An Optimization of GCN-GRU Crime Prediction Network using Lyrebird Optimization Algorithm

Crime detection is an important aspect of helping law enforcement agencies to prevent new crimes by finding patterns. Nonetheless, the dynamic aspect of criminality and the swiftness of crime occasion creates immense challenges in precise classification of crimes in the future. To overcome this problem, different Deep Learning (DL) models have been examined in crime prediction activities. The Graph Convolutional Network with Gated Recurrent Unit (GCN-GRU) has been promising among them and can capture the spatial and temporal characteristics of crime. GCN models the local and global spatial dependencies by dynamically refining graph topologies, providing resistance to noisy or incomplete data and providing overfitting resistance. Concurrently, GRU which has lightweight architecture, learns effectively both short as well as long-term temporal dependencies in sequential crime information. The classification performance of the model is further enhanced with the Cross-Entropy Loss, which gives greater confidence to the correct crime types. Despite these advantages, the GCN-GRU model suffers from high computational complexity and strong dependence on data quality. Additionally, hyperparameters can be manually tuned and this is time consuming and can lead to suboptimal performance. In order to address these shortcomings, this paper presents an Optimization-based GCN-GRU (O-GCN-GRU) model that could be used to optimize the hyperparameters of the GCN-GRU network to maximize crime prediction results. This model uses a new metaheuristic optimization algorithm called Lyrebird Optimization Algorithm (LOA), inspired by the manner in which lyrebirds in the wild react to threats in their environment. The LOA has two phases, of exploration and exploitation whereby lyrebirds replicate an escape strategy to search a large number of possible solutions and a hiding strategy to search intensively within promising areas. The best hyperparameters of GCN-GRU model are determined based on these processes to make crime prediction with crime data effective. Through the application of LOA, this model is effective in fine-tuning the properties of the GCN-GRU network, which leads to both high predictive ability and less computation costs. Lastly, the experimental findings indicate that the O-GCN-GRU model has an accuracy rate of 96.8%, which is higher than the other models of crime predictions in terms of performance and reliability.