Keywords: Co-training, Temporal Graphs, Dynamic Graphs, Semi-Supervised Learning, Graph Neural Network, Traffic Forecasting, Pseudo-values, Regression
Abstract: Graphs are widely in use to model related instances of data attributed with properties providing rich spatial information. Although a lot of classical graph-related problems have been solved with the advent of Graph Neural Networks (GNN), Spatio-Temporal data poses a new challenge. We propose GraphCoReg: a novel methodology to perform regression on spatio-temporal data, in a Semi-Supervised Learning (SSL) setting using co-training. Our co-training approach exploits two model-based views of the dataset using two temporal Graph Neural Networks (GNNs) - an Attention-based GNN (ASTGCN) and a Long Short Term Memory GNN (GCLSTM). Additionally, methodologies to incrementally add the pseudo-targets to training data have been described. We finally compare the performance of the semi-supervised model with equivalent supervised models. This approach has been tested on the MetrLA dataset for traffic forecasting. This is a work-in-progress to investigate the performance of GraphCoReg on multiple benchmark spatio-temporal datasets for the task of regression on temporal graphs.