Struct HgnnLayer 

pub struct HgnnLayer { /* private fields */ }

A single HGNN convolution layer.

Stores the learnable weight matrix Θ of shape [in_dim × out_dim]. Optionally stores attention weight vectors when use_attention = true.

Implementations

impl HgnnLayer

pub fn new(config: HgnnLayerConfig, seed: u64) -> Self

Create a new HGNN layer with Xavier-uniform initialised weights.

Arguments

• config – layer configuration
• seed – RNG seed for reproducible weight initialisation

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

Forward pass: X' = Ã X Θ

à = D_v^{-1/2} H diag(w / D_e) H^T D_v^{-1/2}

Arguments

• incidence – incidence matrix H of shape [n_nodes × n_edges]
• node_feats – node feature matrix X of shape [n_nodes × in_dim]
• edge_weights – per-hyperedge weight vector (length = n_edges); if None, all weights default to 1.0

Returns

Node feature matrix of shape [n_nodes × out_dim]

pub fn relu(x: Array2<f64>) -> Array2<f64>

Element-wise ReLU activation.

pub fn n_params(&self) -> usize

Number of learnable parameters in this layer.