A Deep Learning-based Traffic Event Detection From Social Media

Traffic events, such as congestion and accidents, lead to driver frustration, longer travel times, blocked emergency vehicles, and increased fuel consumption. Timely detection of traffic events is necessary for trip planning and restoring smooth operating conditions. Social media platforms allow users to post about their interests and experiences. Traffic is one of the topics for which users generate freely and publicly available content. Thus, extracting and analyzing such user-generated content can yield valuable traffic information. Detecting traffic information from social media data using machine learning, in the literature, is mostly limited to detecting whether or not they contain traffic-related information. This study goes one step further to detect specific traffic-related information from social media, such as accidents, congestion, disabled vehicle, road closure, road hazard, work/construction zone, and traffic condition information. To this aim, by adjusting kernel size and the number of filters, the experimental results showed that, with the proper setting of the kernel size and number filters, convolutional neural network outperforms bidirectional long-short term memory network, multi-layer perceptron, random forest, and support vector machines, in terms of F1 score and generalization accuracy.