SDDETECT training a detector given a target class and model PD=SDDETECT(DATA,TCLASS,MODEL,options) INPUT DATA a data set TCLASS target class name MODEL untrained pipeline OUTPUT PD detector pipeline OPTIONS 'reject',N Fraction or number of target class samples to reject 'target',name Name of the target decision in the final detector 'non-target',name Name of the non-target decision in the final detector 'test',TS External test set used to set detector threshold with ROC 'nodisplay' Do not display class mapping table 'measures',M Specify ROC measures to be estimated like for sdroc function 'confmat' Store confmat in the ROC (useful for visualization/cost optim) DESCRIPTION SDDETECT provides one-command training of a TCLASS detector on DATA using MODEL. By default it splits DATA into 80% which is used for training the MODEL on target class TCLASS. Decision threshold is fixed using ROC analysis performed on the remaining 20% of data. If external test set is provided with 'test' option, it is used for ROC analysis (complete DATA is used for training a model). If TCLASS is not present in DATA, detector is trained on all DATA samples. In this one-class scenario, SDDETECT by default accepts all samples in DATA. Using 'reject' option, we may specify the number or fraction of rejected objects. SDDETECT PD returns 'TCLASS' and 'non-TCLASS' decisions (this may be changed using 'target' and 'non-target' options). EXAMPLES Train Gaussian detector on 'banana'. If other classes are present in data, use ROC to set threshold. Otherwise, accept all training samples: pd=sddetect(a,'banana',sdgauss) Train k-NN detector on all samples in a and set its threshold by rejecting 1% of samples: pd=sddetect(a,'all',sdknn([],'k',7),'reject',0.01) Estimate true positive rate and precision in the internally-built ROC: pd=sddetect(a,'apple',sdgauss,'measures',{'TPr','apple','precision','apple'}) SEE ALSO SDREJECT, SDGAUSS, SDPARZEN, SDKNN, SDROC READ MORE http://perclass.com/doc/guide/classifiers.html#detector
sddetect
is referenced in examples: