Feature Fusion Models for Deep Autoencoders: Application to Traffic Flow Prediction

Abstract

Due to reduction in dimensionality and extraction of the definitive features of input data, deep architectures have achieved significant success in various machine learning applications. Considering their successful applications in speech recognition and image classification, the main goal of this research is to investigate the performance of the sparse autoencoders utilized in regression analysis. To this end, deep sparse autoencoders with the standard method of training, cascaded, and partially cascaded architectures, fed with the fusion of low- and high-level features, are proposed and implemented. The regression task is to forecast the vehicular flow rate of a location on an arterial highway using different traffic variables of several locations ahead in the Twin Cities Metro area of Minneapolis. The results demonstrate that the partially cascaded model exhibits advancements in yielding more accurate results than the other two architectures fed with the features that correlate the most to the traffic flow rate.