Set-Oriented Dimension Reduction: Localizing Principal Component Analysis Via Hidden Markov Models