search

lauerfab / MLweb

Machine learning and scientific computing (linear algebra, statistics, optimization) javascript libraries, with an online lab.
http://mlweb.loria.fr
GNU General Public License v2.0
85 stars 22 forks source link

LDA Options #4

Closed DelfsEngineering closed 7 years ago

DelfsEngineering commented 7 years ago

We would like to use LDA with Mahalanobis Distance and Leave One Out cross validation. Can we use your code base as it is for this type of LDA?

If not, could this be added?

Cheers

lauerfab commented 7 years ago

It depends what you mean by LDA with Mahalanobis distance. As you can see from the description of the LDA method (by clicking on the link "more information on this topic" from the help of LDA in the online lab at http://mlweb.loria.fr/lalolab/ ), LDA is based on probability density functions of normal distributions, which naturally involve the squared Mahalanobis distance. So plain LDA might be what you are looking for.

Regarding Leave-one-out (LOO), this is implemented by the cross-validation function applicable to any Classifier/Regression model. By default, this performs a 5-fold cross-validation, but you can specify the number of folds in a 3rd argument. For training data in the matrix X and vector Y, this could be done with

model = new Classifier(LDA) RecRate = model.cv(X, Y, Y.length)

To make this easier, I just added a wrapper for that, so now you can just call

RecRate = model.loo(X, Y)

I hope this answers your questions.