You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This is a revive of the draft PR #86 (please look into it for further information) because I think it is worth to look into this more given that CUR outperforms FPS by far in regression quality and is often not used because it is so expensive to compute.
The core idea is to update the eigenvectors after a selection instead of recomputing them by an eigendecomposition. @ceriottm mentioned in a discussion that it was mathematically unstable for eigenvectors corresponding to degenerated eigenvalues. So this deserves some dedicated time look into this in detail.
the issue here is that for a big matrix we don't want to compute ALL eigenvectors, but only the top ones. this can be done efficiently, but the problem is then how to do a rank-one update of only some of the eigenvectors. this is the only article I had found for this, but couldn't find an implementation https://epubs.siam.org/doi/epdf/10.1137/18M1172120
This is a revive of the draft PR #86 (please look into it for further information) because I think it is worth to look into this more given that CUR outperforms FPS by far in regression quality and is often not used because it is so expensive to compute.
The core idea is to update the eigenvectors after a selection instead of recomputing them by an eigendecomposition. @ceriottm mentioned in a discussion that it was mathematically unstable for eigenvectors corresponding to degenerated eigenvalues. So this deserves some dedicated time look into this in detail.
Links:
The text was updated successfully, but these errors were encountered: