A convolutional neural network for traffic information sensing from social media text

Abstract

Mining social media data to obtain traffic relevant information is an emerging topic due to the real-time and ubiquitous features of social media. In this paper, we focus on a specific issue in social media mining that concerns to extract traffic relevant microblogs from the Sina Weibo platform, which is the first and essential step to further extract detailed traffic information, such as the location of a traffic incident. It is transformed into a machine learning problem of short text classification. We employ deep neural networks to classify microblogs into traffic relevant and traffic irrelevant ones. More specifically, we firstly adopt the continuous bag-of-word (CBOW) model to learn word embedding representations based on the dataset of three billion unlabeled microblogs. Next we use a convolutional neural network (CNN) to learn the abstract features of traffic relevant and traffic irrelevant microblogs. The key advances in this paper are: use of semantics of words and deployment of deep neural networks to extract traffic information from social media text. Experiments show that the proposed deep learning method has superior performance over support vector machine (SVM) based method and multi-layer perceptron (MLP) based method.