vO.1.6 Release Notes

This release focuses on several BlueGraph’s dependency issues, fixes several bugs and adds new features to the bluegraph.downstream.similarity module.

Downstream tasks

Similarity search

Added new similarity metrics for similarity indices:

• Wasserstein distance

• Kullback–Leibler divergence

These metrics can be used with ScikitLearnSimilarityIndex (based on ball-trees). The following snippets of code illustrate the usage of new metrics.

from bluegraph.downstream.similarity import ScikitLearnSimilarityIndex, SimilarityProcessor

vectors = [
    [0, 0, 1],
    [1, 0, 0],
    [0, 1, 1]
]

ws_index = ScikitLearnSimilarityIndex(
    dimension=3,
    similarity="wasserstein",
    initial_vectors=vectors)

distances, indices = ws_index.search([[1, 1, 0]], k=1)
closest_vectors = ws_index.reconstruct(indices)

>>> closest_vectors
array([[[1., 0., 0.]]])

The following code illustrates the usage of the Kullback–Leibler divergence:

kl_index = ScikitLearnSimilarityIndex(
    dimension=3,
    similarity="kl",
    initial_vectors=vectors)

distances, indices = kl_index.search([[1, 1, 0]], k=1)
closest_vectors = kl_index.reconstruct(indices)

>>> closest_vectors
array([[[0., 0., 1.]]])