Good and Bad Neighborhood Approximations for Outlier Detection Ensembles

Outlier detection methods have used approximate neighborhoods in filter-refinement approaches. Outlier detection ensembles have used artificially obfuscated neighborhoods to achieve diverse ensemble members. Here we argue that outlier detection models could be based on approximate neighborhoods in the first place, thus gaining in both efficiency and effectiveness. It depends, however, on the type of approximation, as only some seem beneficial for the task of outlier detection, while no (large) benefit can be seen for others. In particular, we argue that space-filling curves are beneficial approximations, as they have a stronger tendency to underestimate the density in sparse regions than in dense regions. In comparison, LSH and NN-Descent do not have such a tendency and do not seem to be beneficial for the construction of outlier detection ensembles.