Discovering patterns in traffic sensor data

We maintain a one of a kind, large-scale and high resolution (both spatially and temporally) traffic sensor dataset collected from the entire Los Angeles County road network. Traffic sensors (installed under the road pavement) are used to measure real-time traffic flows through road segments. In this paper, we exploit this dataset to rigorously verify two popular instinctive understandings about traffic flows on road segments: 1) each road segment has a typical traffic flow (known by local travelers) and one can often categorize road segments based on the similarity of their traffic flows, and 2) the road segments within each category not only have similar traffic flows but also are similar in their other characteristics (such as locality, connectivity). Toward this end, we developed a hypothesis analysis framework based on a variety of clustering and correlation evaluation techniques and leveraged this framework to respectively show the following. First, the set of road segments can indeed be partitioned into a set of distinct subpartitions with similar traffic flows, and there is a limited number of signature traffic patterns/labels each of which can accurately represent all traffic flows of a subpartition of the road segments. Second, all segments within each subpartition (represented by one signature) are also highly similar in three other characteristics, namely, direction, connectivity and locality. Our experiments verify our observations with high confidence.