vspinu / simdist

High performance similarity and distance metrics for sparse representations

High performance distances and similarities for various dense and sparse representations with primary focus on applications in NLP and recommender systems.

Supported and Planned Object Types

• matrix from base R
• dgCMatrix, dgRMatrix and dgTMatrix from Matrix package
• simple_triplet_matrix from slam package
• data.frames in primary-secondary-value (psv) format
• list of named numeric or character vectors

Distances for 2D Representations

	matrix	dgCMatrix	dgRMatrix	dgTMatrix	slam	psv	list
cosine	✔	✔	✔	✔		✔
euclidean	✔	✔	✔	✔		✔
mahalanobis
jaccard

Aggregation Distances for 3D Representations

	dgCMatrix	dgRMatrix	dgTMatrix	slam	psv	list
centroid	✔	✔	✔		✔
semantic_min_max1	✔	✔	✔		✔
semantic_min_sum2	✔	✔	✔		✔

[1] More commonly known as "Relaxed Word Mover Distance" (RWMD) proposed in Kusner et. al. ‘From Word Embeddings To Document Distances’ (2015).

[2] Similar to RWMD measure, proposed in Mihalcea et.al. 'Corpus-Based and Knowledge-Based Measures of Text Semantic Similarity' (2006)

Transformations

norm_l1, norm_l2.

simdist's Issues

Extend support for binary representations.

Allow for custom dist matrix to `adist`s