Covariate Shift Weights

Problem

Compute covariate shift importance weights for training samples: (n_train/n_test) * p / (1 - p), where p is the adversarial probability of belonging to the test set, clipping p to [1e-12, 1 - 1e-12].

Return zeros if n_test is 0; empty input returns [].

Hint

Samples that look like the test set get larger training weight.

Test cases

case 1

in: [[0.5,0.8],100,50]

expected: [2,8.000000000000002]

case 2

in: [[0.2],200,100]

expected: [0.5]

case 3

in: [[0.5],10,5]

expected: [2]

case 4

in: [[],10,5]

expected: []

Starter codePython

def covariate_shift_weights(probs, n_train, n_test):
    # Your code here
    pass

4 test cases · run your code against them in the editor