Efficient Generative Adversarial Networks with Recurrent Convolutional Network for Prediction of Crimes

Abstract

Accurate crime prediction is paramount for the enhamcement of public safety;however ,the exisiting models frequently exhibit deficiencies in their capacity to accurately anticipate crime occurances,primarily due to their inadequacy in accomodating intricate spacial and temporal data,which encompasses interdependencies and multi-scale patterns.To mitigate these shortcomings, an innovative Crime Recurrent Graph Generative Adversial Network (CRGGAN) has been introduced, which incorporates a Graph Convolution Network (GCN) functioning as its discriminator to proficiently model spatial relationships and spillover effects,alongside a bi-directional Long Short-Term Memory (bi-LSTM) network within the generator to effectively capture complex temporal dynamics across diverse scales.This integrative methodology sustantially augments the model's capacity to amalgamate spatial and temporal patterns,thereby facilitating more precise predictions and enhanced crime prevention strategies.The results from the experiments show that CRGGAN achieves an impressive perfromance level,with an accuracy rate of 98.6% ,precision at 97.6%,recall hitting 98.1% and a F1-score of 97.2%.Forthermore the model displays a minimal running time and improved precision, recall,F1-score and specificity in its outcomes,with a training duration of 42 seconds for 50 epochs and an inference time of 1.5 milliseconds per sample,rendering it exceptionally suitable for application in real-world scenarios.