ts-classify

Fast text classification via SVM, compiled to WebAssembly.

Features

• Sparse vectors - efficient for high-dimensional data (text, trigrams)
• Binary serialization - train once, load instantly
• Written in Rust, compiled to WASM for near-native performance in the browser

Installation

npm install ts-classify

Classifiers

All classifiers use the same sparse data format: flat Float64Array of [index, value, index, value, ...] pairs.

import init, { JsOneVsRestSVM, JsMulticlassSVM, JsNearestCentroid, JsSVM } from 'ts-classify';

await init();

JsOneVsRestSVM

Multiclass SVM using one-vs-rest strategy. Trains K classifiers for K classes. Best for large numbers of classes.

const svm = new JsOneVsRestSVM();
svm.train(samplesFlat, sampleLengths, labels);
const prediction = svm.predict(sampleFlat);
const margins = svm.margins(sampleFlat); // [class0, margin0, class1, margin1, ...]

JsMulticlassSVM

Multiclass SVM using one-vs-one strategy. Trains K(K-1)/2 classifiers, predicts by voting. Often more accurate for smaller numbers of classes.

const svm = new JsMulticlassSVM();
svm.train(samplesFlat, sampleLengths, labels);
const prediction = svm.predict(sampleFlat);

JsNearestCentroid

Computes the centroid of each class and predicts by cosine similarity. Very fast to train, no hyperparameters.

const nc = new JsNearestCentroid();
nc.train(samplesFlat, sampleLengths, labels);
const prediction = nc.predict(sampleFlat);
const similarities = nc.margins(sampleFlat); // [class0, sim0, class1, sim1, ...]

JsSVM

Binary SVM for two-class problems. Labels are -1.0 or 1.0.

const svm = new JsSVM();
svm.train(samplesFlat, sampleLengths, labels); // labels: Float64Array of -1.0/1.0
const prediction = svm.predict(sampleFlat);     // returns -1.0 or 1.0
const margin = svm.margin(sampleFlat);           // raw decision value

Data format

// Flat sparse format: [idx, val, idx, val, ...]
const samplesFlat = new Float64Array([
    0, -2.0, 1, 0.0,  // sample 0: 2 pairs
    0, 1.0, 1, 1.0,   // sample 1: 2 pairs
]);
const sampleLengths = new Uint32Array([2, 2]);  // pairs per sample
const labels = new Int32Array([0, 1]);           // class IDs (or Float64Array for JsSVM)

Serialization

All classifiers support binary serialization:

const bytes = svm.to_bytes();
const loaded = JsOneVsRestSVM.from_bytes(bytes);

Performance

Scaling (coordinate descent, OvR):

| Classes | Classifiers | Train | Predict | |---------|-------------|-------|---------| | 10 | 10 | 511µs | 144ns | | 50 | 50 | 560µs | 743ns | | 100 | 100 | 1.4ms | 1.6µs | | 5000 | 5000 | ~50-100ms | ~80µs |

Building from source

Requires Rust and wasm-pack.

wasm-pack build --target web --no-default-features

Run tests (native):

cargo test --release

License

MIT