https://stackoverflow.com/questions/51852415/weird-svm-behavior-in-r-e1071#

exp.py

import numpy as np
from sklearn import svm
import sklearn as sk

def exp(majority_vote=False, n=1000):
    
    # data
    y = np.array([1,2] * 14)
    x = np.random.rand(len(y), 100)

    # prediction results 
    results = np.zeros((len(y), n))
    results[:] = np.nan

    # AUC values of each iteration
    aucs = np.zeros(n)
    aucs[:] = np.nan
    
    # precision values of each iteration
    precs = np.zeros(n)
    precs[:] = np.nan
    
    # accuracy values of each iteration
    accs = np.zeros(n)
    accs[:] = np.nan

    bias_in_train = 0
    majority = 0
    for i in range(0, n):

        #print("Iter ", i)

        # select train and test set

        perm = np.random.permutation(range(0, len(y)))

        train = perm[0:int(len(y)/2)]
        test = perm[int(len(y)/2):]
        #print("Train indices: ", train)
        #print("Test indices:  ", test)

        x_train = x[train,]
        y_train = y[train]

        x_test = x[test,]
        y_test = y[test]

        if majority_vote:
            # train and predict (majority)
            counts = np.unique(y_train, return_counts=True)
            pred = counts[0][np.argmax(counts[1])]
            prediction = np.repeat(pred, len(y_test))
        else:
            # train and predict (SVM)
            clf = svm.SVC()
            clf.fit(x_train, y_train)
            prediction = clf.predict(x_test)
        
        # check majority vote
        y_u = np.unique(y_train, return_counts=True)
        p_u = np.unique(prediction, return_counts=True)
        #print(y_u, p_u)
        if y_u[1][0] != y_u[1][1]:
            bias_in_train += 1
            idx = np.argmax(y_u[1])
            val = y_u[0][idx]
            if len(p_u[0]) is 1 and p_u[0][0] == val:
                #print("True")
                majority += 1
        #else:
            #print("False")
        #print("Train:      ", y_train)
        #print(np.unique(y_train, return_counts=True))
        #print("Test:       ", y_test)
        #print("Prediction: ", prediction)

        # store results
        results[test, i] = prediction

        # calculate AUC
        aucs[i] =  sk.metrics.roc_auc_score(y_test == 2, prediction)
        precs[i] =  sk.metrics.precision_score(y_test == 2, prediction == 2)
        accs[i] =  sk.metrics.accuracy_score(y_test == 2, prediction == 2)
        
        #print(aucs[i], y_test == 2, prediction)

    classification_mean = np.nanmean(results, axis=1)
    auc = sk.metrics.roc_auc_score(y == 2, np.round(means))
    auc_mean = aucs.mean()
    
    prec = sk.metrics.precision_score(y == 2, np.round(means) == 2)
    prec_mean = precs.mean()
    
    acc = sk.metrics.accuracy_score(y == 2, np.round(means) == 2)
    acc_mean = accs.mean()
    
    # calculate precision using the result matrix
    tp = (results == 2) & (np.transpose(np.repeat(np.array([y]), n, axis=0)) == 2)
    fp = (results == 2) & (np.transpose(np.repeat(np.array([y]), n, axis=0)) == 1)
    prec_correct = np.nan_to_num((tp.sum(axis=0) / (tp.sum(axis=0) + fp.sum(axis=0)))).mean()
    prec_overall = (tp.sum() / (tp.sum() + fp.sum()))
    
    # calculate accuracy using the result matrix
    correct = np.equal(results, np.transpose(np.repeat(np.array([y]), n, axis=0))).sum(axis=0)
    overall = np.sum(np.nan_to_num(results) > 0, axis=0)
    acc_correct = (correct / overall).mean()
    acc_overall = (correct.sum() / overall.sum())
    
    print("---")
    print("AUC:      {:.4f} (meanRows) / {:.4f} (independent)".format(auc, auc_mean))
    print("Prec:     {:.4f} (meanRows) / {:.4f} (independent) / {:.4f} (from matrix, independent) / {:.4f} (from matrix, overall)".format(prec, prec_mean, prec_correct, prec_overall))
    print("Acc:      {:.4f} (meanRows) / {:.4f} (independent) / {:.4f} (from matrix, independent) / {:.4f} (from matrix, overall)".format(acc, acc_mean, acc_correct, acc_overall))
    print("Runs: {}, Class bias in train: {}, Majority vote: {}".format(n, bias_in_train, majority))
    return auc, auc_mean

exp(majority_vote=False, n=100)