compute_hub_score

compute_hub_score(
    embeddings,
    k=30,
    n_iterations=30,
    neighbors=None,
    similarities=None,
)

Compute hub score for each point in the embedding space.

Hub score identifies structural hub nodes that hold the embedding space together, as opposed to popular instances that fill it. The score combines: - Eigenvector centrality: rewards being in a tightly woven neighborhood - Distance to global centroid: rewards central position in the space

General/structural concepts score high on both signals; specific instances score low on both.

Parameters

Name	Type	Description	Default
embeddings	np.ndarray	Normalized embeddings (n, d)	required
k	int	Number of neighbors for k-NN graph	30
n_iterations	int	Iterations for eigenvector power method	30
neighbors	Optional[np.ndarray]	Pre-computed k-NN indices (n, k+1). If None, computed internally.	None
similarities	Optional[np.ndarray]	Pre-computed similarities (n, k+1). If None, computed internally.	None

Returns

Name	Type	Description
	HubScoreResult	HubScoreResult with scores and component signals.

Example

result = compute_hub_score(embeddings) hub_nodes = result.get_hub_nodes(threshold=1.0) print(f”Found {len(hub_nodes)} hub nodes”)

Use for taxonomy direction

if result.scores[idx_animal] > result.scores[idx_dog]: … print(“Animal is more general than Dog”)