Spatio-Temporal Crime Prediction Using Dynamic Mode Decomposition and CNN-LSTM

Crime is an issue that has affected all levels of society. A reactive approach to combating crime through Law Enforcement could potentially be more effective, as opposed to a proactive approach. This study delved into the use of crime prediction models to reinforce existing strategies, through employing Dynamic Mode Decomposition (DMD) and a Convolutional Neural Network Long Short-Term Memory (CNN-LSTM) architecture. Both models were leveraged to capture the spatial and temporal dimensions of crime. By adopting an epidemiological framework, the researchers approached crime as a dynamic, systemic phenomenon akin to the spread of diseases, offering new insights into its patterns. Originally designed for fluid dynamics, DMD was adapted to uncover spatial and temporal trends in crime data, providing critical insights into the evolution and concentration of criminal activities. On the other hand, CNN-LSTM used the combination of the spatial feature extraction capabilities of CNNs with the temporal forecasting strength of LSTMs. Using a publicly available crime dataset with geographic and temporal markers, the researchers evaluated the models’ performance through standard regression metrics such as Mean Absolute Error, Mean Squared Error, and Root Mean Squared Error. Notably, the findings revealed that DMD outperformed CNN-LSTM in its ability to predict crime patterns across both space and time as it had a lesser margin of error per cluster and day across all predicted days.