Predicting density of serious crime incidents using a Multiple-Input Hidden Markov Maximization a posteriori model

Abstract

Accurately predicting the displacement of crime from a given state such as cold to another state such as warm or hot, facilitates the efficient allocation of resources and the mitigation of crime threats. In this study, a crime forecasting model was developed, based on Spearman’s Correlations and a clustering technique (DBSCAN), which captures significant groupings in a geospatial dataset. A Multi-Input Hidden Markov Model (MI-HMM) machine learning framework was developed to train the dataset. The results from the MI-HMM were then used to make a Maximum a Posteriori (MAP) decision over the possible state of crime for the next month. This novel model, MI-HMM-MAP, was used to predict the density of crime including criminal hot spots over time. The model was evaluated using real-world dataset. Findings show an average of 72.5% accuracy and 81.7% correctness. The model was compared to 5 classical predictive models. Results show that our model significantly outperforms a linear regression model, a neural network model, and two machine learning approaches. It slightly outperforms a deep learning approach as demonstrated statistically by an application to the crime of murder in Trinidad and Tobago.