outliers because we square errors, a few points with large errors can have large effect on fitted...

4
Outliers • Because we square errors, a few points with large errors can have large effect on fitted response surface • Both in simulations and in experiments there is potential for preposterous results due to failures of algorithms or tests • Points with large deviations from fit are called outliers. • Key question is how to distinguish between outliers that should be removed and ones that should be kept.

Upload: cheyenne-kemery

Post on 31-Mar-2015

219 views

Category:

Documents


2 download

TRANSCRIPT

  • Slide 1

Slide 2 Outliers Because we square errors, a few points with large errors can have large effect on fitted response surface Both in simulations and in experiments there is potential for preposterous results due to failures of algorithms or tests Points with large deviations from fit are called outliers. Key question is how to distinguish between outliers that should be removed and ones that should be kept. Slide 3 Weighted least squares Weighted least squares was developed to allow us to assign weights to data based on confidence or relevance. Most popular use of weighted least squares is for moving least squares, where we refit data for each prediction point with high weights for nearby data. Linear interpolation from a table is an extreme form. Error measure Normal equations Slide 4 Determination of weights Slide 5 Example x = (1:10)'; y = 10 - 2*x + randn(10,1); y(10) = 0; bls = regress(y,[ones(10,1) x]) brob = robustfit(x,y) bls = 7.8518 -1.3644 brob = 8.4504 -1.5278 >> scatter(x,y,'filled'); grid on; hold on plot(x,bls(1)+bls(2)*x,'r','LineWidth',2); plot(x,brob(1)+brob(2)*x,'g','LineWidth',2) legend('Data','Ordinary Least Squares','Robust Regression')