Module graph_neural_network 

Graph Neural Network (GNN) with message-passing, edge weighting, and multi-layer propagation.

This module implements a full message-passing GNN where nodes aggregate neighbour features, transform them through stacked linear layers (with bias and activation), and iteratively refine node embeddings.

Quick example

use ipfrs_tensorlogic::graph_neural_network::{
    GraphNeuralNetwork, GnnConfig, GnnLayer, GnnActivation, GnnAggregation,
};

// One hidden layer: 2-dim input -> 4-dim output, ReLU
let layer = GnnLayer {
    weights: vec![
        vec![0.5, -0.3],
        vec![-0.1,  0.8],
        vec![ 0.2,  0.4],
        vec![-0.6,  0.1],
    ],
    bias: vec![0.0, 0.0, 0.0, 0.0],
    activation: GnnActivation::Relu,
};

let config = GnnConfig {
    layers: vec![layer],
    aggregation: GnnAggregation::Mean,
    num_iterations: 2,
};

let mut gnn = GraphNeuralNetwork::new(config);
let a = gnn.add_node(vec![1.0, 0.0]);
let b = gnn.add_node(vec![0.0, 1.0]);
gnn.add_edge(a, b, 1.0).expect("example: should succeed in docs");

let embeddings = gnn.forward();
assert_eq!(embeddings.len(), 2);
assert_eq!(embeddings[0].len(), 4);

Structs

GnnConfig

Full configuration for a GraphNeuralNetwork.

GnnEdge

Directed edge in the graph with an optional scalar weight.

GnnLayer

A single linear + bias + activation layer inside the GNN.

GnnNodeId

Newtype wrapper for node indices.

GnnStats

Snapshot of graph and model statistics.

GraphNeuralNetwork

Message-passing Graph Neural Network.

NodeFeatures

Feature vector associated with a single node.

Enums

GnnActivation

Activation function applied element-wise in a GnnLayer.

GnnAggregation

Neighbour feature aggregation strategy used during message passing.

GnnError

Errors that can arise when working with a GraphNeuralNetwork.

Functions

xorshift64

Xorshift64 pseudo-random number generator for test vector generation.