Identification of Freeway Traffic Congestion From Social Media Using a Hybrid Deep Learning Method: A Case Study

Abstract

The massive real-time data shared by Internet users provides a potentially rich resource for detecting traffic congestion. Targeting China's predominant social network platform, ‘Sina Weibo’, this paper proposes a hybrid deep learning method to mine valuable freeway traffic congestion information. Specifically, original microblog data is extracted and filtered via a customised web crawler coupled with geographical anchors. Afterwards, the selected microblogs undergo rigorous preprocessing, wherein a domain-specific Word2Vec model is trained to represent textual information as high-dimensional word embeddings. To effectively identify congestion-related microblogs, this study develops a ConvBILSTM model that integrates a TextCNN layer for capturing local textual features and a BILSTM layer for modelling global context dependencies. Extensive experimental evaluations demonstrate the superiority of the proposed method compared to benchmark approaches, achieving a recall of 0.8519 and an F1-score of 0.8415. Furthermore, the congestion-prone locations extracted from congestion-related microblogs based on Document Frequency scores are highly consistent with ground-truth data. Overall, this research facilitates timely and accurate reporting of traffic congestion, providing a valuable supplement or alternative to conventional freeway traffic surveillance methods.