SVMKFolds.py

#-*- coding: utf-8 -*-

from sklearn import svm
import scipy

def SVMKFolds(k, sentences, dictionary, kernel = 'linear', C = 1.0, gamma = 0.001, times = 1):
	precisions = []
	recalls = []
	accuracies = []
	f1s = []
	testsize = len(sentences)/k
	# Correct Prediction, True Positive, True Negative, Incorrect Prediction, False Positive, False Negative
	counts = [{"CP":0, "TP": 0, "TN":0, "IP":0, "FP":0, "FN":0} for t in xrange(k*times)]
	for t in xrange(k*times):
		numpy.random.shuffle(sentences)
		X, y = Vectorize(sentences, dictionary)
		#Define classifier
		clf = svm.SVC(kernel = kernel, C = C, gamma = gamma)
		clf.fit(X[:-testsize],y[:-testsize])
		#Test data
		for i in range(1, testsize+1):
			predicted = clf.predict(X[-i].reshape(1,-1))[0]
			true_value = y[-i]
			if (predicted == true_value): # test data
				counts[t]["CP"] += 1
				if predicted == 1:
					counts[t]["TP"] += 1
				else:
					counts[t]["TN"] += 1
			else:
				counts[t]["IP"] += 1
				if predicted == 1:
					counts[t]["FP"] += 1
				else:
					counts[t]["FN"] += 1
		# precision = true_positives / (true_positives + false_positives)
		# recall = true_positives / (true_positives + false_negatives)
		# accuracy = (true_positives + true_negatives) / (true_positives + true_negatives + false_positives + false_negatives)
		if counts[t]["TP"]+counts[t]["FP"]>0:
			precision = counts[t]["TP"] / ((counts[t]["TP"] + counts[t]["FP"])*1.0)
		else:
			precision = 0
		if counts[t]["TP"]+counts[t]["FN"]>0:
			recall = counts[t]["TP"] / ((counts[t]["TP"] + counts[t]["FN"])*1.0)
		else:
			recall = 0
		accuracy = counts[t]["CP"]/(testsize*1.0)
		if precision>0 or recall>0:
            F1 = 2* ((precision*recall)/(precision+recall))
        else:
            F1 = 0
		#
		accuracies.append(accuracy)
		recalls.append(recall)
		precisions.append(precision)
		f1s.append(F1)
	avpr = sum(precisions)/(len(precisions)*1.0)
	stpr = scipy.std(precisions)
	avre = sum(recalls)/(len(recalls)*1.0)
	stre = scipy.std(recalls)
	avac = sum(accuracies)/(len(accuracies)*1.0)
	stac = scipy.std(accuracies)
	avf1 = sum(f1s)/(len(f1s)*1.0)
	stf1 = scipy.std(f1s)
	results = [[avpr, stpr, precisions], [avre, stre, recalls], [avac, stac, accuracies], [avf1, stf1, f1s], counts]
	return results

if __name__ == "__main__":
	pass