PDF version of this entire document

... 4[*]
It may be safer to attempt this with 5 NRR algorithms, then choosing just 4 that are most distinguishable, potentially by hand-tweaking various free parameters. Without significant difference in quality, it would be hard to back the results using ground-truth label overlap.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.