1
|
|
|
"""Evaluation metrics for Annif""" |
2
|
|
|
|
3
|
|
|
from __future__ import annotations |
4
|
|
|
|
5
|
|
|
import warnings |
6
|
|
|
from typing import TYPE_CHECKING |
7
|
|
|
|
8
|
|
|
import numpy as np |
9
|
|
|
import scipy.sparse |
10
|
|
|
from sklearn.metrics import f1_score, precision_score, recall_score |
11
|
|
|
|
12
|
|
|
from annif.exception import NotSupportedException |
13
|
|
|
from annif.suggestion import SuggestionBatch, filter_suggestion |
14
|
|
|
|
15
|
|
|
if TYPE_CHECKING: |
16
|
|
|
from collections.abc import Iterable, Iterator, Sequence |
17
|
|
|
from io import TextIOWrapper |
18
|
|
|
|
19
|
|
|
from click.utils import LazyFile |
20
|
|
|
from scipy.sparse._arrays import csr_array |
21
|
|
|
|
22
|
|
|
from annif.corpus.subject import SubjectIndex, SubjectSet |
23
|
|
|
from annif.suggestion import SubjectSuggestion |
24
|
|
|
|
25
|
|
|
|
26
|
|
|
def true_positives(y_true: csr_array, y_pred: csr_array) -> int: |
27
|
|
|
"""calculate the number of true positives using bitwise operations, |
28
|
|
|
emulating the way sklearn evaluation metric functions work""" |
29
|
|
|
return int((y_true.multiply(y_pred)).sum()) |
30
|
|
|
|
31
|
|
|
|
32
|
|
|
def false_positives(y_true: csr_array, y_pred: csr_array) -> int: |
33
|
|
|
"""calculate the number of false positives using bitwise operations, |
34
|
|
|
emulating the way sklearn evaluation metric functions work""" |
35
|
|
|
return int((y_true < y_pred).sum()) |
36
|
|
|
|
37
|
|
|
|
38
|
|
|
def false_negatives(y_true: csr_array, y_pred: csr_array) -> int: |
39
|
|
|
"""calculate the number of false negatives using bitwise operations, |
40
|
|
|
emulating the way sklearn evaluation metric functions work""" |
41
|
|
|
return int((y_true > y_pred).sum()) |
42
|
|
|
|
43
|
|
|
|
44
|
|
|
def dcg_score( |
45
|
|
|
y_true: csr_array, y_pred: csr_array, limit: int | None = None |
46
|
|
|
) -> np.float64: |
47
|
|
|
"""return the discounted cumulative gain (DCG) score for the selected |
48
|
|
|
labels vs. relevant labels""" |
49
|
|
|
|
50
|
|
|
n_pred = y_pred.count_nonzero() |
51
|
|
|
if limit is not None: |
52
|
|
|
n_pred = min(limit, n_pred) |
53
|
|
|
|
54
|
|
|
top_k = y_pred.data.argsort()[-n_pred:][::-1] |
55
|
|
|
order = y_pred.indices[top_k] |
56
|
|
|
gain = y_true[:, order] |
57
|
|
|
discount = np.log2(np.arange(1, n_pred + 1) + 1) |
58
|
|
|
return (gain / discount).sum() |
59
|
|
|
|
60
|
|
|
|
61
|
|
|
def ndcg_score(y_true: csr_array, y_pred: csr_array, limit: int | None = None) -> float: |
62
|
|
|
"""return the normalized discounted cumulative gain (nDCG) score for the |
63
|
|
|
selected labels vs. relevant labels""" |
64
|
|
|
|
65
|
|
|
scores = np.ones(y_true.shape[0], dtype=np.float32) |
66
|
|
|
for i in range(y_true.shape[0]): |
67
|
|
|
true = y_true[[i]] |
68
|
|
|
idcg = dcg_score(true, true, limit) |
69
|
|
|
if idcg > 0: |
70
|
|
|
pred = y_pred[[i]] |
71
|
|
|
dcg = dcg_score(true, pred, limit) |
72
|
|
|
scores[i] = dcg / idcg |
73
|
|
|
|
74
|
|
|
return float(scores.mean()) |
75
|
|
|
|
76
|
|
|
|
77
|
|
|
class EvaluationBatch: |
78
|
|
|
"""A class for evaluating batches of results using all available metrics. |
79
|
|
|
The evaluate() method is called once per document in the batch or evaluate_many() |
80
|
|
|
for a list of documents of the batch. Final results can be queried using the |
81
|
|
|
results() method.""" |
82
|
|
|
|
83
|
|
|
def __init__(self, subject_index: SubjectIndex) -> None: |
84
|
|
|
self._subject_index = subject_index |
85
|
|
|
self._suggestion_arrays = [] |
86
|
|
|
self._gold_subject_arrays = [] |
87
|
|
|
|
88
|
|
|
def evaluate_many( |
89
|
|
|
self, |
90
|
|
|
suggestion_batch: ( |
91
|
|
|
list[list[SubjectSuggestion]] | SuggestionBatch | list[Iterator] |
92
|
|
|
), |
93
|
|
|
gold_subject_batch: Sequence[SubjectSet], |
94
|
|
|
) -> None: |
95
|
|
|
if not isinstance(suggestion_batch, SuggestionBatch): |
96
|
|
|
suggestion_batch = SuggestionBatch.from_sequence( |
97
|
|
|
suggestion_batch, self._subject_index |
98
|
|
|
) |
99
|
|
|
self._suggestion_arrays.append(suggestion_batch.array) |
100
|
|
|
|
101
|
|
|
# convert gold_subject_batch to sparse matrix |
102
|
|
|
ar = scipy.sparse.dok_array( |
103
|
|
|
(len(gold_subject_batch), len(self._subject_index)), dtype=bool |
104
|
|
|
) |
105
|
|
|
for idx, subject_set in enumerate(gold_subject_batch): |
106
|
|
|
for subject_id in subject_set: |
107
|
|
|
ar[idx, subject_id] = True |
108
|
|
|
self._gold_subject_arrays.append(ar.tocsr()) |
109
|
|
|
|
110
|
|
|
def _evaluate_samples( |
111
|
|
|
self, |
112
|
|
|
y_true: csr_array, |
113
|
|
|
y_pred: csr_array, |
114
|
|
|
metrics: Iterable[str] = [], |
115
|
|
|
) -> dict[str, float]: |
116
|
|
|
y_pred_binary = y_pred > 0.0 |
117
|
|
|
|
118
|
|
|
# define the available metrics as lazy lambda functions |
119
|
|
|
# so we can execute only the ones actually requested |
120
|
|
|
all_metrics = { |
121
|
|
|
"Precision (doc avg)": lambda: precision_score( |
122
|
|
|
y_true, y_pred_binary, average="samples" |
123
|
|
|
), |
124
|
|
|
"Recall (doc avg)": lambda: recall_score( |
125
|
|
|
y_true, y_pred_binary, average="samples" |
126
|
|
|
), |
127
|
|
|
"F1 score (doc avg)": lambda: f1_score( |
128
|
|
|
y_true, y_pred_binary, average="samples" |
129
|
|
|
), |
130
|
|
|
"Precision (subj avg)": lambda: precision_score( |
131
|
|
|
y_true, y_pred_binary, average="macro" |
132
|
|
|
), |
133
|
|
|
"Recall (subj avg)": lambda: recall_score( |
134
|
|
|
y_true, y_pred_binary, average="macro" |
135
|
|
|
), |
136
|
|
|
"F1 score (subj avg)": lambda: f1_score( |
137
|
|
|
y_true, y_pred_binary, average="macro" |
138
|
|
|
), |
139
|
|
|
"Precision (weighted subj avg)": lambda: precision_score( |
140
|
|
|
y_true, y_pred_binary, average="weighted" |
141
|
|
|
), |
142
|
|
|
"Recall (weighted subj avg)": lambda: recall_score( |
143
|
|
|
y_true, y_pred_binary, average="weighted" |
144
|
|
|
), |
145
|
|
|
"F1 score (weighted subj avg)": lambda: f1_score( |
146
|
|
|
y_true, y_pred_binary, average="weighted" |
147
|
|
|
), |
148
|
|
|
"Precision (microavg)": lambda: precision_score( |
149
|
|
|
y_true, y_pred_binary, average="micro" |
150
|
|
|
), |
151
|
|
|
"Recall (microavg)": lambda: recall_score( |
152
|
|
|
y_true, y_pred_binary, average="micro" |
153
|
|
|
), |
154
|
|
|
"F1 score (microavg)": lambda: f1_score( |
155
|
|
|
y_true, y_pred_binary, average="micro" |
156
|
|
|
), |
157
|
|
|
"F1@5": lambda: f1_score( |
158
|
|
|
y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples" |
159
|
|
|
), |
160
|
|
|
"NDCG": lambda: ndcg_score(y_true, y_pred), |
161
|
|
|
"NDCG@5": lambda: ndcg_score(y_true, y_pred, limit=5), |
162
|
|
|
"NDCG@10": lambda: ndcg_score(y_true, y_pred, limit=10), |
163
|
|
|
"Precision@1": lambda: precision_score( |
164
|
|
|
y_true, filter_suggestion(y_pred, 1) > 0.0, average="samples" |
165
|
|
|
), |
166
|
|
|
"Precision@3": lambda: precision_score( |
167
|
|
|
y_true, filter_suggestion(y_pred, 3) > 0.0, average="samples" |
168
|
|
|
), |
169
|
|
|
"Precision@5": lambda: precision_score( |
170
|
|
|
y_true, filter_suggestion(y_pred, 5) > 0.0, average="samples" |
171
|
|
|
), |
172
|
|
|
"True positives": lambda: true_positives(y_true, y_pred_binary), |
173
|
|
|
"False positives": lambda: false_positives(y_true, y_pred_binary), |
174
|
|
|
"False negatives": lambda: false_negatives(y_true, y_pred_binary), |
175
|
|
|
} |
176
|
|
|
|
177
|
|
|
if not metrics: |
178
|
|
|
metrics = all_metrics.keys() |
179
|
|
|
|
180
|
|
|
with warnings.catch_warnings(): |
181
|
|
|
warnings.simplefilter("ignore") |
182
|
|
|
|
183
|
|
|
return {metric: all_metrics[metric]() for metric in metrics} |
184
|
|
|
|
185
|
|
|
def _result_per_subject_header( |
186
|
|
|
self, results_file: LazyFile | TextIOWrapper |
187
|
|
|
) -> None: |
188
|
|
|
print( |
189
|
|
|
"\t".join( |
190
|
|
|
[ |
191
|
|
|
"URI", |
192
|
|
|
"Label", |
193
|
|
|
"Support", |
194
|
|
|
"True_positives", |
195
|
|
|
"False_positives", |
196
|
|
|
"False_negatives", |
197
|
|
|
"Precision", |
198
|
|
|
"Recall", |
199
|
|
|
"F1_score", |
200
|
|
|
] |
201
|
|
|
), |
202
|
|
|
file=results_file, |
203
|
|
|
) |
204
|
|
|
|
205
|
|
|
def _result_per_subject_body( |
206
|
|
|
self, zipped_results: zip, results_file: LazyFile | TextIOWrapper |
207
|
|
|
) -> None: |
208
|
|
|
for row in zipped_results: |
209
|
|
|
print("\t".join((str(e) for e in row)), file=results_file) |
|
|
|
|
210
|
|
|
|
211
|
|
|
def output_result_per_subject( |
212
|
|
|
self, |
213
|
|
|
y_true: csr_array, |
214
|
|
|
y_pred: csr_array, |
215
|
|
|
results_file: TextIOWrapper | LazyFile, |
216
|
|
|
language: str, |
217
|
|
|
) -> None: |
218
|
|
|
"""Write results per subject (non-aggregated) |
219
|
|
|
to outputfile results_file, using labels in the given language""" |
220
|
|
|
|
221
|
|
|
y_pred = y_pred.T > 0.0 |
222
|
|
|
y_true = y_true.T |
223
|
|
|
|
224
|
|
|
true_pos = y_true.multiply(y_pred).sum(axis=1) |
225
|
|
|
false_pos = (y_true < y_pred).sum(axis=1) |
226
|
|
|
false_neg = (y_true > y_pred).sum(axis=1) |
227
|
|
|
|
228
|
|
|
with np.errstate(invalid="ignore"): |
229
|
|
|
precision = np.nan_to_num(true_pos / (true_pos + false_pos)) |
230
|
|
|
recall = np.nan_to_num(true_pos / (true_pos + false_neg)) |
231
|
|
|
f1_score = np.nan_to_num(2 * (precision * recall) / (precision + recall)) |
232
|
|
|
|
233
|
|
|
zipped = zip( |
234
|
|
|
[subj.uri for subj in self._subject_index], # URI |
235
|
|
|
[subj.labels[language] for subj in self._subject_index], # Label |
236
|
|
|
y_true.sum(axis=1), # Support |
237
|
|
|
true_pos, # True positives |
238
|
|
|
false_pos, # False positives |
239
|
|
|
false_neg, # False negatives |
240
|
|
|
precision, # Precision |
241
|
|
|
recall, # Recall |
242
|
|
|
f1_score, # F1 score |
243
|
|
|
) |
244
|
|
|
self._result_per_subject_header(results_file) |
245
|
|
|
self._result_per_subject_body(zipped, results_file) |
246
|
|
|
|
247
|
|
|
def results( |
248
|
|
|
self, |
249
|
|
|
metrics: Iterable[str] = [], |
250
|
|
|
results_file: LazyFile | TextIOWrapper | None = None, |
251
|
|
|
language: str | None = None, |
252
|
|
|
) -> dict[str, float]: |
253
|
|
|
"""evaluate a set of selected subjects against a gold standard using |
254
|
|
|
different metrics. If metrics is empty, use all available metrics. |
255
|
|
|
If results_file (file object) given, write results per subject to it |
256
|
|
|
with labels expressed in the given language.""" |
257
|
|
|
|
258
|
|
|
if not self._suggestion_arrays: |
259
|
|
|
raise NotSupportedException("cannot evaluate empty corpus") |
260
|
|
|
|
261
|
|
|
y_pred = scipy.sparse.csr_array(scipy.sparse.vstack(self._suggestion_arrays)) |
262
|
|
|
y_true = scipy.sparse.csr_array(scipy.sparse.vstack(self._gold_subject_arrays)) |
263
|
|
|
|
264
|
|
|
results = self._evaluate_samples(y_true, y_pred, metrics) |
265
|
|
|
results["Documents evaluated"] = int(y_true.shape[0]) |
266
|
|
|
|
267
|
|
|
if results_file: |
268
|
|
|
self.output_result_per_subject(y_true, y_pred, results_file, language) |
269
|
|
|
return results |
270
|
|
|
|