Modeling Urban Walking Risk Using Spatial-Temporal Machine Learning

A machine learning model is developed to predict the perceived risk of urban walking routes using San Francisco police incident data. The approach uses Uber's H3 hexagonal grid system for geospatial representation and cyclical encoding for temporal features like time of day and day of the week. An LLM is utilized to assign severity scores to different incident types, creating a nuanced risk signal that goes beyond simple incident counts. The final model employs XGBoost with a Tweedie regression objective, which is well-suited for modeling the zero-inflated and right-skewed distribution of incident risk data.