postprocessing.py

#!/usr/bin/env python
# coding: utf-8

# In[ ]:


### NAIVE BAYES FUSION

# Inputs: ytest, ymodels should have multiple model results
# ymodels = []
# ymodels.append(ymodel_dt)
# ymodels.append(ymodel_LR)
# etc 

import numpy as np
from itertools import product
import random


def NB_fusion(ytest, ymodels):
    class0prob = []  # label probabilities for true class 0 for each of the different classifiers
    class1prob = []  # label probabilities for true class 1 for each of the different classifiers

    prior0 = len(ytest[ytest == 0]) / len(ytest)  # prior probability for class 0
    prior1 = len(ytest[ytest == 1]) / len(ytest)  # prior probability for class 1

    for i in range(len(ymodels)):
        p0 = []
        p1 = []

        true0 = ymodels[i][np.where(ytest == 0)]  # samples where true value is class 0
        true0_lab0 = true0[true0 == 0]  # true value is class 0 and label is 0
        true0_lab1 = true0[true0 == 1]  # true value is 0 and label is 1

        p0.append(len(true0_lab0) / len(true0))  # p0[0], probability of label 0 given true 0
        p0.append(len(true0_lab1) / len(true0))  # p0[1], probablity of label 1 given true 0

        class0prob.append(p0)

        true1 = ymodels[i][np.where(ytest == 1)]  # samples where true value is class 1
        true1_lab0 = true1[true1 == 0]  # samples where true value is 1 and label is 0
        true1_lab1 = true1[true1 == 1]  # samples where true value is 1 and label is 1

        p1.append(len(true1_lab0) / len(true1))  # p1[0], probability of label 0 given true 1
        p1.append(len(true1_lab1) / len(true1))  # p1[1], probability of label 1 given true 1

        class1prob.append(p1)

    # possible combinations for labels as an array   
    classifier_labels = np.array(list(product([0, 1], repeat=len(ymodels))))

    final_labels = []
    post_probs = []
    for c in classifier_labels:  # for each combination of classifier labels

        post = []
        cond0 = 1
        cond1 = 1

        for i in range(len(class0prob)):
            cond0 *= class0prob[i][c[i]]

        post.append(prior0 * cond0)  # post[0], calculate posterior probability of class 0

        for i in range(len(class1prob)):
            cond1 *= class1prob[i][c[i]]

        post.append(prior1 * cond1)  # post[1], calculate posterior probability of class 1

        post = np.array(post)
        post_probs.append(
            post)  # posterior probabilities for each class for all combinations of classifier labels
        final_labels.append(post.argmax())  # final label for each combination of classifier labels

    fused_label = []
    ymodels_t = np.transpose(np.array(ymodels))
    for i in range(len(ymodels_t)):
        for c in range(len(classifier_labels)):
            if np.all(ymodels_t[i] == classifier_labels[c]):
                fused_label.append(final_labels[c])

    fused_label = np.array(fused_label)

    return fused_label


def BKS(ytest, ymodels):
    # All possible combinations of classifier labels

    classifier_labels = np.array(list(product([0, 1], repeat=len(ymodels))))

    true0 = np.transpose(np.array(ymodels))[
        np.where(ytest == 0)]  # ymodel results where true class is 0
    true1 = np.transpose(np.array(ymodels))[
        np.where(ytest == 1)]  # ymodel results where true class is 1

    n0 = []
    n1 = []
    final_labels = []

    for c in range(len(classifier_labels)):  # Count number of samples from each TRUE CLASS
        # that are labeled with each combination
        num0 = 0
        num1 = 0

        for i in range(len(true0)):
            if np.all(true0[i] == classifier_labels[c]):
                num0 += 1

        n0.append(num0)

        for i in range(len(true1)):
            if np.all(true1[i] == classifier_labels[c]):
                num1 += 1

        n1.append(num1)

        if num1 > num0:
            final_labels.append(1)

        elif num0 > num1:
            final_labels.append(0)

        elif num0 == num1:
            final_labels.append(random.choice([0, 1]))

    fused_label = []
    ymodels_t = np.transpose(np.array(ymodels))

    for i in range(len(ymodels_t)):
        for c in range(len(classifier_labels)):
            if np.all(ymodels_t[i] == classifier_labels[c]):
                fused_label.append(final_labels[c])

    return fused_label