Struct GraphSageLayer 

pub struct GraphSageLayer {
    pub weights: Array2<f64>,
    pub bias: Option<Array1<f64>>,
    pub in_dim: usize,
    pub out_dim: usize,
    pub aggregation: SageAggregation,
    pub use_relu: bool,
    pub normalize: bool,
}

A single GraphSAGE layer.

Concatenates each node’s own representation with the aggregated neighbor representation, then applies a linear transformation:

  h_v = σ( W · concat(h_v, AGG({h_u : u ∈ N(v)})) + b )

The output is L2-normalized (per-node) following the original paper.

Fields

weights: Array2<f64>

Weight matrix [2 * in_dim, out_dim]

bias: Option<Array1<f64>>

Optional bias [out_dim]

in_dim: usize

Input feature dimension

out_dim: usize

Output feature dimension

aggregation: SageAggregation

Aggregation strategy

use_relu: bool

Apply ReLU activation

normalize: bool

Apply L2 normalization on output embeddings

Implementations

impl GraphSageLayer

pub fn new(in_dim: usize, out_dim: usize) -> Self

Create a new GraphSAGE layer with Glorot-uniform initialization.

Arguments

• in_dim – Input feature dimension per node.
• out_dim – Output embedding dimension.

pub fn with_aggregation(self, aggr: SageAggregation) -> Self

Use a specific aggregation strategy.

pub fn without_normalize(self) -> Self

Disable L2 normalization on output.

pub fn without_activation(self) -> Self

Disable ReLU activation.

pub fn forward( &self, adj: &CsrMatrix, features: &Array2<f64>, ) -> Result<Array2<f64>>

Forward pass.

Arguments

• adj – Sparse adjacency (row i → outgoing neighbors of i).
• features – Node feature matrix [n_nodes, in_dim].

Trait Implementations

impl Clone for GraphSageLayer

fn clone(&self) -> GraphSageLayer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for GraphSageLayer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.