ChapterZero

K-means as a matrix optimization problem

I came across this formulation of k-means while trying to figure out what variants of nonnegative matrix factorization are used in practice:

The basic kernel k-means problem, given datapoints $\{x_i\}_{i=1}^n$ , is to find clusters $C_1, \ldots, C_k$ which minimize the sum of the distances of the points in the clusters from the cluster centers:
$\displaystyle \min_{C_1, \ldots, C_k} \sum_{i=1}^r \sum_{x_i\in C_r} \left\|\phi(x_i)-\frac{1}{|C_r|}\sum_{x_j \in C_r} \phi(x_j)\right\|^2$

Here $\phi$ maps the input points into some feature space; the kernel $k(x, y) = \langle \phi(x), \phi(y) \rangle$ measures the similarity of the features of its inputs. Using the kernel, we can rewrite the above as
$\displaystyle \min_{C_1, \ldots, C_k} \sum_{i=1}^r \left( \sum_{x_i \in C_r} k(x_i, x_i)-\frac{1}{|C_r|} \sum_{x_i \in C_r} \sum_{x_j \in C_r} k(x_i, x_j) \right),$
and since the first term is a constant, this is equivalent to
$\displaystyle \max_{C_1, \ldots, C_k} \sum_{i=1}^r \frac{1}{|C_r|} \sum_{x_i, x_j \in C_r} k(x_i, x_j).$

Now let $F_{ij} = 1/|C_r|$ if $i,j \in C_r$ and 0 otherwise, and let $K_{ij} = k(x_i, x_j)$ and write the above as
$\displaystyle \max_{\text{appropriate } F} \sum_{ij} F_{ij} K_{ij} = \max_{\text{appropriate } F} \text{Tr}(FK).$
The question is, what is the appropriate set to maximize over?

The paper shows that F = GG^t where is a rank matrix with orthonormal columns and entries all nonnegative and F1 = 1 . This is easy to see: by reordering the points x_i in a manner corresponding to the optimal cluster choice, so that each cluster is represented by a contiguous subsequence of the ordering, we find an equation relating the corresponding $\hat{F}$ to :
$\displaystyle \sum_{ij} F_{ij} K_{ij} = \sum_{ij} \hat{F}_{ij}K_{\sigma(i), \sigma(j)} \Leftrightarrow F = P_\sigma \hat{F} P_\sigma^t$
where $P_\sigma$ is the permuation matrix that maps the th standard basis vector to the $\sigma(i)$ th.

By construction $\hat{F}$ is block diagonal with blocks each of the form $\hat{F}_i = \lambda_i 11^t$ such that $\hat{F}_i 1 = 1$ . Now take to be the matrix whose th column is the indicator vector for C_r scaled by $1/\sqrt{|C_r|}$ . It’s clear that $\hat{F} = GG^t$ satisfies the above characterization, and conjugation with a permutation matrix doesn’t change this, so satisfies the above characterization.

What I don’t see is how they justify the statement that this is the set to optimize over. Sure, each clustering is represented by some in this set, but how do we know that the set of clustering matrices isn’t a proper subset of the set? Unless I’m missing something in their argument, it’s incomplete until we know that if F=GG^t where is a rank matrix with orthonormal columns and all nonnegative entries and F1 = 1 , then $F = P\hat{F}P^t$ for some permutation matrix and block diagonal $\hat{F}$ that has blocks, each of which satisfies $\hat{F}_i = \lambda_i 11^t$ and $\hat{F}_i1 = 1$ . (Well this, or that if you have a matrix in this set which doesn’t represent a clustering, there’s always a matrix in the set which represents a clustering and has a higher value of the objective)

Update It turns out that these sets are equivalent. It follows easily from the fact that ‘s entries are all nonnegative and has orthogonal columns that there is a permutation matrix so that (PG)(PG)^t is block diagonal ( swaps the rows of , and from the fact G^tG = I , we know there’s only one nonzero entry on each row of ). From F1 = GG^t 1=1 , we know that (PG)(PG)^t1 = 1 , so the blocks on the diagonal are doubly stochastic (since the matrix is symmetric). Now notice that each block matrix in is of the form xx^t where x >0 and xx^t 1 = 1 . This implies is a multiple of — otherwise, since the rows look like $x_i ( [x_1, \ldots, x_n] )$ , the row corresponding the smallest x_i would have a smaller sum than the others, so xx^t 1 would not be .

It would’ve been nice if the paper spelled this out.

Possibly relevant posts:

A failed attempt to use an SDP to find a nonorthogonal factorization (1/17/2010)
Full rankedness of random submatrices of orthogonal matrices (5/26/2010)
Three problems motivated by Completely Positive decompositions (1/13/2010)

Tags: factorization, matrix, nonnegative, norm, optimization

K-means as a matrix optimization problem

Possibly relevant posts:

Leave a comment

Archives

intelligently designed

light and not filling

Tags