Why don't similar-looking reads cluster together in IsoSeq?

Frequently an IsoSeq run produces a large number (hundreds) of clusters, far in excess of the expected number of unique isoforms present. This population includes many clusters which are very similar or even identical, which one would expect to have clustered together.

OTOH, combining slightly different reads could cause small differences to be ignored, when they may in fact be real.

How do we decide where to draw the line? Can we control the degree of similarity required for clustering?