nevali: "All gone. It was a mistake for several reasons. PushState or bust."
danbri: "While the Laplacian spectrum has many attractive theoretical properties which make it very useful for partitioning large sparse arrays, it has at least one drawback: the Fiedler eigenvector is associated with the second smallest eigenvalue."
danbri: "The current method of choice for finding a few eigenvectors of very large sparse matrices is the Lanczos method, which tends not to find small eigenvectors as quickly as it finds the largest eigenvectors."
danbri: Compare Mahout's Lanczos documentation, " Which k? It's actually a spread across the entire spectrum: the first few will most certainly be the largest singular values, and the bottom few will be the smallest."
danbri: ...which means -if I understand correctly- a long wait for 2nd smallest, if dataset is big.
danbri: "The current method of choice for finding a few eigenvectors of very large sparse matrices is the Lanczos method, which tends not to find small eigenvectors as quickly as it finds the largest eigenvectors."
danbri: Compare Mahout's Lanczos documentation, " Which k? It's actually a spread across the entire spectrum: the first few will most certainly be the largest singular values, and the bottom few will be the smallest."
danbri: ...which means -if I understand correctly- a long wait for 2nd smallest, if dataset is big.
danbri: "We present some recent interesting material from mathematical physics and graph theory that has not appeared in social network literature" (1997)