UMAP segmentation fault

umap.UMAP(n_neighbors=11, metric='jaccard', 
                      min_dist=0.5).fit(X)

Error - segmentation fault

umap-learn 0.5.1 - Bad umap-learn 0.5.0 - Bad umap-learn 0.4.6 - Good

But if you increase it n_neighbors to 500, then everything works on new versions.

[lmcinnes]

So this was actually quite complicated because it is a subtle issue. In the end it really comes down to the nature of your data and the construction of random projection trees to initialize a nearest neighbour search combining with floating point error accumulation that ends up causing bad things to happen (tree splits not occurring properly). A long term fix for that is a little harder, but there is a very simple short term fix:

Since you are using Jaccard distance, which only cares about whether entries are non-zero or not, you can simply pass in a binarized version of your data and everything should work.

umap.UMAP(n_neighbors=11, metric='jaccard', 
                      min_dist=0.5).fit(X != 0)

[lmcinnes]

Any angular metric will have similar issues. Your data is very skewed. Try using StandardScaler or similar on your data first to at least get the different columns in the same value ranges.