Text hashingvectorizeroptions

HashingVectorizerOptions

How many columns the hash lands in, and what it does with signs.

public sealed record HashingVectorizerOptions

Properties — NumFeatures (default 1048576, which is 2^20) is how many columns the matrix has, and therefore how often two different terms collide into one. AlternateSign (default true) gives half the terms a negative weight, so that collisions tend to cancel rather than accumulate. Norm (default SparseNorm.L2) is the norm each row is divided by, and is nullable: null leaves the rows unnormalized, which is scikit-learn's norm=None. Count carries the tokenization settings — the token pattern, lowercasing, n-gram range and the rest — because hashing changes only what happens after a document is cut into terms.

Example — a deliberately tiny feature space, so collisions are certain.

using Lodestar.Text.Vectorization;

var hv = new HashingVectorizer(new HashingVectorizerOptions { NumFeatures = 16 });
CsrMatrix hashed = hv.Transform(["the cat eats", "the dog eats", "the cat and the dog"]);

int columns = hashed.ColumnCount;  // => 16
double rowLength = hashed.RowL2Norm(0);  // => 1

Remarks — NumFeatures is the whole trade. Large enough and collisions are rare and the matrix is wide; small enough and the matrix is compact and two unrelated terms share a column. The default of 2^20 is scikit-learn's n_features=2**20, chosen so that collisions are negligible for most corpora while the matrix stays sparse — the columns cost nothing until something lands in them.

AlternateSign is the part that looks like a trick and is not. When two terms collide, one of them carrying a negative sign means the collision partly cancels instead of doubling, so the inner product between two documents stays close to what it would have been without the collision. It is alternate_sign=True in scikit-learn and on by default for the same reason.

There is no MinDf or MaxDf here, and there cannot be: both need document frequencies, and counting those would mean the pass over the corpus that this vectorizer exists to avoid.

Applies to — net10.0, netstandard2.0.

See also — HashingVectorizer, CountVectorizerOptions, SparseNorm, the Python equivalence table.