Introduction to Kernel PCA and other Spectral Methods Applied to Unsupervised Learning

In this work, the techniques of Kernel Principal Component Analysis (Kernel PCA or KPCA) and Spectral Clustering are introduced along with some illustrative examples. This work focuses on studying the effects of applying PCA as a preprocessing stage for clustering data. Several tests are carried out on real data to establish the pertinence of including PCA. The use of these methods requires of additional procedures such as parameter tuning; the kernel alignment is presented as an alternative for it. The results of kernel alignment expose a high level of agreement between the timing curves their respective Rand indexes. Finally, the study shows that the success of PCA is problem-dependent and no general criteria can be established.