Cosine Similarity: Explained and Examples

  • Thread starter Thread starter daveronan
  • Start date Start date
  • Tags Tags
    Cosine
Click For Summary
Cosine similarity measures the angle between two vectors, with a result of 1 indicating that the vectors are equal, 0 indicating orthogonality, and -1 indicating they are oppositely directed. When comparing vectors, particularly in text search, the focus is typically on non-negative weights, making negative similarity values irrelevant. However, in domains like audio feature comparison, vectors can have negative values, allowing for a broader range of similarity assessments. Scalar multiples of vectors indicate they lie in the same one-dimensional subspace, while differing signs denote dissimilarity. Understanding cosine similarity is essential for accurately interpreting relationships between vectors in various applications.
daveronan
Messages
13
Reaction score
0
Just for clarification...

If I take the cosine similarity of two vectors and i get an answer of 1, then bother vectors are equal and the same.

If I do the same again with another two vectors and get an answer of 0, then the vectors are at an angle of 90 degrees to each other and dissimilar.

If I do the same again with another two vectors and get an answer of -1, then the vectors are at an angle of 180 degrees to each other and are even more dissimilar than they were at 90 degree or should I always take the absolute value of the answer. The equation I'm referring to can be found here.

http://upload.wikimedia.org/math/f/3/6/f369863aa2814d6e283f859986a1574d.png

Thanks!
 
Physics news on Phys.org
If the cosine of two vectors is 1 then they are scalar multiples of each other. If it is -1 then it is same except the scalar is negative. If the cosine is 0 they are orthogonal. Don't take the absolute value. The greater the number the greater the similarity.

As far as similarity goes, I'm guessing you are looking at the vector space model for text search since they represents documents as term weight vectors and similarity between documents is taken as the cosine between their corresponding vectors. This similarity is a separate notion that leverages the cosine and is not inherent in the cosine of two vectors. With that said, term weight vectors consist of non-negative weights and so all vectors will be within 90 degrees or less with each other and negative similarity is not possible so you only consider 0 to 1.
 
Hi TheOldHag,

First of, thanks for getting back to me.

I'm comparing audio features (MFCCs), so it's possible for me to get vectors with negative values.

Thanks for your answer. You can't beat clarification! :)
 
Vectors that are scalar multiples of each other are in the same one dimensional subspace of each other. I think similarity is domain specific so not sure how to answer the question. They are similar because they are scalar multiples and they are different because they are oppositely signed.

FYI, I'm answering questions because I'm in the process of learning myself but there are limits to what I know.
 
I am studying the mathematical formalism behind non-commutative geometry approach to quantum gravity. I was reading about Hopf algebras and their Drinfeld twist with a specific example of the Moyal-Weyl twist defined as F=exp(-iλ/2θ^(μν)∂_μ⊗∂_ν) where λ is a constant parametar and θ antisymmetric constant tensor. {∂_μ} is the basis of the tangent vector space over the underlying spacetime Now, from my understanding the enveloping algebra which appears in the definition of the Hopf algebra...

Similar threads

  • · Replies 26 ·
Replies
26
Views
3K
  • · Replies 33 ·
2
Replies
33
Views
4K
Replies
5
Views
2K
  • · Replies 3 ·
Replies
3
Views
2K
  • · Replies 9 ·
Replies
9
Views
2K
  • · Replies 9 ·
Replies
9
Views
3K
  • · Replies 19 ·
Replies
19
Views
4K
Replies
14
Views
2K
  • · Replies 25 ·
Replies
25
Views
3K
  • · Replies 3 ·
Replies
3
Views
834