This is with reference to the PR #6727 and is discussed in the comments in #5774.

Description

The current implementation of LabelPropagation, in semi_supervised module, creates directed k-NN graphs when the knn kernel is chosen.

One way to fix it is as shown in musically-ut/semi_supervised.

Steps to Reproduce / Expected Results / Actual Results

This is part of the reason I didn't submit it as a separate issue: it is difficult to come up with test cases which are obviously correct or incorrect.

However, none of the papers about semi-supervised learning cited in #6727 considers directed graphs. Using directed graphs has some obvious problems, like rendering parts of the graph completely unreachable, thus limiting the propagation of labels only from roots to descendants.