multivariate outlier detection

Question

Hello,

I have the following question from a customer (Freescale) about the JMP12 outliers detection and about multivariate robust outliers:

is there a way to know which are the variables who have the biggest impact on Mahalanobis distance (as a discriminante with stepwise)?

Is this way equivalent to do a PCA followed by a mahalanobis distance ? -> in my sense, it is a little bit different due to robust analysis.

KNN multivariate outlier: is it equivalent to do a cluster analysis with a k-mean eval to 1 (more option in clustering platform)?

Thank for your reply,

Florence

malcolm_moore1 · Answer

Hi Florence,

It looks like Freescale may already be doing this or something similar, one way of taming this high dimensional multivariate outlier detection problem might be:

1) Use Mahalanobis plot to detect outlier events, then

2) Look at individuals SPC chart of each PCA Cluster to determine which cluster(s) are outliers for each outlier event in (1), then

3) Look at individuals SPC chart of variables within outlier cluster(s) for each outlier event in (2) to identify outlier variables

If this procedure works for them, they may want to script the workflow to make it easier for a wider set of users to apply.

Malcolm

flo_kussener · Answer

Hey Peter ! Thank a lot for your suggestion. The main issue with the scatterplot matrix is the fact that she is in semiconductor area so she has more than 300 variables. Not so easy to use graphical tools… Regarding PCA, she has JMP Pro, so I think that she using the clustering option… Florence De : peter.bartell@jmp.com Envoyé : lundi 8 juin 2015 17:26 À : Florence Kussener Objet : Re: - multivariate outlier detection JMP User Community multivariate outlier detection reply from peter.bartell@jmp.com in JMP SE Community - View the full discussion

Peter_Bartell · Answer

Florence: Here's a couple thoughts for your customer...wrt to the Mahalanobis plot, if there are outliers on that plot, a simple click on the outlier point will highlight that row in the data table, and a dynamic view of the scatterplot matrix too should also reveal one or more variables that are outside the bi-variate ellipses. This should give some hints as to which variables are the culprit(s).

My next thought is to your second question...a simple PCA report will show the bi-plot and if there are multivariate outliers, chances are good they will show up as outlying points on the bi-plot since the bi-plot is basically a residual plot for variability left over after the first two principle components are calculated.