For my current project I favour the QR approach because I need the covariance matrix (for the confidence bounds on the prediction). For dense on heap matrices the current QR code is fast enough, but for distributed matrices it's very slow. I think that the the current implementation (with the two for loops) is causing too much unnecessary blocking.
Kind regards Pieters PS: For some ill conditioned problems (eg. polynomial regression), the additional numerical stability of Householder compared to MGS is favourable. -- Sent from: http://apache-ignite-users.70518.x6.nabble.com/
