I measured the score_samples performance for Isolation Forest (which is in essence the base for everything else) and found the following:
Here n_samples is the shape for score_samples argument. As you can see, there is very significant multi threading overhead for data sets smaller than 2048 instances. Please note that our default n_jobs is -1. The overhead is responsible for slow AAD optimizations since AAD calls score_samples for known data subset (very little, <100). The measurements were carried out using Python profiling.
Hello,
I measured the
score_samplesperformance for Isolation Forest (which is in essence the base for everything else) and found the following:Here
n_samplesis the shape forscore_samplesargument. As you can see, there is very significant multi threading overhead for data sets smaller than 2048 instances. Please note that our defaultn_jobsis-1. The overhead is responsible for slow AAD optimizations since AAD callsscore_samplesfor known data subset (very little, <100). The measurements were carried out using Python profiling.