Unsupervised and Supervised Principal Component Analysis: Tutorial Benyamin Ghojogh
[email protected] Department of Electrical and Computer Engineering, Machine Learning Laboratory, University of Waterloo, Waterloo, ON, Canada Mark Crowley
[email protected] Department of Electrical and Computer Engineering, Machine Learning Laboratory, University of Waterloo, Waterloo, ON, Canada Abstract space and manifold learning (Friedman et al., 2009). This This is a detailed tutorial paper which explains method, which is also used for feature extraction (Ghojogh the Principal Component Analysis (PCA), Su- et al., 2019b), was first proposed by (Pearson, 1901). In or- pervised PCA (SPCA), kernel PCA, and kernel der to learn a nonlinear sub-manifold, kernel PCA was pro- SPCA. We start with projection, PCA with eigen- posed, first by (Scholkopf¨ et al., 1997; 1998), which maps decomposition, PCA with one and multiple pro- the data to high dimensional feature space hoping to fall on jection directions, properties of the projection a linear manifold in that space. matrix, reconstruction error minimization, and The PCA and kernel PCA are unsupervised methods for we connect to auto-encoder. Then, PCA with sin- subspace learning. To use the class labels in PCA, super- gular value decomposition, dual PCA, and ker- vised PCA was proposed (Bair et al., 2006) which scores nel PCA are covered. SPCA using both scor- the features of the X and reduces the features before ap- ing and Hilbert-Schmidt independence criterion plying PCA. This type of SPCA were mostly used in bioin- are explained. Kernel SPCA using both direct formatics (Ma & Dai, 2011). and dual approaches are then introduced.