A Model for Criminal Behaviour Prediction Using Machine Learning in the UAE Security System

Abstract

Crime poses a significant societal challenge, impacting both individual well-being and economic stability. Despite a general decline in global crime rates, certain crime types are on the rise, necessitating innovative strategies and advanced technologies for prediction and prevention by law enforcement agencies. Machine learning, with its growing application in predictive modeling, plays a critical role in this context. By processing data through advanced algorithms, these models can forecast crime, detect hotspots, identify criminals, and aid in other crucial tasks. This study leverages unstructured crime data from the United Arab Emirates (UAE) Police Department. A thorough pre-processing regimen is applied to structure the data, converting various text elements into categorical data. A significant aspect of this process involves using the textual crime descriptions to generate quantitative features through natural language processing (NLP). The dataset is annotated with crime types through a process combining NLP and label induction, primarily using unsupervised learning and clustering. This approach ensures a representative and reliable label set. Predictive modeling employs a range of classification algorithms, selected based on their proven effectiveness in relevant literature. Two distinct experiments were conducted: the first focused on socio-demographic features of criminals to predict criminal behavior, while the second augmented this feature set with data derived from crime descriptions using NLP techniques. This was done to evaluate whether the addition of crime-specific information could enhance prediction accuracy. The performance of these predictive models is rigorously evaluated using 10-fold crossvalidation. The results indicate that models incorporating crime descriptions alongside sociodemographic data significantly outperform those without such descriptions. Specifically, the XGBoost model demonstrated superior classification accuracy in the first experiment, while the Multi-Layer Perceptron model excelled in the second. These findings underscore the value of integrating comprehensive data sets in predictive modeling for effective crime prediction and prevention.

Date of Award	2023
Original language	English
Awarding Institution	Aberystwyth University