Robust regression

Ordinary least squares can move sharply when targets contain outliers. HuberRegressor softens large residuals, RANSACRegressor fits a consensus set, TheilSenRegressor aggregates many subsample fits, and QuantileRegressor estimates a conditional quantile instead of a mean.

JavaScript implementation

@kanaries/ml exposes sklearn-style constructors and fit/predict methods that run in browser or Node.js. Use Huber for moderate contamination, RANSAC for clear gross outliers, Theil-Sen for small robust datasets, and quantile regression for medians or prediction bands.

Interactive robust regression playground

Switch between Huber, RANSAC, Theil–Sen, and quantile regression, then try the outlier dataset or add extreme points yourself. Each curve is trained live by the selected @kanaries/ml estimator.

Quick start

import { Linear } from '@kanaries/ml';

const model = new Linear.RANSACRegressor({ randomState: 42 });
model.fit([[0], [1], [2], [3]], [0, 1, 20, 3]);
console.log(model.predict([[4]]));

Detailed API reference

• HuberRegressor({ epsilon=1.35, alpha=0.0001, maxIter=100, tol=1e-5, fitIntercept=true }) exposes coef, intercept, scale, and outliers.
• RANSACRegressor({ estimator, minSamples, residualThreshold, maxTrials=100, stopProbability=.99, randomState }) exposes estimatorFitted, inlierMask, and nTrials.
• TheilSenRegressor({ nSubsamples, maxSubpopulation=10000, maxIter=300, tol=1e-3, randomState }) exposes coef, intercept, and nIter.
• QuantileRegressor({ quantile=.5, alpha=1, fitIntercept=true, maxIter=5000, tol=1e-7 }) solves the pinball-loss problem with an ADMM optimizer.