|
1
|
|
|
"""Evaluation metrics for Annif""" |
|
2
|
|
|
|
|
3
|
|
|
import collections |
|
4
|
|
|
import statistics |
|
5
|
|
|
import warnings |
|
6
|
|
|
import numpy as np |
|
7
|
|
|
from sklearn.metrics import precision_score, recall_score, f1_score |
|
8
|
|
|
from sklearn.metrics import label_ranking_average_precision_score |
|
9
|
|
|
|
|
10
|
|
|
|
|
11
|
|
|
def true_positives(y_true, y_pred): |
|
12
|
|
|
"""calculate the number of true positives using bitwise operations, |
|
13
|
|
|
emulating the way sklearn evaluation metric functions work""" |
|
14
|
|
|
return (y_true & y_pred).sum() |
|
15
|
|
|
|
|
16
|
|
|
|
|
17
|
|
|
def false_positives(y_true, y_pred): |
|
18
|
|
|
"""calculate the number of false positives using bitwise operations, |
|
19
|
|
|
emulating the way sklearn evaluation metric functions work""" |
|
20
|
|
|
return (~y_true & y_pred).sum() |
|
21
|
|
|
|
|
22
|
|
|
|
|
23
|
|
|
def false_negatives(y_true, y_pred): |
|
24
|
|
|
"""calculate the number of false negatives using bitwise operations, |
|
25
|
|
|
emulating the way sklearn evaluation metric functions work""" |
|
26
|
|
|
return (y_true & ~y_pred).sum() |
|
27
|
|
|
|
|
28
|
|
|
|
|
29
|
|
|
def precision_at_k_score(y_true, y_pred, limit): |
|
30
|
|
|
"""calculate the precision at K, i.e. the number of relevant items |
|
31
|
|
|
among the top K predicted ones""" |
|
32
|
|
|
scores = [] |
|
33
|
|
|
for true, pred in zip(y_true, y_pred): |
|
34
|
|
|
order = pred.argsort()[::-1] |
|
35
|
|
|
limit = min(limit, np.count_nonzero(pred)) |
|
36
|
|
|
order = order[:limit] |
|
37
|
|
|
gain = true[order] |
|
38
|
|
|
scores.append(gain.sum() / limit) |
|
39
|
|
|
return statistics.mean(scores) |
|
40
|
|
|
|
|
41
|
|
|
|
|
42
|
|
|
def dcg_score(y_true, y_pred, limit=None): |
|
43
|
|
|
"""return the discounted cumulative gain (DCG) score for the selected |
|
44
|
|
|
labels vs. relevant labels""" |
|
45
|
|
|
order = y_pred.argsort()[::-1] |
|
46
|
|
|
n_pred = np.count_nonzero(y_pred) |
|
47
|
|
|
if limit is not None: |
|
48
|
|
|
n_pred = min(limit, n_pred) |
|
49
|
|
|
order = order[:n_pred] |
|
50
|
|
|
gain = y_true[order] |
|
51
|
|
|
discount = np.log2(np.arange(order.size) + 2) |
|
52
|
|
|
|
|
53
|
|
|
return (gain / discount).sum() |
|
54
|
|
|
|
|
55
|
|
|
|
|
56
|
|
|
def ndcg_score(y_true, y_pred, limit=None): |
|
57
|
|
|
"""return the normalized discounted cumulative gain (nDCG) score for the |
|
58
|
|
|
selected labels vs. relevant labels""" |
|
59
|
|
|
scores = [] |
|
60
|
|
|
for true, pred in zip(y_true, y_pred): |
|
61
|
|
|
idcg = dcg_score(true, true, limit) |
|
62
|
|
|
dcg = dcg_score(true, pred, limit) |
|
63
|
|
|
if idcg > 0: |
|
64
|
|
|
scores.append(dcg / idcg) |
|
65
|
|
|
else: |
|
66
|
|
|
scores.append(1.0) # perfect score for no relevant hits case |
|
67
|
|
|
return statistics.mean(scores) |
|
68
|
|
|
|
|
69
|
|
|
|
|
70
|
|
|
class EvaluationBatch: |
|
71
|
|
|
"""A class for evaluating batches of results using all available metrics. |
|
72
|
|
|
The evaluate() method is called once per document in the batch. |
|
73
|
|
|
Final results can be queried using the results() method.""" |
|
74
|
|
|
|
|
75
|
|
|
def __init__(self, subject_index): |
|
76
|
|
|
self._subject_index = subject_index |
|
77
|
|
|
self._samples = [] |
|
78
|
|
|
|
|
79
|
|
|
def evaluate(self, hits, gold_subjects): |
|
80
|
|
|
self._samples.append((hits, gold_subjects)) |
|
81
|
|
|
|
|
82
|
|
|
def _evaluate_samples(self, y_true, y_pred, metrics='all'): |
|
83
|
|
|
y_pred_binary = y_pred > 0.0 |
|
84
|
|
|
results = collections.OrderedDict() |
|
85
|
|
|
with warnings.catch_warnings(): |
|
86
|
|
|
warnings.simplefilter('ignore') |
|
87
|
|
|
|
|
88
|
|
|
results['Precision (doc avg)'] = precision_score( |
|
89
|
|
|
y_true, y_pred_binary, average='samples') |
|
90
|
|
|
results['Recall (doc avg)'] = recall_score( |
|
91
|
|
|
y_true, y_pred_binary, average='samples') |
|
92
|
|
|
results['F1 score (doc avg)'] = f1_score( |
|
93
|
|
|
y_true, y_pred_binary, average='samples') |
|
94
|
|
|
if metrics == 'all': |
|
95
|
|
|
results['Precision (conc avg)'] = precision_score( |
|
96
|
|
|
y_true, y_pred_binary, average='macro') |
|
97
|
|
|
results['Recall (conc avg)'] = recall_score( |
|
98
|
|
|
y_true, y_pred_binary, average='macro') |
|
99
|
|
|
results['F1 score (conc avg)'] = f1_score( |
|
100
|
|
|
y_true, y_pred_binary, average='macro') |
|
101
|
|
|
results['Precision (microavg)'] = precision_score( |
|
102
|
|
|
y_true, y_pred_binary, average='micro') |
|
103
|
|
|
results['Recall (microavg)'] = recall_score( |
|
104
|
|
|
y_true, y_pred_binary, average='micro') |
|
105
|
|
|
results['F1 score (microavg)'] = f1_score( |
|
106
|
|
|
y_true, y_pred_binary, average='micro') |
|
107
|
|
|
results['NDCG'] = ndcg_score(y_true, y_pred) |
|
108
|
|
|
results['NDCG@5'] = ndcg_score(y_true, y_pred, limit=5) |
|
109
|
|
|
results['NDCG@10'] = ndcg_score(y_true, y_pred, limit=10) |
|
110
|
|
|
if metrics == 'all': |
|
111
|
|
|
results['Precision@1'] = precision_at_k_score( |
|
112
|
|
|
y_true, y_pred, limit=1) |
|
113
|
|
|
results['Precision@3'] = precision_at_k_score( |
|
114
|
|
|
y_true, y_pred, limit=3) |
|
115
|
|
|
results['Precision@5'] = precision_at_k_score( |
|
116
|
|
|
y_true, y_pred, limit=5) |
|
117
|
|
|
results['LRAP'] = label_ranking_average_precision_score( |
|
118
|
|
|
y_true, y_pred) |
|
119
|
|
|
results['True positives'] = true_positives( |
|
120
|
|
|
y_true, y_pred_binary) |
|
121
|
|
|
results['False positives'] = false_positives( |
|
122
|
|
|
y_true, y_pred_binary) |
|
123
|
|
|
results['False negatives'] = false_negatives( |
|
124
|
|
|
y_true, y_pred_binary) |
|
125
|
|
|
|
|
126
|
|
|
return results |
|
127
|
|
|
|
|
128
|
|
|
def results(self, metrics='all'): |
|
129
|
|
|
"""evaluate a set of selected subjects against a gold standard using |
|
130
|
|
|
different metrics. The set of metrics can be either 'all' or |
|
131
|
|
|
'simple'.""" |
|
132
|
|
|
|
|
133
|
|
|
y_true = np.array([gold_subjects.as_vector(self._subject_index) |
|
134
|
|
|
for hits, gold_subjects in self._samples]) |
|
135
|
|
|
y_pred = np.array([hits.vector |
|
136
|
|
|
for hits, gold_subjects in self._samples]) |
|
137
|
|
|
|
|
138
|
|
|
results = self._evaluate_samples( |
|
139
|
|
|
y_true, y_pred, metrics) |
|
140
|
|
|
results['Documents evaluated'] = y_true.shape[0] |
|
141
|
|
|
return results |
|
142
|
|
|
|
NatLibFi /
Annif
