Cosine Similarity Estimation Using FracMinHash: Theoretical Analysis, Safety Conditions, and Implementation

The increasing number and volume of genomic and metagenomic data necessitates scalable and robust computational models for precise analysis. Sketching techniques utilizing $$ -mers from a biological sample have proven to be useful for large-scale analyses. In recent years, FracMinHash has emerged as a popular sketching technique and has been used in several useful applications. Recent studies on FracMinHash proved unbiased estimators for the containment and Jaccard indices. However, theoretical investigations for other metrics, such as the cosine similarity, are still lacking.

In this paper, we present a theoretical framework for estimating cosine similarity from FracMinHash sketches. We establish conditions under which this estimation is sound, and recommend a minimum scale factor $$ for accurate results. Experimental evidence supports our theoretical findings.

We also present frac-kmc, a fast and efficient FracMinHash sketch generator program. frac-kmc is the fastest known FracMinHash sketch generator, delivering accurate and precise results for cosine similarity estimation on real data. We show that by computing FracMinHash sketches using frac-kmc, we can estimate pairwise cosine similarity speedily and accurately on real data. frac-kmc is freely available here: https://github.com/KoslickiLab/frac-kmc/.