I am not very familiar with LDA but was interested in using it to find multiple, orthogonal discriminant dimensions for data with 2 class labels. I understand that standard that the point of LDA is to find a single axis that discriminates 2 classes so it will necessarily find a single discriminant dimension but I was wondering whether, having found this axis, I would be able to factor it out of the data

i.e. say $\beta$ $\in \mathbb{R}^{m} $are the LDA coefficients that project the data $D \in \mathbb{R}^{n x m}$ where $n$ is the number of samples and $m$ is the number of features of each sample, onto the discriminant axis. Then calculate

$$D2 = D - \left( \frac{\beta\cdot(\beta^T\cdot D^T)}{(\beta^T\cdot\beta)} \right) ^T $$

and rerun LDA on $D2$ again to find a second orthogonal discriminant dimension. I am interested in using this approach mainly for data visualisation but am not sure whether it make sense? If so is there a name for this method?


1 Answer 1


Finding the discriminant directions is equivalent to finding the eigen values of the matrix $S_w^{-1} S_b$, where $S_w$ is the within scatter matrix and $S_b$ the between scatter matrix. For $C$ classes, this matrix has rank $C-1$, which means that you can only find $C-1$ discriminant directions.

In your case of only two classes, you cannot find more than one direction.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.