Improving Robustness of Optimized Parameters Gradient Tree Boosting for Crime Forecast Model

The rise in crime rate poses significant problems to societies around the world. Departments responsible for public security spend significant time looking for ways of combating crime and adapting policing techniques to curb rising crime rates. The application of artificial intelligence for crime forecasting through optimized parameters of gradient tree boosting using dragonfly algorithm (DA-GTB) has recently attracted research attention. However, the DA-GTB is affected by a loss function that calculates pseudo-responses-value, in which by default uses the least absolute deviation (LAD) loss function. Motivated by the role and limitation of loss function, this research sought to determine DA-GTB crime forecast model using a loss function that improves the model’s predictive performance. Three crime type data from Tanzania Police force that cover 10 years was used and four-candidate loss functions (Wang et al., 2020, Annals of Data Science, 9, 187–212), were identified and applied experimentally to determine the best performing loss function. The evaluation showed that Huber loss function with optimal alpha (α): 0.15, 0.09, and 0.15 parameters with respect to theft, robbery, and burglary crime types, respectively, outperformed other loss functions by an average of 2%, achieving the lowest error rate. Thus, the Huber loss function with the attained α parameters improves the DA-GTB model crime forecast performance.