Predicting Traffic Congestions with Global Signatures Discovered by Frequent Pattern Mining

We propose a traffic jam prediction method based on mining frequent patterns correlated to traffic jams. For traffic jam prediction at a given sensor, first, we apply a one-dimensional clustering scheme to identify automatically which sensors are and in what degree correlated to the given sensor in terms that certain volume values with a compact distribution co-occur frequently with the traffic jams of a certain time lag at the given sensor. Then, such co-occurred frequent patterns are represented in an abstract way using Gaussian models. Finally, we score the jam possibility via the weighted sum of the probability that every sensor data belongs to the corresponding Gaussian model. By applying the proposed method, we found that signatures related to traffic jams exist widely in the road network, not a local region, where over 3000 sensors provide information contributive to traffic congestion prediction at every given sensor, and some low-volume patterns act also as signals to warn upcoming traffic jams. The mechanism of the proposed method is different from the existing methods in that the proposed method seeks signatures of traffic jams from globally unbalanced traffic flow distribution.