Crime Type Identification Using High-Order Deep Residual Network with Multiple Attention Algorithm

Crime type identification is crucial for improving public safety through more accurate prevention and efficient responses. However, practical applications often suffer from a significant lack of effective samples features, making it difficult to focus on the most informative aspects during identification. This study addresses these challenges by proposing a novel crime type identification method that leverages a deep neural network enhanced with multiple attention mechanisms. The approach includes a tailored data processing method involving target encoding to convert categorical data into numerical form, L2 normalizer to standardize data and ensure balanced feature contribution, and variance threshold feature selection to remove low-variance features. Additionally, a High-Order Deep Residual Network with Multiple Attention (HO-ResNet-MA) is developed, featuring an optimized Huta68 block (Huta-6(8)-MA ResBlock) with an enhanced Contextual Transformer (CoT) unit for local attention and a queue-and-exclusion layer for global attention. To validate the effectiveness of the proposed method, homicide reports data and Chicago crimes data are processed and fed into the crime type identification model, resulting in accuracies of over 84.1% and 99.5%, respectively. This study makes contributions to the field of crime analysis by validating the practical applicability of these approaches, and enhancing the efficiency of public safety workers.