Short term traffic speed prediction using different feature sets and sensor clusters

Accurate short term traffic speed prediction has been one of the most important issues of Intelligent Traffic Systems. In literature there have been many works that apply prediction models on few amounts of sensors' data for a short time training and test period. Unlike most of the previous works, in this paper we used 122 speed sensors' data from Istanbul that was collected from 1st January to 31st December 2014. We extracted four different feature sets for regression algorithms. Then we clustered the sensors into different groups and trained a model for each group. Prediction results are obtained by using decision tree and KNN based regression algorithms. Experimental results show that KNN based regression algorithm generally outperforms decision trees and training a single KNN model for each sensor is better than training a general model for all sensors. On the other hand, incorporating weather data doesn't help to improve the performance beyond our expectations.