2
$\begingroup$

I have a given data set $D = \{ x_i, y_i \}_{i=1}^n$ for a regression problem. When I plot the data, it looks like there is an underlying parabola (2nd order linear model) and some outliers.

I want to design an approach using a probabilistic model with a latent binary variable $\{ 0,1 \}$ indicating whether a data point is an outlier or not.

Currently I have no idea what I could do, what would the parameters be in this cause and how are they optimized? Is Expectation Maximization an idea?

$\endgroup$
1
  • $\begingroup$ Are there known outliers in your data? Why isn't using the $\frac{3}{2} IQR$ cutoff sufficient? $\endgroup$
    – user17794
    Jun 16, 2012 at 22:41

1 Answer 1

1
$\begingroup$

My recommendation is to use robust regression. It is simpler and downweights the outliers.

$\endgroup$
4
  • $\begingroup$ But this doesn't involve a probabilistic model, does it? $\endgroup$
    – Mahoni
    Jun 16, 2012 at 22:02
  • $\begingroup$ Why do you want a probabilistic model for outliers? $\endgroup$ Jun 16, 2012 at 22:23
  • $\begingroup$ @MichaelChernick : Your suggestion makes sense only if outliers are bad. What if the whole purpose is to find outliers? The outliers may be the places where you should drill for oil. $\endgroup$ Jun 16, 2012 at 22:38
  • $\begingroup$ @MichaelHardy Mahoni said that the model fit a quadratic function except for a few outliers. That indicates that the outliers were hurting the fit. But if you want to detect the outliers the larger residuals from a robust regression fit will be the outliers. $\endgroup$ Jun 16, 2012 at 22:50

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .