Increase the max_iter for LabelPropagation. #9441
Conversation
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
|
Does this correspond, approximately, to 30 in LabelSpreading, or is it much more generous? |
|
Btw, I think we should change this for the 0.19 release, seeing as we're basically overturning |
|
@musically-ut: If you can briefly illustrate the convergence |
|
I have tried with the scaled digits dataset with 90% unlabeled data and the RBF kernel and for good cross-validated values of gamma (e.g. gamma between 1 and 20), the effective number of iterations can vary between 2 and 3000 (I had to increase tol to 0.01). So indeed |
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from #5893.
|
Backported to the 0.19.X branch as c51aee8. |
|
Oh, cool. Thanks @ogrisel! I too was composing a benchmark with the datasets; I noticed that we may have to change the example of digit learning in light of the changed underlying algorithm. |
|
The digit learning example seems to be fine on the dev branch: |
|
Yes, the example works with However, is it only me or is the table much harder to read on the |
|
This is a matplotlib issue, we should use |
|
@ogrisel I think I prefer backporting just before the release, because otherwise it's hard to keep track of what to backport. |
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
LabelPropagation converges much slower than LabelSpreading. The default of max_iter=30 works well for LabelSpreading but not for LabelPropagation. This was extracted from scikit-learn#5893.
In practice,
LabelPropagationconverges much slower thanLabelSpreading. The defaultof
max_iter=30works well forLabelSpreadingbut not forLabelPropagation. This PR changesmax_iterforLabelPropagationto 1000.This was extracted from #5893.