Deep Neural Networks for Grid-Based Elusive Crime Prediction Using a Private Dataset Obtained from Japanese Municipalities

Abstract

People have the potential to be victims of elusive crimes such as stalking and indecent exposure anytime. To prevent the incidents, proposing an elusive crime prediction technique is a challenging work in Japan. This study assesses the efficiency of deep neural networks (DNNs) for grid-based elusive crime prediction using a private dataset obtained from Japanese municipalities that contains three crime categories (stalking, indecent exposure, and suspicious behavior) in five prefectures (Aichi, Fukuoka, Kanagawa, Osaka, and Tokyo) for 20 months (from July 2017 to February 2019). Through incremental training evaluation method that did not use future information of the testing 1-month data, the DNN-based technique using spatio-temporal and geographical information showed significant superior prediction performances (Mean ± SD%: 88.2 ± 3.0, 85.5 ± 4.5, and 85.8 ± 3.2 for stalking, indecent exposure, and suspicious behavior) to a random forest-based technique (81.9 ± 3.5, 83.3 ± 3.7, and 82.3 ± 2.1).