pyACA.ToolSeqFeatureSel Namespace Reference

Functions
	ToolSeqFeatureSel (V, ClassIdx, iNumFeatures2Select=-1)
	helper function: sequential forward feature selection

Function Documentation

ToolSeqFeatureSel()

ToolSeqFeatureSel	(		V,
			ClassIdx,
			iNumFeatures2Select = -1 )

helper function: sequential forward feature selection

Parameters

V	feature matrix with all observations (dimension iNumFeatures x iNumObservations)
ClassIdx	class labels (length: iNumObservations)
iNumFeatures2Select	target number of features (default: all)

Returns

selFeatureIdx: vector with ordered feature indices (length: iNumFeatures2Select)

AccPerSubset: accuracy for each subset

Definition at line 16 of file ToolSeqFeatureSel.py.

def ToolSeqFeatureSel(V, ClassIdx, iNumFeatures2Select=-1):
 
    iNumFeatures = V.shape[0]
    if iNumFeatures2Select <= 0 or iNumFeatures2Select > iNumFeatures:
        iNumFeatures2Select = iNumFeatures
 
    # initialize
    selFeatureIdx = -1 * np.ones(iNumFeatures2Select).astype(int)
    unselFeatures = np.ones(iNumFeatures).astype(bool)
    AccPerSubset = np.zeros(iNumFeatures2Select)
    accTmp = np.zeros(iNumFeatures)
 
    # find single best feature
    for f in range(iNumFeatures):
        # get accuracy of selected features plus current feature f
        accTmp[f], fold_accuracies, conf_mat = ToolLooCrossVal(V[f, :][None, :], ClassIdx)
    selFeatureIdx[0] = np.argmax(accTmp, axis=0).astype(int)
    unselFeatures[selFeatureIdx[0]] = False
    AccPerSubset[0] = accTmp[selFeatureIdx[0]]
 
    # iterate until target number of features is reached
    for i in np.arange(1, iNumFeatures2Select):
        # iterate over all features not yet selected
        for f in range(iNumFeatures):
            if unselFeatures[f]:
                # get accuracy of selected features plus current feature f
                subset = np.zeros(i+1).astype(int)
                subset[:i] = selFeatureIdx[:i].astype(int)
                subset[i] = int(f)
                accTmp[f], fold_accuracies, conf_mat = ToolLooCrossVal(V[subset, :], ClassIdx)
            else:
                accTmp[f] = -1
                continue
        
        # identify feature maximizing the accuracy
        # move feature from unselected to selected
        selFeatureIdx[i] = np.argmax(accTmp, axis=0).astype(int)
        unselFeatures[selFeatureIdx[i]] = False
        AccPerSubset[i] = accTmp[selFeatureIdx[i]]
 
    return selFeatureIdx, AccPerSubset