Particle Swarm Optimization of Deep Learning Models for the Determination of Vehicle Hijacking Tweets

Abstract The problem of vehicle hijacking is increasingly prevalent in many parts of the world, including South Africa, causing travelers to live in constant fear of becoming victims. To address this, this work employs the usage of social media data to identify hijacking incidents in urban environments by using the city of Cape Town in South Africa as the case study. More specifically, the incorporated data is obtained from the Twitter social media platform. However, this form of data tends to be informal and unstructured. Hence, this led to the issue of data relevance, which was addressed by using Machine Learning algorithms to classify tweets. Two deep learning algorithms were employed: a Convolutional Neural Network (CNN), and a Multi-layer Feed-forward Neural Network (MLFNN). However, due to model complexity, Particle Swarm Optimization (PSO) was also implemented to produce the model variations, PSOCNN and PSOMLFNN. It was found that the PSO-optimized models outperformed the standard non-PSO models. The PSOMLFNN approach was also identified to be the best-performing model amongst the employed selection by obtaining an accuracy of 96.15%. Hence, it was successfully employed along with the Tweet data to produce vehicle hijacking maps. In future work, further machine learning and optimization algorithms will be tested to identify the best-performing algorithms for this task as well as extend the hijacking maps to a mobile application for drivers on the South African roads.