GATC and DeepCut: Deep spatiotemporal feature extraction and clustering for large-scale transportation network partition

The network partition is an important method for many key transport problems, e.g., transport network zoning, parallel computing of traffic assignment problem, and analysis of the macroscopic fundamental diagram, to name a few. This paper designs two partition frameworks called GATC (Graph attention auto-encoder for clustering) and DeepCut, which can partition the transportation network into several components. These two frameworks combine unsupervised deep learning and clustering, taking into account both temporal factors and spatial factors. Firstly, the traffic flow time series data is encoded by graph attention auto-encoder, with graph structure and content considered. Secondly, the normalized cut method is used to partition the transportation network into several homogeneous sub-networks. DeepCut encodes the input data by a simple encoder, and the normalized cut method is used to partition the transportation network. The proposed methods are verified by a numerical example, which demonstrates the rationality and effectiveness of GATC and DeepCut for transportation network partition.