About distances about annoy

[erikbern]

cosine_distance = sqrt(2-2*cos_similarity)

so cosine_similarity = 1 - cosine_distance^2/2

[erikbern]

Cosine distance can be between 0 and 2. The distance between x and -x should be exactly 2

http://docs.scipy.org/doc/scipy-0.16.0/reference/generated/scipy.spatial.distance.cosine.html

I will add a unit tests though

[anhldbk]

@erikbern Thank you Erik

[harsh157]

Shouldn't the distance range be 0 to 4? as it is computing 2*(1-cos). That is if we have vectors in opposing directions also.

[MrChrisJohnson]

Why 2_(1-cos)? Distance should just be 1-cos which is equivalent to 1 - u_v / ||u|| * ||v||. So, distance would range from 0 (angle of 0) to 2 (angle of 180). See the link that @erikbern links to earlier.

[Irtza]

@erikbern @MrChrisJohnson Can we convert to cosine similarity using cos_sim = 1-cos_distance ?