Analyze Crime using a Deep LSTM-based Autoencoder

This study presents a deep learning framework for analyzing crime data using an LSTM-based autoencoder to detect anomalies in criminal patterns and investigative outcomes. The methodology begins with normalizing multidimensional data, including financial activities, behavioral patterns, and criminal records, followed by temporal segmentation into fixed-length sliding windows to preserve sequential dependencies. The autoencoder architecture employs LSTM layers in both encoder and decoder components, trained with Mean squared error loss and optimized using Adam with early stopping regularization to prevent overfitting. Analysis of South Yorkshire Police data from December 2024 revealed critical patterns in case outcomes. The model demonstrated effective anomaly detection through reconstruction error thresholds, with geospatial visualization showing anomalous crime records clustered outside typical crime hotspots. This framework provides actionable insights for improving investigative workflows and data quality controls in law enforcement agencies. By automating the identification of irregular patterns, it highlights procedural inconsistencies, incomplete documentation, or emerging crime trends requiring policy interventions. The study underscores the potential of deep learning techniques to enhance the accuracy and efficiency of crime data analysis and investigative processes.