Multi-layer Generalized GCN Model (Hamilton 2020) — model_gcn

Stacks multiple layer_gcn_general() layers, each of which keeps separate weights for a node's own features and for its aggregated neighbors.

Usage

model_gcn_general(
  in_features,
  hidden_dims,
  out_features,
  activation = nnf_relu,
  out_activation = NULL,
  dropout = 0,
  normalize = TRUE
)

Arguments

in_features: Integer. Number of input features per node
hidden_dims: Integer vector. Dimensions of hidden layers (length = L)
out_features: Integer. Number of output features (typically 1 for regression)
activation: Function. Activation for hidden layers. Default: nnf_relu
out_activation: Function or NULL. Activation for output layer. Default: NULL
dropout: Numeric. Dropout rate (0-1) applied after each hidden layer. Default: 0
normalize: Logical. Whether to add self-loops and row-normalize the adjacency matrix. Default: TRUE

Value

Tensor n_nodes x out_features. Final predictions

Details

Architecture:

L hidden generalized GCN layers with configurable activation
1 output layer with optional output activation
Total layers = length(hidden_dims) + 1

When normalize = TRUE, self-loops are added and the adjacency matrix is row-normalized once in the forward pass, before any layer is applied.

Forward pass

model(x, adj)

x: Tensor n_nodes x in_features. Node feature matrix.
adj: Sparse COO tensor n_nodes x n_nodes. Adjacency matrix defining graph structure.

References

Hamilton, W. L. (2020). Graph Representation Learning. In Synthesis Lectures on Artificial Intelligence and Machine Learning. Springer International Publishing. doi:10.1007/978-3-031-01588-5

Examples

if (FALSE) { # torch::torch_is_installed()
adj <- adj_from_edgelist(from = c(1, 2, 3, 4), to = c(2, 3, 4, 1))
x <- torch::torch_randn(4, 14)

model <- model_gcn_general(14, c(32, 16), 1)
model(x, adj)

# Supply a pre-normalized adjacency matrix instead
adj_norm <- adj_row_normalize(add_graph_self_loops(adj))
model <- model_gcn_general(14, c(32, 16), 1, normalize = FALSE)
model(x, adj_norm)
}