[WIP] Make knn kernel undirected. #9439
Conversation
|
do we have reference to guide this change? (sorry if they're listed in the
issue)
…On 24 Jul 2017 3:28 am, "Utkarsh Upadhyay" ***@***.***> wrote:
This is a very simple way of making the knn kernel matrix symmetric.
However, this slightly changes the meaning of n_neighbors: the *actual*
number of neighbors may vary from n_neighbors to 2 * n_neighbors in the
worst case.
- Update tests after #5893
<#5893> is merged.
Fixes #8008 <#8008>.
------------------------------
You can view, comment on, or merge this pull request online at:
#9439
Commit Summary
- Make knn kernel undirected.
File Changes
- *M* sklearn/semi_supervised/label_propagation.py
<https://github.com/scikit-learn/scikit-learn/pull/9439/files#diff-0>
(13)
- *M* sklearn/semi_supervised/tests/test_label_propagation.py
<https://github.com/scikit-learn/scikit-learn/pull/9439/files#diff-1>
(40)
Patch Links:
- https://github.com/scikit-learn/scikit-learn/pull/9439.patch
- https://github.com/scikit-learn/scikit-learn/pull/9439.diff
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#9439>, or mute the
thread
<https://github.com/notifications/unsubscribe-auth/AAEz64rv6gXp9bvPcQNKetvMyKrQW-Uhks5sQ4K5gaJpZM4Ogg-1>
.
|
|
There is some discussion of directed graphs in Section 6.6 of Semi-Supervised Learning Literature Survey. This paper explicitly deals with the problem, but using a different On a cursory look, I couldn't find any explicit discussion about using kNN to build the underlying graphs in the two references mentioned for the present code; they only seem to use k-NN as baselines. |
|
Would it be sensible to name this kernel by a different string and leave the theoretically dubious one there? |
Also, fix/update the tests. Fixes scikit-learn#8008.
3d4c1f3 to
d99d853
Compare
|
Interesting option. I see that it will preserve backwards compatibility but I am not sure whether the current If we are planning on leaving the old
|
|
Hence I would close it because:
|
This is a very simple way of making the
knnkernel matrix symmetric. However, this slightly changes the meaning ofn_neighbors: the actual number of neighbors may vary fromn_neighborsto2 * n_neighborsin the worst case.Fixes #8008.