Fig 1.
The overall structure diagram of the model.
The time series of input data is duplicated into two copies: (1)one passing through the dilated convolution model on the top to learn temporal features; (2)the other passing through the RNN model on the bottom to learn spatial features.
Fig 2.
The temporal convolutional feature extraction module consists of two parallel convolutional branches: A standard 1D convolution branch and a dilated 1D convolution branch.
The input sequence is processed by both branches simultaneously to capture multi-scale temporal patterns. The outputs from both branches are concatenated to form a feature matrix, which is then passed through a ReLU activation function to generate the final temporal feature matrix.
Fig 3.
The Graph Attention (GAT) layer takes the temporal feature matrix and the spatial feature matrix as inputs.
The GAT layer computes attention scores between nodes using learnable weight matrices. The attention scores are used to weight the importance of neighboring nodes, enabling the model to adaptively capture complex spatial dependencies. The output of the GAT layer is an integrated feature matrix that combines both temporal and spatial information.
Table 1.
Overview of the datasets used in the experiments.
Table 2.
RMSE and PCC metrics for different models on the first three datasets with forecast weeks set at 2, 3, 5, 10, 15.
Table 3.
RMSE metrics for different models on the last dataset with forecast weeks set at 3, 5, 7, 11, 15.
Table 4.
Ablation experiments on graph structures and RMSE metrics for the first three groups of influenza datasets with forecast weeks at 2 and 10.