IDENTIFICATION OF SIMILARITIES AND PREDICTION OF UNKNOWN FEATURES IN AN URBAN STREET NETWORK
Keywords: urban traffic, traffic analysis, machine learning, clustering, spatio-temporal data, data integration, data mining, floating-car-data
Abstract. Accurate predictions of the characteristics of urban streets in particular with respect to the typical traffic situations are crucial for numerous real world applications such as navigation, scheduling of logistic and public transportation services as well as high-level planning of infrastructure which may include planning of construction sites or even changes of the road topology. However, this information may be hard to obtain, especially in complex urban road networks where interdependencies between roads are highly present. In addition, accurate and recent traffic data is not always available, especially for uncommon situations like large-scale public events, traffic accidents or construction sites. This work demonstrates how to employ historical traffic datasets in conjunction with other, infrastructure related data, to derive a deeper understanding of urban traffic behaviour. In particular this paper provides the following contributions: (1) the generation of meaningful features to describe the segments in urban road networks; (2) an unsupervised machine learning approach that identifies similar segments based on those features; (3) a supervised approach to predict unknown features of the segments and, finally, (4) an extensive evaluation of the extracted road characteristics and the proposed methods using real-world data. The resulting clusters reveal the similarities of the street segments and give a different perspective on the road network and the traffic situation, respectively. The experiments on the classification approach demonstrate that unknown features can be predicted with a good quality.