抄録
PCA-guided k-Means derives a continuous (relaxed) solution of the cluster membership indicators in k-Means.
A robust k-Means model can be formulated based on a fuzzy PCA-guided clustering procedure where a responsibility weight of each sample in k-Means process is estimated based on the noise fuzzy clustering mechanism, and cluster membership indicators in k-Means process are derived as fuzzy principal components considering the responsibility weights in fuzzy PCA.
In this paper, kernel method is applied to the fuzzy PCA-guided k-Means in order to extract a larger number of clusters than the dimensionality of a data set.
Considering mapping to a high dimensional space, we estimate meaningful principal components that are used for capturing nonlinear classification boundaries.
Cluster structures can be visually assessed by the spectral ordering, in which samples are arranged based on mutual connectivity weights, and are emphasized in the diagonal block structure in the connectivity matrix.
