1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
|
from collections import defaultdict
from sklearn.metrics import auc, precision_recall_curve
import numpy as np
def adjusted_aupr(y_true, y_score, force_monotonic=False):
# Adapted from https://github.com/yesimon/metax_bakeoff_2019/blob/master/plotting/Metagenomics%20Bench.ipynb
original_precision, recall, thresholds = precision_recall_curve(y_true, y_score)
precision = original_precision
if force_monotonic:
# adjusts precision per each recall valu on the curve
# it guarantees that the curve is monotonic decreasing
precision_max_per_recall = defaultdict(float)
for r, p in zip(recall, original_precision):
precision_max_per_recall[r] = max(precision_max_per_recall[r], p)
adjusted_precision = []
for r, p in zip(recall, original_precision):
adjusted_precision.append(precision_max_per_recall[r])
precision = adjusted_precision
# force start at zero
if thresholds[0] == 0:
precision[0] = 0
recall[0] = recall[1]
recall = np.insert(recall, 0, 1)
precision = np.insert(precision, 0, 0)
aupr = auc(recall, precision)
return {k: v for k, v in {
"aupr": aupr,
"original_precision": original_precision if force_monotonic else None,
"recall": recall,
"precision": precision,
"thresholds": thresholds,
}.items() if v is not None}
|
