Uma análise dos impactos de hiperparâmetros aplicados ao algoritmo kde para a geração de mapas de hotspots criminais

Abstract

The use of machine learning is growing every day and has applications in several areas of everyday life. An example of the use of machine learning is in the area of public safety, where there is a need to determine risky locations so that patrolling reinforcements can be directed. Several methods have already been proposed to solve this problem. For the most part, these techniques need to be manually configured parameters so that they can obtain better results. In this need, there is the problem of hyperparameter optimization. This work seeks to analyze the impacts of different hyperparameters on the model result, for this, the grid search and random search hyperparameter optimization techniques are used, having the problem as a use case. for generating crime hotspot maps using the KDE technique. Furthermore, this work proposes an extension of KDE, combining kernel functions in space and time as optimization hyperparameters to improve results. The experiments demonstrate that with the use of random search, results were obtained with an evaluation up to 7.63% better than what is obtained using the grid search, mainly due to the use of a lower bandwidth (318m). It is also observed that adding a kernel time function to KDE that decreases the relevance of older events results in a 4.31% improvement in grid search results.