clarification of distance metric

[matanox]

Hi,

This is a coolest project, really awesome :-) Would you care to kindly comment on the distance metric/s implemented and the rationale thereof?

In many machine learning scenarios we pick e.g. cosine similarity on normalized vectors (so that we're working in a multi-dimensional sphere). Is this here very different in that you look at the plain vector distance?

Thanks in advance for your commenting!!

[anvaka]

I'm using spotify/annoy library for the index. By default its distance metric is set to 'angular':

Annoy uses Euclidean distance of normalized vectors for its angular distance, which for two vectors u,v is equal to sqrt(2(1-cos(u,v)))

And that is exactly what I'm using here. So, this is regular euclidean distance, implemented by law of cosines on normalized vectors

[matanox]

hey thanks. great work.