This model uses a GNN architecture for multivariate time series forecasting. The adjacency matrix is learned from the data using a neural network, hence pre-defined graph topologies is not required.
The input for the model is a set of features at certain time points for a given number of nodes. A sliding window approach is used to create the labels for training. Depending on the window size and forecast length, the data is sliced to create the training data points.
ajayrkpe / gnn_multivariate Goto Github PK
View Code? Open in Web Editor NEWLicense: GNU General Public License v3.0