Wednesday, July 18, 2012

Co-variance and Gram Matrix

You can think of your data as a matrix, where the rows represent the instances and the columns represent the features. Let's call this matrix $A$. Now $A^TA$ represents feature $\times$ feature, or as it is commonly called: the co-variance matrix. On the other hand,  $AA^T$ represents instance $\times$ instance, i.e. the similarity! This matrix is called the gram matrix.  

This is useful in computing SVD. SVD is a method to decompose a matrix $A$ into 3 components:

$A=U\Sigma V^T$

U is obtained from the eigenvectors of the gram matrix ( $AA^T$), while V is obtained from the eigenvectors of the co-variance matrix ( $A^TA$). 

2 comments:

  1. hello! I just discover this website and it's wonderful...everything is well explained, great job. Is there gonna be more?

    ReplyDelete
  2. I will try to update it more frequently! Thanks for the comment.

    ReplyDelete