Lodestar.Text 0.6.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

LshBanding

How a MinHash signature is cut into bands for locality-sensitive hashing.

public readonly record struct LshBanding(int Bands, int RowsPerBand)

Parameters — Bands is b, how many bands the signature is cut into. RowsPerBand is r, how many signature slots each band holds.

Example — the banding a threshold of 0.8 asks for over 128 permutations.

using Lodestar.Text.Similarity;

LshBanding banding = LshBanding.Solve(0.8, 128);

int bands = banding.Bands;  // => 9
int rows = banding.RowsPerBand;  // => 13
int used = banding.Permutations;  // => 117

Members — one page each.

Member	What it does
LshBanding.Solve	The banding minimizing the weighted error at a threshold
LshBanding.CollisionProbability	The S-curve this banding produces

Remarks — b and r are inputs, and the solve that picks them is a function you can call. The reference hides the same solve inside a constructor, which is what makes the trade-off hard to see: a threshold does not select a banding on its own — it selects one given how much a caller minds a false positive against a false negative.

The solve rarely uses every slot. Nine bands of thirteen is 117 of the 128 available, because b × r has to divide into the length and the best pair usually leaves a remainder. LshIndex therefore accepts a signature longer than the banding needs and ignores the tail rather than refusing it.

More bands finds more and verifies more; more rows per band finds less and wastes less. The S-curve 1 - (1 - s^r)^b is the whole story, and CollisionProbability draws it.

Applies to — net10.0, netstandard2.0.

See also — LshIndex, MinHash, the set-similarity index.