Spatio-Temporal Graph Convolutional Networks: A Deep Learning Framework for Traffic Forecasting. https://arxiv.org/abs/1709.04875

It seems here are 3 major parameters we can decide in the paper, temporal conv channels in the Output block CTO, kernel size/radius in the temporal conv Kt as well as graph conv Ks.

So we conduct a hyperparameter optimization experiment that set the searching space as list follow:

CTO: [32, 64, 128]

Kt: [2, 3]

Ks: [2, 3, 4]

and run on metr-la dataset, 12 his point to 3 pred point, 15 epoch per combination, 7:1:2 train:val:test radio.

After about 620 experiments, as shown in the following figure we select the top 1% result in the test data of all experiments, and it's clearly show that a (CTO:64, Kt:2, Ks:4) combination can achieve better result.