Introduced on day 38
Hubness
In high-dimensional nearest-neighbour search, a few points appear in a disproportionate number of other points' neighbour lists while many appear in none.
In detail
Hubs crowd out relevant results and can be reduced by centering, normalisation or neighbour re-scaling. It is a property of the geometry and the data distribution, so you check it by counting how often each item is retrieved, not by inspecting individual queries.