def _entropy_scores(cm): """Given a ClusteringMetrics object, calculate three entropy-based metrics (Alternative implementation for testing) """ H_C = fentropy(cm.row_totals) H_K = fentropy(cm.col_totals) H_CK = sum(fentropy(col) for col in cm.iter_cols()) H_KC = sum(fentropy(row) for row in cm.iter_rows()) # The '<=' comparisons below both prevent division by zero errors # and ensure that the scores are non-negative. homogeneity = 0.0 if H_C <= H_CK else (H_C - H_CK) / H_C completeness = 0.0 if H_K <= H_KC else (H_K - H_KC) / H_K nmi_score = harmonic_mean(homogeneity, completeness) return homogeneity, completeness, nmi_score
def test_2x2_invariants(): """Alternative implementations should coincide for 2x2 matrices """ for _ in xrange(100): cm = ConfusionMatrix2.from_random_counts(low=0, high=10) # object idempotency assert_equal( cm.to_ccw(), ConfusionMatrix2.from_ccw(*cm.to_ccw()).to_ccw(), msg="must be able to convert to tuple and create from tuple") # pairwise H, C, V h, c, v = cm.pairwise_hcv()[:3] check_with_nans(v, geometric_mean(h, c), ensure_nans=False) # informedness actual_info = cm.informedness() expected_info_1 = cm.TPR() + cm.TNR() - 1.0 expected_info_2 = cm.TPR() - cm.FPR() check_with_nans(actual_info, expected_info_1, 4, ensure_nans=False) check_with_nans(actual_info, expected_info_2, 4, ensure_nans=False) # markedness actual_mark = cm.markedness() expected_mark_1 = cm.PPV() + cm.NPV() - 1.0 expected_mark_2 = cm.PPV() - cm.FOR() check_with_nans(actual_mark, expected_mark_1, 4, ensure_nans=False) check_with_nans(actual_mark, expected_mark_2, 4, ensure_nans=False) # matthews corr coeff # actual_mcc = cm.matthews_corr() # expected_mcc = geometric_mean(actual_info, actual_mark) # check_with_nans(actual_mcc, expected_mcc, 4, ensure_nans=False) # kappas actual_kappa = cm.kappa() # kappa is the same as harmonic mean of kappa components expected_kappa_1 = harmonic_mean(*cm.kappas()[:2]) check_with_nans(actual_kappa, expected_kappa_1, 4, ensure_nans=False) # kappa is the same as accuracy adjusted for chance expected_kappa_2 = harmonic_mean(*cm.adjust_to_null(cm.accuracy, model='m3')) check_with_nans(actual_kappa, expected_kappa_2, 4, ensure_nans=False) # kappa is the same as Dice coeff adjusted for chance expected_kappa_3 = harmonic_mean(*cm.adjust_to_null(cm.dice_coeff, model='m3')) check_with_nans(actual_kappa, expected_kappa_3, 4, ensure_nans=False) # odds ratio and Yule's Q actual_odds_ratio = cm.DOR() actual_yule_q = cm.yule_q() expected_yule_q = _div(actual_odds_ratio - 1.0, actual_odds_ratio + 1.0) expected_odds_ratio = _div(cm.PLL(), cm.NLL()) check_with_nans(actual_odds_ratio, expected_odds_ratio, 4, ensure_nans=False) check_with_nans(actual_yule_q, expected_yule_q, 4, ensure_nans=False) # F-score and Dice expected_f = harmonic_mean(cm.precision(), cm.recall()) actual_f = cm.fscore() check_with_nans(expected_f, actual_f, 6) check_with_nans(expected_f, cm.dice_coeff(), 6, ensure_nans=False) # association coefficients (1) dice = cm.dice_coeff() expected_jaccard = _div(dice, 2.0 - dice) actual_jaccard = cm.jaccard_coeff() check_with_nans(actual_jaccard, expected_jaccard, 6, ensure_nans=False) # association coefficients (2) jaccard = cm.jaccard_coeff() expected_ss2 = _div(jaccard, 2.0 - jaccard) actual_ss2 = cm.sokal_sneath_coeff() check_with_nans(actual_ss2, expected_ss2, 6, ensure_nans=False) # adjusted ochiai actual = cm.ochiai_coeff_adj() expected = harmonic_mean(*cm.adjust_to_null(cm.ochiai_coeff, model='m3')) check_with_nans(actual, expected, 6, ensure_nans=False)