Spectral Clustering a Brief Tutorial
Graph based clustering problems are treating clustering as graph partitioning. In this case, it treats cluster points using eigenvectors.
- good for sparse data
- don’t make assumption on statistics
D(degree matrix) = d(v_i) for i = j, 0 otherwise
which has degree of each vertex
Affinity matrix is i,j pair of affinity
Laplacian L = D - A
Spectrum Matrix is k eigen vectors of L with dimension N*k, N is dimension of eigen vector.
After getting spectral matrix, we calculate k means in a much lower dimension
image segmentation, motion segmentation, image clustering etc.