@pnwatt/loess

loess

Fork of a JavaScript implementation of the Locally-Weighted Regression package originally written in C by Cleveland, Grosse and Shyu (1992). This fork addresses security issues with project dependencies. The dependencies have been updated to fix the security issues.

Getting started

First install the package:

npm install @pnwatt/loess --save

Load in your data:

var data = require("./myData.json");

Instantiate a LOESS model with the data:

var Loess = require("loess");
var options = { span: 0.5, band: 0.8, degree: 1 };
var model = new Loess(data, options);

Fit model by calling the .predict( ) method on the model object:

var fit = model.predict();
console.log(fit.fitted);
// do something else with fit.fitted

To fit model on a new set of points, pass a data object into .predict( )

var newData = {
  x: [1, 2, 3, 4, 5],
  x2: [6, 7, 8, 9, 10],
};

fit = model.predict(newData);

var upperLimit = fit.fitted.map((yhat, idx) => yhat + fit.halfwidth[idx]);
var lowerLimit = fit.fitted.map((yhat, idx) => yhat - fit.halfwidth[idx]);
// plot upperLimit and lowerLimit

Alternatively, use .grid( ) method to generate a grid of equally spaced points:

newData = model.grid([20, 20]);

fit = model.predict(newData);

Usage

Find out more by visiting my demo app:

https://loess.netlify.com/

Documentation

class Loess {
  constructor (data: object, options: object) {
    // arguments
    data /*required*/ = {
      y: [number],
      x: [number],
      x2: [number], // optional
      w: [number]  // optional
    }

    options /*optional*/ = {
      span: number, // 0 to inf, default 0.75
      band: number, // 0 to 1, default 0
      degree: [0, 1, 2] || ['constant', 'linear', 'quadratic'] // default 2
      normalize: boolean, // default true if degree > 1, false otherwise
      robust: boolean, // default false
      iterations: integer //default 4 if robust = true, 1 otherwise
    }

    // return a LOESS model object with the following properties
    this.y = data.y
    this.x = [data.x, data.x2] // predictor matrix
    this.n = this.y.length // number of data points
    this.d = this.x.length // dimension of predictors
    this.bandwidth = options.span * this.n // number of data points used in local regression
    this.options = options
  }

  predict (data: object) {
    // arguments
    data /*optional*/ = {
      x: [number],
      x2: [number]
    } // default this.x

    return {
      fitted: [number], // fitted values for the specified data points
      halfwidth: [number] // fitted +- halfwidth is the uncertainty band
    }
  }

  grid (cuts: [integer]) {
    return {
      x_cut: [number], // equally-spaced data points
      x_cut2: [number],
      x: [number], // all combination of x_cut and x_cut2, forming a grid
      x2: [number]
    }
  }
}

Note:

• data should be passed into the constructor function as json with keys y, x and optionally x2 and w. Values being the arrays of response, predictor variables, and observation weights.
• If no data is supplied to .predict( ) method, default is to perform fitting on the original dataset the model is constructed with.
• span refers to the percentage number of neighboring points used in local regression.
• band specifies how wide the uncertainty band should be. The higher the value, the greater number of points encompassed by the uncertainty band. Setting to 0 will return only fitted values.
• By default LOESS model will perform local fitting using the quadratic function. Overwrite this by setting the degree option to "linear" or "constant". Lower degree fitting function computes faster.
• For multivariate data, normalize option defaults to true. This means normalization is applied before performing proximity calculation. Data is transformed by dividing the factors by their 10% trimmed sample standard deviation. Turn off this option if dealing with geographical data.
• Set robust option to true to turn on iterative robust fitting procedure. Applicable for estimates that have non-Gaussian errors. More iterations requires longer computation time.
• When using .grid( ), cuts refers to the number of equally spaced points required along each axis.
• dist is intentionally checked in to avoid the need to build when installing directly from GitHub.

Credits

William S. Cleveland, Susan J. Devlin
Locally Weighted Regression: An Approach to Regression Analysis by Local Fitting
Journal of the American Statistical Association, Vol. 83, No. 403. (Sep., 1988), pp. 596-610.

William S. Cleveland, Eric Grosse, Ming-Jen Shyu
A Package of C and Fortran Routines for Fitting Local Regression Models (20 August 1992)
Source code available at http://www.netlib.org/a/dloess