[Bug]: Incorrect cosine similarity returned in the search results

[jpowie01]

[Bug]: Incorrect cosine similarity returned in the search results

[yanliang567]

/assign @liliu-z
please take a look
/unassign

[xiaofan-luan]

don't quite understand it.
What is the current problem?
The distance result seems to be correct

[yanliang567]

I think searching an exact vector for himself should return 0 as distance, but it returns -1 if cosine.

[xiaofan-luan]

I think searching an exact vector for himself should return 0 as distance, but it returns -1 if cosine.

exact vector should be 1. 0 means not related at all

[xiaofan-luan]

it's opposite direction, then the distance is -1.

[jpowie01]

Yes, that's exactly the problem. The values I get back from the search query are incorrect.

If you look closer to the example I provided above, I've added three vectors to the database ([1, 1, 1, 1], [1, 1, -1, -1], [-1, -1, -1, -1]) and then searched for a vector [-1, -1, -1, -1]. As a result, I got distance values as following:

• [1, 1, 1, 1] -> 1.0
• [1, 1, -1, -1] -> 0.0
• [-1, -1, -1, -1] -> -1.0 (exact search)

Those are incorrect values, neither for cosine similarity nor cosine distance. From the behaviour I'm seeing, those values are representing cosine similarity multiplied by -1. But... why?

Here is a snippet of code using scipy & scikit-learn computing those metrics on the same vectors:

>>> from scipy.spatial import distance
>>> distance.cosine([-1, -1, -1, -1], [1, 1, 1, 1])
2.0
>>> distance.cosine([-1, -1, -1, -1], [1, 1, -1, -1])
1.0
>>> distance.cosine([-1, -1, -1, -1], [-1, -1, -1, -1])
0.0

>>> from sklearn.metrics.pairwise import cosine_similarity
>>> cosine_similarity([[-1, -1, -1, -1]], [[1, 1, 1, 1]])
array([[-1.]])
>>> cosine_similarity([[-1, -1, -1, -1]], [[1, 1, -1, -1]])
array([[0.]])
>>> cosine_similarity([[-1, -1, -1, -1]], [[-1, -1, -1, -1]])
array([[1.]])

[xiaofan-luan]

/assign @liliu-z