eval-ttest-mslr.py 文件源码

def ttest(filename1, filename2):
    qids1, values1 = load_evaluation_file(arguments.filename1)
    qids2, values2 = load_evaluation_file(arguments.filename2)

    if qids1.shape[0] != qids2.shape[0]:
        raise ValueError('number of queries in files do not match (%d != %d)'\
                         % (qids1.shape[0], qids2.shape[0]))

    qids1_sort_idxs = np.argsort(qids1)
    qids2_sort_idxs = np.argsort(qids2)

    qids1 = qids1[qids1_sort_idxs]
    qids2 = qids2[qids2_sort_idxs]

    if np.any(qids1 != qids2):
        raise ValueError('files do not contain the same queries')

    values1 = values1[qids1_sort_idxs]
    values2 = values2[qids2_sort_idxs]

    mean1 = np.mean(values1)
    mean2 = np.mean(values2)

    t_statistic, p_value = ttest_rel(values1, values2)    

    return values1.shape[0], mean1, mean2, t_statistic, p_value