A Data Mining Approach to Creating Fundamental Traffic Flow Diagram

Abstract This paper investigates application of clustering techniques in partitioning traffic flow data to congested and free flow regimes. Clustering techniques identify the similarities and dissimilarities between data, and classify the data into groups with similar characteristics. Such techniques have been successfully used in market research, astronomy, psychiatry, and transportation. A framework is proposed for clustering traffic data based on fundamental traffic flow variables. Three types of clustering techniques are investigated: 1) connectivity-based clustering, 2) centroid-based clustering, and 3) distribution-based clustering. Specifically, hierarchical clustering, K-means clustering and general mixture model (GMM) were investigated. Traffic sensor data from three freeway bottleneck locations in two major U.S. metropolitan areas, St. Louis, Missouri, and Twin Cities, Minnesota, were used in the study. Various combinations of traffic variables were investigated for all three clustering techniques. The results indicated that the clustering is an effective way to partition traffic data into the free flow and congested flow regimes. Partitioned traffic data can be used to create fundamental traffic flow diagrams and macroscopic traffic stream models. Using speeds, or both speeds and occupancies as input variables produced the best clustering results. The performance of K-means and hierarchical clustering techniques were comparable to each other and they outperformed GMM clustering.