Advancing urban traffic accident forecasting through sparse spatio-temporal dynamic learning.

Area of Science:

• Computer Science
• Urban Planning
• Public Health

Background:

• Traffic accidents pose significant public health and safety challenges.
• Accurate prediction is vital for informed urban planning and public confidence.
• Traditional models struggle with sparse and unevenly distributed accident data.

Purpose of the Study:

• To develop an advanced framework for accurate traffic accident prediction.
• To address limitations of existing models in handling sparse spatio-temporal data.
• To enhance understanding of complex dependencies in accident data.

Main Methods:

• Proposed a Sparse Spatio-Temporal Dynamic Hypergraph Learning (SST-DHL) framework.
• Integrated multi-view spatiotemporal convolution for local correlations.
• Employed cross-regional dynamic hypergraph learning for global dependencies.
• Utilized a two-supervised self-learning paradigm for robust pattern capture.

Main Results:

• SST-DHL demonstrated significant improvements over baseline models on New York City and London datasets.
• The framework showed enhanced performance across various data sparsity levels.
• Achieved superior accuracy in predicting traffic accidents compared to existing methods.

Conclusions:

• The SST-DHL framework effectively captures higher-order dependencies in sparse traffic accident data.
• It offers improved accuracy and interpretability for traffic accident prediction.
• This approach enhances public safety and trust through reliable predictions.