Truncated SVD

• Keep only k largest eigenvalues

  • 0 out others

• Discard eigenvectors with (new) 0 eigenvalue

  • keep k columns of U and V

• Gives new rank-k matrix

• Closest rank-k approximation to A

• Terms, documents combinations of eigenvecs

  • “optimal” projection of terms, documents into a k-dimensional subspace

Previous slide

Next slide

Back to first slide

View graphic version