SDPCA Principal Component Analysis P=SDPCA(DATA) % remove directions with zero variance P=SDPCA(DATA,DIM) % project to DIM-D subspace P=SDPCA(DATA,FRAC) % project preserving FRAC of variance [P,RES]=SDPCA(DATA,PPL) % optimize PCA dimensionality for classifier PPL INPUT DATA SDDATA set or data matrix DIM Output dimensionality FRAC Fraction of preserved variance (0,1) PPL Untrained classifier pipeline (such as sdlinear) OPTIONS 'no display' Do not print any output 'test',TS When optimizing dimensionality, do not split data but use externally provided set TS 'tsfrac',F If optimizing dimensionality by classifier error, set the fraction of DATA used for testing/validation (default: 0.2) OUTPUT P PCA projection RES Structure with details on optimization PARAMETERS 'weights' relative input feature importance in the projection DESCRIPTION SDPCA implements Principal Component Analysis projection maximizing variance in the data set DATA. SDPCA training is unsupervised meaning that the class labels are not used. If called without additional parameter, SDPCA projects data to a subspace with non-zero eigenvalues. This approach is also useful to ensure orthogonalization of data (useful e.g. for decision tree or random forest classifiers) SDPCA may also optimize dimensionality for a specific classifier. The DATA is split into training and validation subsets. PPL classifier is trained on the training subset and evaluated on the validation part. Dimensionality yielding lowest mean error is found. Details on the optimization process are returned in RES structure. EXAMPLES p=sdpca(data) % projection removing dimensions with zero variance p=sdpca(data,10) % projection to 10D subspace p=sdpca(data,0.99) % projection to a subspace preserving 99% of variance p=sdpca(data,sdlinear) % optimizing dimensionality minimizing SDLINEAR error READ MORE http://perclass.com/doc/latest/guide/dimensionality_reduction.html#sdpca SEE ALSO SDLDA
sdpca
is referenced in examples: