Enhancing short-term crime prediction with human mobility flows and deep learning architectures

Area of Science:

• Criminology
• Data Science
• Urban Studies

Background:

• Place-based short-term crime prediction models utilize historical crime data and spatio-temporal patterns.
• Crime opportunity theory posits that human mobility influences crime generation, making mobility data relevant for prediction.
• Previous studies demonstrated that mobility metrics from sources like call detail records (CDR) can improve crime prediction accuracy.

Purpose of the Study:

• To investigate the impact of fine-grained human mobility features on place-based short-term crime prediction.
• To evaluate the effectiveness of incorporating human mobility data into various deep learning models for crime forecasting.
• To assess prediction improvements across diverse urban environments and crime categories.

Main Methods:

• Utilized a publicly available, fine-grained human mobility dataset from a location intelligence company.
• Conducted comprehensive evaluations across multiple cities with varied demographics and crime types.
• Employed various deep learning models and compared them with non-neural regression models.

Main Results:

• Adding human mobility flow features to historical crime data improved F1 scores for neural crime prediction models by 2% to 7% across different cities and crime types.
• Neural network architectures demonstrated a slight performance improvement (up to 2%) compared to non-neural regression models.
• The study confirmed the value of human mobility data in enhancing the accuracy of short-term crime prediction.

Conclusions:

• Fine-grained human mobility data is a valuable feature for improving place-based short-term crime prediction models.
• The integration of mobility data offers tangible accuracy gains, particularly when using advanced neural network architectures.
• Future research should continue exploring mobility patterns and advanced modeling techniques for more effective crime forecasting.