Struct GatEmbedder 

pub struct GatEmbedder { /* private fields */ }

Multi-head graph attention network embedder trained on RDF triple lists.

Architecture follows Veličković et al. (2018): for each node the model attends over its in-neighbourhood using learned Q/K/V projections, applies LeakyReLU-gated softmax attention, concatenates the num_heads outputs, projects through W_out, applies ReLU, and L2-normalises.

Training uses a margin-ranking loss and sign-SGD with gradient clipping.

Implementations

impl GatEmbedder

pub fn new(config: GatEmbedderConfig) -> Self

Create a new, un-trained GAT embedder.

pub fn fit( &mut self, triples: &[(String, String, String)], ) -> Result<(), EmbeddingError>

Train on (subject_iri, predicate_iri, object_iri) triples.

Steps:

1. Build entity index and adjacency map (undirected: s→o and o→s).
2. Xavier-initialise all Q/K/V and W_out weight matrices.
3. Initialise random L2-normalised entity feature vectors.
4. For each epoch: attention forward-pass → margin-ranking loss → sign-SGD weight update.
5. Cache final embeddings.

pub fn embed_entity(&self, entity: &str) -> Vec<f64>

Return the embedding for a known entity IRI. Returns a zero vector for unseen entities (inductive fallback — never panics).

pub fn attention_forward( &self, entity_idx: usize, adj: &HashMap<usize, Vec<usize>>, embeddings: &[Vec<f64>], layer_idx: usize, ) -> Vec<f64>

Multi-head attention forward pass for a single entity.

For each head h: Q_h = W_query[h] · e_i K_h = W_key[h] · e_j for each neighbour j V_h = W_value[h] · e_j score_j = LeakyReLU(Q_h · K_h_j / √head_dim) α_j = softmax({score_j}) head_out_h = Σ_j α_j · V_h_j

Heads are concatenated → projected by W_out → ReLU → L2-normalised.

pub fn is_trained(&self) -> bool

Whether fit has been called successfully.

pub fn num_entities(&self) -> usize

Number of distinct entities seen during training.

pub fn embedding_dim(&self) -> usize

Dimension of each output embedding.

Trait Implementations

impl Debug for GatEmbedder

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

Formats the value using the given formatter.