Urban crime prediction based on hierarchical classification and long short-term memory networks

The accelerated pace of urbanization has resulted in increasingly complex and unpredictable patterns of criminal activity, thereby posing ongoing challenges for public safety management. In response to these issues, a hierarchical time-series crime prediction framework has been formulated to enhance the capabilities of urban security systems. Crimes are initially classified into three tiers according to legal definitions and degrees of severity. Subsequently, specialized Long Short-Term Memory (LSTM) models are constructed for each crime tier to effectively capture the unique temporal dynamics associated with each category. Empirical validation, conducted using multi-year crime data from a major city in China, demonstrates the framework's robust fitting performance and high predictive accuracy across multiple crime types. The findings provide both theoretical insights and practical guidance for the development of stratified and differentiated early-warning systems aimed at facilitating intelligent crime prevention and governance in urban environments.