Crime Prediction Model using Deep Neural Networks

Abstract

This project investigates the feasibility of using machine learning techniques, specifically neural networks, to make prediction on criminal behavior based on the history of the arrest bookings. The experiment will have to handle imbalanced data frequencies. To combat the challenge, data augmentation and weighted loss function is being developed to extract information from the minority classes. For this project, we have focused on how neural networks can be advantageous in classification of crime prediction. The specific kind of neural network that has been used in the project is a deep fully connected neural network. Fully connected neural networks are suitable for problems where domain knowledge is limited and many to many relations between features are important. As this report shows, machine learning techniques could definitely be of use for classification of criminal behavior, and we recommend exploring the discussed data augmentation and modeling methods more thoroughly to improve on the results and find new patterns.