Random Forest (JavaScript + WebAssembly)

This package includes a WebAssebly port of the native C++ package RandomForests wrapped in a JavaScript API. It works in most modern browsers and Node.js. There's experimental support of async computation and multithreading in the browsers, using WebWorkers. Unfortunately, no tests yet. You can find examples in the example folder.

Supports regression and classification tasks
Faster than existing JS implementations
Supports sync and async modes
Permutation feature importance
Threads (WebWorkers)
CommonJS module

Why Random Forests

Compared to R, Python, or Julia, the JavaScript ecosystem doesn't have many machine learning packages. Of course, there's Tensorflow.js, but sometimes neural nets are not such a good option (e.g., small tabular data). Random forest is a robust ensemble method that fits multiple decision trees under the hood using different subsets of original data. The algorithm has a pretty simple nature but is very powerful, catching non-linear dependencies between variables. It's also ideal for parallelization and implementation in a native language, such as C or C++. Luckily, WebAssembly makes it possible to take an efficient and fast native lib and port it to browsers and Node.js.

Install

npm install -S random-forest

Sync mode

Init

const { RandomForestClassifier, RandomForestRegressor } = require('random-forest')

const rf = new RandomForestClassifier({
  nEstimators: 100,
  maxDepth: 10,
  maxFeatures: 'auto',
  minSamplesLeaf: 5,
  minInfoGain: 0,
})

Training, Predicting

rf.train(Xtrain, ytrain)
const ypred = rf.predict(Xtest)

Saving, Loading models

const model = rf.save()
fs.writeFileSync('example.model', model)
const modelLoaded = new Uint8Array(fs.readFileSync('example.model'))
rf.load(modelLoaded)
const ypred = rf.predict(Xtest)

Some browsers doesn't allow running WebAssembly in a sync mode. In such case, you can try async mode described below.

Async mode

const { RandomForestClassifier, RandomForestRegressor } = require('random-forest/async')

// ! Don't miss /async part in require('random-forest/async') !

;(async function f () {
  const rf = new RandomForestClassifier({
    nEstimators: 100,
    maxDepth: 10,
    maxFeatures: 'auto',
    minSamplesLeaf: 5,
    minInfoGain: 0,
    nJobs: 4 // Control the number of threads (workers) with this param
  })
  await rf.init()
  await rf.train(Xtrain, ytrain)
  const ypred = await rf.predict(Xtest)
  console.log(ypred, ytest)
})()

Currently the async mode doesn't support loading/saving models.

Web demo (feature importance)

The random-forest lib is used to calculate variable importance and peform feature selection on StatSim Select and StatSim Vis

Development

Contributions are very welcomed. Some insights on how everything works:

Building steps:

The native code is loaded from the native-forest repo, a fork from RandomForests, a C++ implementation of random forests
Custom C++ interfaces are in src/api.cpp and src/api.h.
Emscripten compiles the native-forest code with defined interfaces into native/native.js and native/native.wasm. Compilation settings located in Makefile
To load WebAssembly in sync mode, prepare-wasm.js script converts the wasm file into a Uint8 array and stores it in the wrappers folder
Then src/base.js loads wrapper/native.bin.js as a regular CommonJS module, initializes it using the native/native.js module utils and then inititalizes native functions with cwrap
That's all what needed for the sync mode to work. Now prepare async version. To make it easier loading and bundling the module, a WebWorker script is bundled, rather than uses importScript. It's also loaded not as a separate file, but Blob. To generate the Blob we need the worker to be compiled first, then loaded as a string
Bundle src/worker.js into dist/worker.js
Use prepare-worker.js to read code of dist/worker.js and save it as a module in wrapper/worker.code.js
Load wrapped code in src/async.js, init Blob, the URL, and WebWorkers
In async mode results are aggregates

syedzayyan/random-forest-wasm