discriminant analysis

User13

I noticed a couple of things in running the linear discriminant analysis function:

1.The predictions on classification performance I get with it are not the same as other tools R MASS and Tanagra ( which do match each other). As a result the error rates and accuracy rates dont compare either

2.When i choose the LDA operator instead of QDA or RDA- when it actually runs, the output says quadratic on the title of the page- not linear. I can only get it to say linear on the title if I run an alpha of 1 and use RDA.



Since most of what is going on behind the scenes is invisible to the user, the fact that results on the same dataset dont match other tools running the same data, i am wondering if there is a bug here, or that I am just not understanding what Rapidminer does behind the scenes- so that I might not be exactly matching the exact test I do in the other tools.



What causes the differences between RapidMiner and the other tools above?



I do like what RapidMiner does and how it does it- but am concerned when I cant get same results with what I think are exactly the same statistical tests.



Thanks

User13

Hi,

it seems to me, that the two strings of the QDA and the LDA model have been confused with each other, but apart from the wrong title, it shouldn't have any further implications.

The slight differences between the models should be minimal. The implementation is pure LDA from the books, but there might be differences in the way how numerical precision effects are handled by the math library in the background. I wouldn't take small deviations in the error rates too serious, as long as there is no significant trend of our algorithm to behave worse than the other implementations. Did you test that on just one dataset? Can you share dataset and results and R code so that we can reproduce?

With kind regards,
Sebastian Land

User13

User13