Struct HgnnNetwork 

pub struct HgnnNetwork { /* private fields */ }

Multi-layer HGNN network.

Layers alternate convolution and ReLU activation; the final layer has no activation applied (the caller is free to apply softmax / sigmoid as appropriate for the downstream task).

Implementations

impl HgnnNetwork

pub fn new(dims: &[usize], use_attention: bool, seed: u64) -> Self

Build a multi-layer HGNN.

Arguments

• dims – sequence of dimensions [in_dim, h1, h2, ..., out_dim]. Must have at least two elements.
• use_attention – whether every layer uses the HGAT attention variant.
• seed – base RNG seed; each layer receives seed + layer_index.

Panics

Panics if dims.len() < 2.

pub fn forward( &self, incidence: &Array2<f64>, node_feats: &Array2<f64>, edge_weights: Option<&Array1<f64>>, ) -> Result<Array2<f64>>

Forward pass through all layers.

ReLU is applied after every hidden layer; the output of the last layer is returned without activation.

Arguments

• incidence – incidence matrix H: [n_nodes × n_edges]
• node_feats – node feature matrix X: [n_nodes × in_dim]
• edge_weights – optional per-hyperedge weight vector

Returns

Output feature matrix: [n_nodes × out_dim]

pub fn n_params(&self) -> usize

Total number of learnable parameters across all layers.

pub fn depth(&self) -> usize

Number of layers.