Debugging build_distance_objects()

[paddytobias]

Dealing with cases were there is no similarity for a document in the dfm and therefore NA is returned by quanteda::textstat_simil(). Fixing this by setting all NAs to 0 as a default. This would be a common problem for sparse DFMs

[kbenoit]

A distance of 0 means the objects are equivalent, such as dist(A, A) = 0. For sparse symmetric matrix objects (of which the Matrix package defines several, the margin can be treated in a special way, at least for distance measure for which this basic axiom of distance is true. (In quanteda we removed all those measures for which it was not true.)

A distance of NA means the distance is undefined, for instance dist(docA, docB) where docB is empty (contains zero counts for all features).

There should imo be a difference. Note that proxy and stats::dist() treat these differently. We went with the stats::dist() approach. See See quanteda/quanteda#1540 where we discussed this at some length.