Forecasting Citywide Traffic Congestion Based on Social Media

This paper makes the first step towards mining citywide traffic congestion correlation by utilizing traffic related information from social media. Traffic congestion correlation mining, namely studying which road segments close to each other are highly likely to occur congestion simultaneously, is especially important to help many real applications, such as traffic prediction, traffic control, and urban transportation planning. Traditional traffic data collected from various sensors and other equipments are costly to obtain and hard to scale up to cover a entire city. With the rising popularity of social media, it is common for the public transportation systems and governments to share real time traffic information with the public through social media. It provides us great opportunities to study the traffic conditions of a city with the rich and easily available online data. However, it is also very difficult to use social media data to mine the citywide traffic congestion correlation due to the following major challenges: (1) Social media data like tweets in Twitter are usually noisy and hard to process, especially those tweets posted by individuals. (2) There lacks a method to study the citywide traffic congestion correlation. In this paper, instead of crawling all the traffic related tweets of a city, we only focus on utilizing the tweets posted by some particular organizations or governments. Tweets posted by them are more accurate and formal, thus it is much easier for traffic information extraction. We regard the traffic congestion correlation mining task as a spatio-temporal frequent pattern mining problem by considering each tweet reporting the traffic congestion of a particular road segment as a spatio-temporal item. A spatio-temporal frequent pattern mining algorithm TC_Apriori is also proposed to discover the road segment co-occurrence patterns in congestion. We use the tweets reporting the traffic information of Chicago to evaluate the proposed approach, and the results show that the proposed approach can effectively discover the road segment co-occurrence patterns in congestion.