exact_sp.py

import numpy as np
from tqdm import tqdm


def get_true_KNN(x_trn, x_tst):
    N = x_trn.shape[0]
    N_tst = x_tst.shape[0]
    x_tst_knn_gt = np.zeros((N_tst, N))
    for i_tst in tqdm(range(N_tst)):
        dist_gt = np.zeros(N)
        for i_trn in range(N):
            dist_gt[i_trn] = np.linalg.norm(x_trn[i_trn, :] - x_tst[i_tst, :], 2)
        x_tst_knn_gt[i_tst, :] = np.argsort(dist_gt)
    return x_tst_knn_gt.astype(int)


def compute_single_unweighted_knn_class_shapley(x_trn, y_trn, x_tst_knn_gt, y_tst, K):
    N = x_trn.shape[0]
    N_tst = x_tst_knn_gt.shape[0]
    sp_gt = np.zeros((N_tst, N))
    for j in tqdm(range(N_tst)):
        sp_gt[j, x_tst_knn_gt[j, -1]] = (y_trn[x_tst_knn_gt[j, -1]] == y_tst[j]) / N
        for i in np.arange(N - 2, -1, -1):
            sp_gt[j, x_tst_knn_gt[j, i]] = sp_gt[j, x_tst_knn_gt[j, i + 1]] + \
                                           (int(y_trn[x_tst_knn_gt[j, i]] == y_tst[j]) -
                                            int(y_trn[x_tst_knn_gt[j, i + 1]] == y_tst[j])) / K * min([K, i + 1]) / (
                                                       i + 1)
    return sp_gt