Congestion prediction using adaptive boosting machine learning classifiers

Congestion is a challenge that commuters have to deal with on a daily basis. Consequently, predicting the future status of a roadway is valuable for travelers in making better travel decisions. The deployment of stationary sensors and the proliferation of mobile vehicle probes provide researchers with a wealth of historical and real-time data that can be used for the automatic prediction of congestion along freeway segments. In this paper we introduce anew algorithm for the automatic prediction of congestion using Adaptive Boosting machine learning classifiers. The proposed algorithm creates the learning dataset by identifying congested sections using a skewed distribution mixture model of speed data to create a binary congestion matrix. The elements of this binary matrix are then used as responses in the training of the classifiers. The predictors for the classifier during the training phase are windows (slots) of the historical spatiotemporal speed matrix. In the real-time running phase, the classifiers use the most recent spatiotemporal speed matrix window to predict the short- and medium-term (up to 100 minutes into the future) status of the roadway. Experimental results using archived data from a 22-mile section of the northbound Interstate 5 (I-5) corridor in the Portland, Oregon, metropolitan region demonstrates promising high true positive and low false positive rates. Specifically, using a relatively large number of weak learners (between 20 and 30 learners) the achieved true positive prediction rate is slightly greater than 0.99 and the false positive rate is less than 0.0001.