The traffic prediction problem has received significant attention in terms of providing solutions to reduce transportation time. Due to the dependency of traffic conditions on time, there is a need for a method that can predict traffic in the shortest possible time. Traffic data can be modeled as temporal graphs. In this thesis, we propose two solutions aimed at improving the execution time of graph neural networks for traffic prediction. Firstly, by employing knowledge distillation with the goal of reducing the execution time of neural networks, we introduce a cost function. This cost function is designed to train a network with fewer parameters using data distilled from a complex network, in such a way that its accuracy remains close to that of the intricate network. Subsequently, we introduce an algorithm that, using this cost function, can concurrently obtain a network. This network is acquired not only by pruning and removing less significant neurons in a complex network ( teacher ), but also by training it using distilled data from the teacher network. The training aims to maintain the accuracy of the pruned network ( student ) at a level close to that of the teacher network, while significantly reducing the number of parameters. These two proposed solutions offer the following advantages over previous methods. A novel approach in training the student network involves utilizing the distillation of spatial-temporal correlations from the teacher network. This is aimed at learning complex spatial and temporal patterns comprehended by the teacher network. Outlier detection in the training dataset is conducted, followed by their replacement with predictions from the teacher network. The pruning algorithm, on one side, is focused on determining the structure of the student network through the removal of low-importance neurons based on the teacher network. On the other side, it simultaneously trains the network resulting from each pruning step using the distillation of student knowledge. Ultimately, we eva‎luate our proposed cost function and algorithm on two datasets, PeMSD7 and PeMSD8. The results demonstrate that each of these two approaches has exhibited improvements over previous methods in the related domain.

مهران صفاياني

عبدالرضا ميرزايي

شادرخ سماوي , زينب مالكي