gizml:cluster

clustering with a predict step: k-means (k-means++ and Lloyd), mini-batch k-means, BIRCH (CF subclusters, then Ward agglomeration by the nearest-neighbor chain), affinity propagation and mean shift. A model is a set of centers with a cluster label each: a row takes the label of its nearest center. Matrices are row-major.

import * as Cluster from "gizml:cluster";

Types

Centers

type Centers = { centers: Vec<f64>; count: i64; features: i64; labels: Vec<i64>; }

The result of clustering: the centers (row-major), their count and feature width, and the center of each sample.

• centers: Vec<f64>
• count: i64
• features: i64
• labels: Vec<i64>

MiniBatch

type MiniBatch = { centers: Vec<f64>; counts: Vec<f64>; k: i64; features: i64; }

Mini-batch k-means fed one batch at a time (scikit-learn's partial_fit): the first batch places the centers by k-means++, and each batch is one step.

• centers: Vec<f64>
• counts: Vec<f64>
• k: i64
• features: i64

Functions

kmeans

function kmeans(x: Vec<f64>, n: i64, d: i64, k: i64, maxIter: i64, tol: f64, rng: mut Random): Centers

k-means: k-means++ seeds, then Lloyd until the squared center shift is at most tol × the mean feature variance.

Effects: fail.

miniBatch

function miniBatch(k: i64, d: i64): MiniBatch

A mini-batch k-means of K centers over D features, before its first batch.

Effects: fail.

miniBatchKmeans

function miniBatchKmeans(x: Vec<f64>, n: i64, d: i64, k: i64, batch: i64, maxIter: i64, rng: mut Random): Centers

Mini-batch k-means: k-means++ on a sample of 3 batches, then batch steps with a per-center rate of 1 / count, until the smoothed batch inertia does not improve for 10 steps.

Effects: fail.

birch

function birch(x: Vec<f64>, n: i64, d: i64, threshold: f64, k: i64): Centers

BIRCH: CF subclusters of radius at most threshold (each row joins its nearest subcluster when the merged radius stays within it), then Ward agglomeration of their centroids into k clusters (k 0: each subcluster is a cluster).

Effects: fail.

affinityPropagation

function affinityPropagation(x: Vec<f64>, n: i64, d: i64, damping: f64, maxIter: i64, convergence: i64, rng: mut Random): Centers

Affinity propagation on negative squared distances, the preference the median similarity. No convergence in maxIter iterations gives no centers, as scikit-learn labels every row -1.

Effects: fail.

meanShift

function meanShift(x: Vec<f64>, n: i64, d: i64, bandwidth0: f64, maxIter: i64): Centers

Mean shift with a flat kernel from every row; the bandwidth (0: estimated as the mean distance to the 30% nearest neighbor), centers closer than the bandwidth to a stronger one removed.

Effects: fail.

Methods

Centers.nearest

function Centers.nearest(self: Centers, row: Vec<f64>): i64

The label of the center nearest to the row; -1 when there are no centers.

Effects: fail.

MiniBatch.step

function MiniBatch.step(self: mut MiniBatch, x: Vec<f64>, n: i64, rng: mut Random): bool throws string

One step of mini-batch k-means on the N rows of X: the first batch seeds the centers by k-means++.

Effects: fail.

MiniBatch.result

function MiniBatch.result(self: MiniBatch): Centers

The centers that the batches so far found, as a fitted model.

Effects: fail.