spatial-Temporal Hypergraph Contrastive Learning Based on Criminal Correlations
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 HCL, a Hawkes-enhanced spatial-temporal hypergraph framework that adds negative-free MSE contrastive learning over type and neighbor spatial crime correlations plus Hawkes-process temporal enhancement, evaluated for crime quantity and occurrence prediction on NYC and Chicago.
Problem
Crime prediction on grid-partitioned urban regions fails to exploit co-occurrence of crime types, neighbor-block danger spillover, and short-term temporal excitation between events.
Approach
Adds two correlation-based negative-free contrastive losses (type and neighbor spatial correlations, aligned with MSE on cosine-normalized embeddings) and a Hawkes-process weighted temporal enhancement on top of existing hypergraph backbones (ST-HSL, ST-SHN) as a plug-and-play module.
Key results
- HCL reduces MAE and MAPE on NYC crime quantity prediction by about 2.96% and 6.20% on average over the best prior model.
- On crime occurrence prediction in NYC, HCL improves Micro-F1 to 0.6301 and Macro-F1 to 0.6304 over ST-SHN (0.6111 / 0.6126).
- As a plug-in on ST-HSL and ST-SHN backbones, HCL yields average 2.84% MAE and 5.47% MAPE gains and 2.08% F1 gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Hawkes temporal enhancement, type spatial contrastive learning, and neighbor spatial contrastive learning | ST-HSL backbone in HCL · NYC and CHI crime quantity prediction | MAE: 2.96% on NYC; 2.72% on CHI | Helped |
| Add Hawkes temporal enhancement, type spatial contrastive learning, and neighbor spatial contrastive learning | ST-SHN backbone in HCL · NYC and CHI crime occurrence prediction | F1: 3.1% and 2.9% | Helped |
| Add Hawkes temporal enhancement (ablation) | HCL · NYC and CHI crime prediction | prediction performance | Helped |
| Add type spatial contrastive learning (ablation) | HCL · NYC and CHI crime prediction | prediction performance | Helped |
| Add neighbor spatial contrastive learning (ablation) | HCL · NYC and CHI crime prediction | prediction performance | Helped |
Limitations
- Adds several hyperparameters (lambda_tc, lambda_nc, delta, s) requiring grid search.
- Extra time consumption from the added module, acknowledged by the authors.
- Gains are modest on Chicago compared with NYC.
Experimental setup
Datasets: NYC crime data (2014-2015) Chicago crime data (2016-2017)
Benchmarks: NYC crime data Chicago crime data
Metrics: MAE MAPE Micro-F1 Macro-F1
Baselines: ST-HSL ST-SHN GMAN DCRNN STGCN ST-MetaNet DeepCrime ST-ResNet
Builds on: ST-HSL ST-SHN BYOL-style negative-free contrastive learning
Closest papers by meaning
- Spatial-Temporal Augmentation for Crime Prediction (Student Abstract) (AAAI 2024 · similarity 0.87)
- Dynamic Neighborhood Modeling via Node-Subgraph Contrastive Learning for Graph-Based Fraud Detection (AAAI 2025 · similarity 0.83)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.81)
- Alleviating Performance Disparity in Adversarial Spatiotemporal Graph Learning Under Zero-Inflated Distribution (AAAI 2025 · similarity 0.80)
- Graph-Aware Contrasting for Multivariate Time-Series Classification (AAAI 2024 · similarity 0.80)
- Graph Structure Learning for Spatial-Temporal Imputation: Adapting to Node and Feature Scales (AAAI 2025 · similarity 0.79)
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.