Optimal Hyperparameter Tuning using Meta-Learning for Big Traffic Datasets

Big traffic data is an emergent issue of Intelligent Transportation System (ITS) since a large of data can be produced every day. Consequently, traffic flow prediction becomes the main component for developing ITS in terms of providing accurate and timely traffic information. In this study, we take the traffic forecasting problem into account using Deep learning (DL) models. Specifically, traffic data are from many devices on the roads in which training by DL models have to face with expensive problems (e.g., time-consuming and human expertise). Therefore, we focus on hyperparameter tuning for training big traffic datasets using meta-learning to improve the automatic learning process and reduce time-consuming tasks. Regarding the experiment, we take data from the Vehicle Detection System (VDS) as the case study for evaluating our approach. Specifically, data have collected from 21 sensors which are located in an urban area. The experiments show promising results of our proposed approach for training multiple traffic datasets.