New Multi-Task Model Boosts Cross-City Accident Prediction Accuracy

Global: New Multi-Task Model Boosts Cross-City Accident Prediction Accuracy

Study Overview

Researchers from multiple institutions have introduced the Mamba Local‑ttention Spatial‑Temporal Network (MLA‑STNet), a unified framework that treats accident risk forecasting as a multi‑task learning problem spanning several cities. The system was evaluated using real‑world traffic accident records from New York City and Chicago, and it demonstrated notable gains in predictive performance compared with existing baselines.

Motivation and Data Challenges

Urban accident datasets are often fragmented, featuring heterogeneous formats, inconsistent reporting standards, and sparse, noisy observations that cluster temporally and spatially. These characteristics have historically impeded the development of scalable, cross‑city safety solutions, prompting the need for a model that can harmonize disparate data sources while preserving city‑specific nuances.

Model Architecture

The proposed MLA‑STNet combines two complementary attention modules. The Spatio‑Temporal Geographical Mamba‑Attention (STG‑MA) attenuates unstable fluctuations in space‑time patterns and reinforces long‑range temporal dependencies. Simultaneously, the Spatio‑Temporal Semantic Mamba‑Attention (STS‑MA) employs a shared‑parameter design to reduce cross‑city heterogeneity, enabling joint training across cities while maintaining distinct semantic representation spaces for each locale.

Experimental Design

To assess the model, the authors conducted 75 experiments covering two forecasting scenarios: full‑day accident prediction and high‑frequency accident periods. Both scenarios leveraged the same underlying datasets from New York City and Chicago, allowing a direct comparison of MLA‑STNet against state‑of‑the‑art baselines under identical conditions.

Performance Outcomes

Across the test suite, MLA‑STNet achieved up to 6% lower root‑mean‑square error (RMSE), 8% higher recall, and 5% higher mean average precision (MAP) relative to competing methods. These improvements were observed consistently across both cities and forecasting horizons.

Robustness to Data Noise

When subjected to 50% input noise, the model’s performance varied by less than 1%, indicating strong resilience to the noisy and incomplete data typical of urban accident reporting systems.

Implications for Urban Safety Management

The results suggest that MLA‑STNet can serve as a scalable, robust, and interpretable component of a coordinated, data‑driven accident prevention strategy. By unifying heterogeneous urban datasets, the approach paves the way for more effective cross‑city collaboration in traffic safety initiatives.This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.