Re: Threads: Mahalanobis distance and Visualising PCA

From: Ross Clement (clemenr_at_wmin.ac.uk)
Date: 10/25/04 

Date: 25 Oct 2004 11:52:26 -0700

Paige Miller <paige.miller@kodak.com> wrote in message news:<cliqq5$hgl$1@news.kodak.com>...

> > PS: As for Mahalanobis distance, I'm now convinced that this is
> > inappropriate for my current data. I followed the advice given in a
> > previous thread to drop one of the dimensions of my [compositional]
> > data. This gave me a covariance matrix that I could easily convert,
> > but tests on both toy and real data show that it performs worse than
> > classifiers based on plain Euclidean distance. And, I tested this by
> > including a run-time option in my program that over-wrote the
> > covariance matrix with an identity matrix. Performance improves.
>
> You know, I NEVER SAID that you use Mahalanobis for classification
> or discrimination. Your original question was about PCA, and
> Mahalanobis is much better for PCA. It is not necessarily better for
> discrimination.

The PS: here was not meant to be on the same topic as the previous bit
where I was discussing what you had written. I.e. I wasn't attributing
recommendations to use Mahalanobis distance to you. It was some people
here at wmin who had said that using Mahalanobis distance in
classification in preference to Euclidean distance was a no-brainer.

Cheers,

Ross=c

Relevant Pages