def eval_cluster_f(self, gold_cluster, test_cluster):
"""Computes the clustering f-score of the gold and test cluster,
where prec = |overlap| / |test|, rec = |overlap| / |gold|.
Returns two tuples of values: the first is (prec, rec, f)
, and the second is (|overlap|, |gold|, |test|)."""
matched = self.by_test[test_cluster][gold_cluster]
proposed = sum(self.by_test[test_cluster].values())
true = self.gold_sizes[gold_cluster]
(p,r,f) = precision_recall_f(matched, true, proposed)
return (p, r, f), (matched, true, proposed)
def macro_average_f(self):
"""Evaluates the macro-average f-score. Macro-averaging adds
number matched and cluster sizes for each pair of clusters,
then takes the f-score at the end. Clusters are matched to
maximize overlap, though this does not necessarily maximize
the metric itself."""
match = 0
prop = 0
true = 0
for gold in self.all_gold:
counts = [self.eval_cluster_f(gold, test)[1]
for test in self.all_test]
best = max(counts, key=lambda x: x[0]) #matched
match += best[0]
true += best[1]
prop += best[2]
(p,r,f) = precision_recall_f(match, true, prop)
return f
def prec_rec(self):
"""Calculates the classification precision, recall and F-score
over edges, with respect to the same-cluster class."""
N00,N11,N01,N10 = self.pairwise_statistics
return precision_recall_f(N11, (N11 + N10), (N11 + N01))
