Abstract: In recent years, graph neural networks have been widely discussed in the field of deep learning. However, most of the researches are based on graph nodes and carry out classification and regression prediction under the premise of multi-dimensional attributes. Forecasting does not produce the desired results on single time-series of feature. This paper proposes a time-series graph convolutional network that can predict features in a complex graph network based on single time-series of feature of the node. By parameterizing the adjacency matrix in the traditional graph convolution network, the algorithm solves the problem of parameter degradation under a single feature condition, and combines the sequential learning method of the LSTM network to integrate the timing information into the training process, which improves the training accuracy. Experiments on the traffic flow data set PeMS and Los show that the prediction accuracy is better than that of mainstream algorithms such as GCN, T-GCN, GRU, LSTM.

Key words: graph convolutional network, single time-series of feature, LSTM, network prediction