forum.alglib.net

I currently use PCA on a large data set where different rows represent data of a different scale than other rows, and I am concerned that when I try to approximate the input data using a few of the basis vectors I am getting data that is skewed toward the larger input scale. What is the recommended approach to reducing the impact of "outlier rows" of data such that all data rows are given similar importance in basis determination?

You can manually re-scale these variables, if you know their scale, so they will have unit scale. After performing PCA you must rescale principal vectors. So, as result, you will get PCA vectors which are orthogonal with respect to inner product given by scale matrix (however, they will be non-orthogonal in the usual Euclidean sense).