1374114 - Graph-Based Clustering and Kriging of Traffic Sensor Networks for Faulty Data Detection

The existing quality control rules for traffic data disregard spatiotemporal correlations among adjacent sensors. Inspired by the power of graph neural networks for spatiotemporal encoding, this study explores an inductive graph neural network kriging (IGNNK) approach for detection of faulty traffic data based on cluster-formed subgraphs, where nodes refer to sensor locations on a highway network and edges encode network connectivity between the sensors governed by network topology. First, subgraphs are generated from the clusters of sensors based on k-mean clustering analysis conducted in the feature space of encodings by a graph auto-encoder (GAE). The GAE encompasses dual encodings from two separate graph neural networks: (1) a graph attention network to reconstruct adjacent matrix, and (2) a diffusion graph convolutional network to reconstruct node features. The GAE is trained by minimizing linear combination of the two reconstruction losses. The cluster-formed subgraphs are used to train the IGNNK model by randomly masking some nodes of the subgraphs and then reconstructing them. Finally, faulty traffic data is detected by evaluating the statistics of kriging errors by the pretrained IGNNK. Our results show the effectiveness of the proposed GAE-based clustering method and the power of the IGNNK model for faulty traffic data detection.