Skip to content

Synthesis Validation

For synthesis validation we have only feature based evaluation metrics. To access the

Feature-based

Quality

pymdma.tabular.measures.synthesis_val.ImprovedPrecision

Improved Precision Metric for accessing fidelity of generative models.

Objective: Fidelity

Parameters:

Name Type Description Default
k int

Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.

 5
metric str

The metric to use when calculating distance between instances. For the available metrics, see the documentation of sklearn.metrics.pairwise_distances.

 "euclidean"
n_workers int

Number of workers for computing pairwise distances. Defaults to 4.

 4
**kwargs

Additional keyword arguments for compatiblilty.

 {}
References

Kynkaanniemi et al., Improved Precision and Recall Metric for Assessing Generative Models (2019). https://arxiv.org/abs/1904.06991

Code adapted from: improved-precision-and-recall-metric: Improved Precision and Recall Metric for Assessing Generative Models — Official TensorFlow Implementation. https://github.com/kynkaat/improved-precision-and-recall-metric

Hypersphere estimation code was taken from: generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models. https://github.com/clovaai/generative-evaluation-prdc

Examples:

>>> improved_precision = ImprovedPrecision()
>>> real_features = np.random.rand(100, 100)
>>> fake_features = np.random.rand(100, 100)
>>> result: MetricResult = improved_precision.compute(real_features, fake_features)
Source code in src/pymdma/general/measures/prdc.py 
 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
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
class ImprovedPrecision(FeatureMetric):
    """Improved Precision Metric for accessing fidelity of generative models.

    **Objective**: Fidelity

    Parameters
    ----------
    k : int, optional
        Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.
    metric : str, optional, default="euclidean"
        The metric to use when calculating distance between instances.
        For the available metrics, see the documentation of `sklearn.metrics.pairwise_distances`.
    n_workers : int, optional
        Number of workers for computing pairwise distances. Defaults to 4.
    **kwargs
        Additional keyword arguments for compatiblilty.

    References
    ----------
    Kynkaanniemi et al., Improved Precision and Recall Metric for Assessing Generative Models (2019).
    https://arxiv.org/abs/1904.06991

    Code adapted from:
    improved-precision-and-recall-metric: Improved Precision and Recall Metric for Assessing Generative Models — Official TensorFlow Implementation.
    https://github.com/kynkaat/improved-precision-and-recall-metric

    Hypersphere estimation code was taken from:
    generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models.
    https://github.com/clovaai/generative-evaluation-prdc

    Examples
    --------
    >>> improved_precision = ImprovedPrecision()
    >>> real_features = np.random.rand(100, 100)
    >>> fake_features = np.random.rand(100, 100)
    >>> result: MetricResult = improved_precision.compute(real_features, fake_features)
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = [EvaluationLevel.INSTANCE, EvaluationLevel.DATASET]
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        k: int = 5,
        metric: str = "euclidean",
        n_workers: int = 4,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.k = k
        self.metric = metric
        self.n_workers = n_workers

    def compute(self, real_features: np.ndarray, fake_features: np.ndarray, **kwargs) -> MetricResult:
        """Compute the Improved Precision metric.

        Parameters
        ----------
        real_features : np.ndarray
            Array of shape (n_samples, n_features) containing the real features.
        fake_features : np.ndarray
            Array of shape (n_samples, n_features) containing the fake features.

        Notes
        -----
        Intermediate computations can be stored in the `context` dictionary of the `kwargs` parameter.
        Usefull when calculating multiple metrics that share the same intermediate computations.

        Returns
        -------
        result: MetricResult
            Dataset-level and instance-level results for the precision metric.
        """
        state = kwargs.get("context", {})
        if "real_nn_distances" not in state:
            state["real_nn_distances"] = compute_nearest_neighbour_distances(
                real_features,
                nearest_k=self.k,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        if "real_fake_distances" not in state:
            state["real_fake_distances"] = compute_pairwise_distance(
                real_features,
                fake_features,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        precision = (
            np.logical_or(
                (state["real_fake_distances"] < np.expand_dims(state["real_nn_distances"], axis=1)),
                np.isclose(state["real_fake_distances"], np.expand_dims(state["real_nn_distances"], axis=1)),
            )
            .any(axis=0)
            .astype(int)
        )

        return MetricResult(
            dataset_level={"dtype": OutputsTypes.NUMERIC, "subtype": "float", "value": precision.mean()},
            instance_level={"dtype": OutputsTypes.ARRAY, "subtype": "int", "value": precision.tolist()},
        )

pymdma.tabular.measures.synthesis_val.ImprovedRecall

Improved Recall Metric for accessing diversity of generative models.

Objective: Diversity

Parameters:

Name Type Description Default
k int

Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.

 5
metric str

The metric to use when calculating distance between instances. For the available metrics, see the documentation of sklearn.metrics.pairwise_distances.

 "euclidean"
n_workers int

Number of workers for computing pairwise distances. Defaults to 4.

 4
**kwargs

Additional keyword arguments for compatiblilty.

 {}
References

Kynkaanniemi et al., Improved Precision and Recall Metric for Assessing Generative Models (2019). https://arxiv.org/abs/1904.06991

Code adapted from: improved-precision-and-recall-metric: Improved Precision and Recall Metric for Assessing Generative Models — Official TensorFlow Implementation. https://github.com/kynkaat/improved-precision-and-recall-metric

Hypersphere estimation code was taken from: generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models. https://github.com/clovaai/generative-evaluation-prdc

Examples:

>>> improved_recall = ImprovedRecall()
>>> real_features = np.random.rand(100, 100)
>>> fake_features = np.random.rand(100, 100)
>>> result: MetricResult = improved_recall.compute(real_features, fake_features)
Source code in src/pymdma/general/measures/prdc.py 
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
class ImprovedRecall(FeatureMetric):
    """Improved Recall Metric for accessing diversity of generative models.

    **Objective**: Diversity

    Parameters
    ----------
    k : int, optional
        Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.
    metric : str, optional, default="euclidean"
        The metric to use when calculating distance between instances.
        For the available metrics, see the documentation of `sklearn.metrics.pairwise_distances`.
    n_workers : int, optional
        Number of workers for computing pairwise distances. Defaults to 4.
    **kwargs
        Additional keyword arguments for compatiblilty.

    References
    ----------
    Kynkaanniemi et al., Improved Precision and Recall Metric for Assessing Generative Models (2019).
    https://arxiv.org/abs/1904.06991

    Code adapted from:
    improved-precision-and-recall-metric: Improved Precision and Recall Metric for Assessing Generative Models — Official TensorFlow Implementation.
    https://github.com/kynkaat/improved-precision-and-recall-metric

    Hypersphere estimation code was taken from:
    generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models.
    https://github.com/clovaai/generative-evaluation-prdc

    Examples
    --------
    >>> improved_recall = ImprovedRecall()
    >>> real_features = np.random.rand(100, 100)
    >>> fake_features = np.random.rand(100, 100)
    >>> result: MetricResult = improved_recall.compute(real_features, fake_features)
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = [EvaluationLevel.INSTANCE, EvaluationLevel.DATASET]
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        k: int = 5,
        metric: str = "euclidean",
        n_workers: int = 4,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.k = k
        self.metric = metric
        self.n_workers = n_workers

    def compute(self, real_features: np.ndarray, fake_features: np.ndarray, **kwargs) -> MetricResult:
        """Compute the Improved Recall metric.

        Parameters
        ----------
        real_features : np.ndarray
            Array of shape (n_samples, n_features) containing the real features.
        fake_features : np.ndarray
            Array of shape (n_samples, n_features) containing the fake features.

        Notes
        -----
        Intermediate computations can be stored in the `context` dictionary of the `kwargs` parameter.
        Usefull when calculating multiple metrics that share the same intermediate computations.

        Returns
        -------
        result: MetricResult
            Dataset-level and instance-level results for the recall metric.
        """
        state = kwargs.get("context", {})
        if "fake_nn_distances" not in state:
            state["fake_nn_distances"] = compute_nearest_neighbour_distances(
                fake_features,
                nearest_k=self.k,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        if "real_fake_distances" not in state:
            state["real_fake_distances"] = compute_pairwise_distance(
                real_features,
                fake_features,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        recall_mask = np.logical_or(
            state["real_fake_distances"] < np.expand_dims(state["fake_nn_distances"], axis=0),
            np.isclose(state["real_fake_distances"], np.expand_dims(state["fake_nn_distances"], axis=0)),
        )
        recall = recall_mask.any(axis=1).astype(int)

        # matrix with (R, F) shape -> .any() -> matrix with (F,) shape
        # an array that indicates for each F sample how many real samples are within its manifold
        recall_counts = recall_mask.sum(axis=0)

        return MetricResult(
            dataset_level={"dtype": OutputsTypes.NUMERIC, "subtype": "float", "value": recall.mean()},
            instance_level={"dtype": OutputsTypes.ARRAY, "subtype": "int", "value": recall_counts.tolist()},
        )

pymdma.tabular.measures.synthesis_val.Density

Density Metric for accessing fidelity of the generated samples. Unlike Improved Precision, it is robust towards outliers in the real/reference data.

Objective: Fidelity

Parameters:

Name Type Description Default
k int

Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.

 5
metric str

The metric to use when calculating distance between instances. For the available metrics, see the documentation of sklearn.metrics.pairwise_distances.

 "euclidean"
n_workers int

Number of workers for computing pairwise distances. Defaults to 4.

 4
**kwargs

Additional keyword arguments for compatibility.

 {}
References

Naeem et al., Reliable Fidelity and Diversity Metrics for Generative Models (2020). https://arxiv.org/abs/2002.09797

Code was adapted from: generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models. https://github.com/clovaai/generative-evaluation-prdc

Examples:

>>> density = Density()
>>> real_features = np.random.rand(100, 100)
>>> fake_features = np.random.rand(100, 100)
>>> result: MetricResult = density.compute(real_features, fake_features)
Source code in src/pymdma/general/measures/prdc.py 
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
class Density(FeatureMetric):
    """Density Metric for accessing fidelity of the generated samples. Unlike
    Improved Precision, it is robust towards outliers in the real/reference
    data.

    **Objective**: Fidelity

    Parameters
    ----------
    k : int, optional
        Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.
    metric : str, optional, default="euclidean"
        The metric to use when calculating distance between instances.
        For the available metrics, see the documentation of `sklearn.metrics.pairwise_distances`.
    n_workers : int, optional
        Number of workers for computing pairwise distances. Defaults to 4.
    **kwargs
        Additional keyword arguments for compatibility.

    References
    ----------
    Naeem et al., Reliable Fidelity and Diversity Metrics for Generative Models (2020).
    https://arxiv.org/abs/2002.09797

    Code was adapted from:
    generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models.
    https://github.com/clovaai/generative-evaluation-prdc

    Examples
    --------
    >>> density = Density()
    >>> real_features = np.random.rand(100, 100)
    >>> fake_features = np.random.rand(100, 100)
    >>> result: MetricResult = density.compute(real_features, fake_features)
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = [EvaluationLevel.INSTANCE, EvaluationLevel.DATASET]
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        k: int = 5,
        metric: str = "euclidean",
        n_workers: int = 4,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.k = k
        self.metric = metric
        self.n_workers = n_workers

    def compute(self, real_features: np.ndarray, fake_features: np.ndarray, **kwargs) -> MetricResult:
        """Compute the Density metric.

        Parameters
        ----------
        real_features : np.ndarray
            Array of shape (n_samples, n_features) containing the real features.
        fake_features : np.ndarray
            Array of shape (n_samples, n_features) containing the fake features.

        Notes
        -----
        Intermediate computations can be stored in the `context` dictionary of the `kwargs` parameter.
        Usefull when calculating multiple metrics that share the same intermediate computations.

        Returns
        -------
        result: MetricResult
            Dataset-level and instance-level results for the density metric.
        """
        state = kwargs.get("context", {})
        if "real_nn_distances" not in state:
            state["real_nn_distances"] = compute_nearest_neighbour_distances(
                real_features,
                nearest_k=self.k,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        if "real_fake_distances" not in state:
            state["real_fake_distances"] = compute_pairwise_distance(
                real_features,
                fake_features,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        density = np.logical_or(
            (state["real_fake_distances"] < np.expand_dims(state["real_nn_distances"], axis=1)),
            np.isclose(state["real_fake_distances"], np.expand_dims(state["real_nn_distances"], axis=1)),
        )
        density = (1.0 / float(self.k)) * density.sum(axis=0)

        return MetricResult(
            dataset_level={"dtype": OutputsTypes.NUMERIC, "subtype": "float", "value": density.mean()},
            instance_level={"dtype": OutputsTypes.ARRAY, "subtype": "float", "value": density.tolist()},
        )

pymdma.tabular.measures.synthesis_val.Coverage

Coverage Metric for accessing diversity of the generated samples. Unlike Improved Recall, it is robust towards outliers in the real/reference data.

Objective: Diversity

Parameters:

Name Type Description Default
k int

Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.

 5
metric str

The metric to use when calculating distance between instances. For the available metrics, see the documentation of sklearn.metrics.pairwise_distances.

 "euclidean"
n_workers int

Number of workers for computing pairwise distances. Defaults to 4.

 4
**kwargs

Additional keyword arguments for compatibility.

 {}
References

Naeem et al., Reliable Fidelity and Diversity Metrics for Generative Models (2020). https://arxiv.org/abs/2002.09797

Code was adapted from: generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models. https://github.com/clovaai/generative-evaluation-prdc

Examples:

>>> coverage = Coverage()
>>> real_features = np.random.rand(100, 100)
>>> fake_features = np.random.rand(100, 100)
>>> result: MetricResult = coverage.compute(real_features, fake_features)
Source code in src/pymdma/general/measures/prdc.py 
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
class Coverage(FeatureMetric):
    """Coverage Metric for accessing diversity of the generated samples. Unlike
    Improved Recall, it is robust towards outliers in the real/reference data.

    **Objective**: Diversity

    Parameters
    ----------
    k : int, optional
        Number of nearest neighbors to consider in the hypersphere estimation. Defaults to 5.
    metric : str, optional, default="euclidean"
        The metric to use when calculating distance between instances.
        For the available metrics, see the documentation of `sklearn.metrics.pairwise_distances`.
    n_workers : int, optional
        Number of workers for computing pairwise distances. Defaults to 4.
    **kwargs
        Additional keyword arguments for compatibility.

    References
    ----------
    Naeem et al., Reliable Fidelity and Diversity Metrics for Generative Models (2020).
    https://arxiv.org/abs/2002.09797

    Code was adapted from:
    generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models.
    https://github.com/clovaai/generative-evaluation-prdc

    Examples
    --------
    >>> coverage = Coverage()
    >>> real_features = np.random.rand(100, 100)
    >>> fake_features = np.random.rand(100, 100)
    >>> result: MetricResult = coverage.compute(real_features, fake_features)
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = [EvaluationLevel.INSTANCE, EvaluationLevel.DATASET]
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        k: int = 5,
        metric: str = "euclidean",
        n_workers: int = 4,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.k = k
        self.metric = metric
        self.n_workers = n_workers

    def compute(self, real_features: np.ndarray, fake_features: np.ndarray, **kwargs) -> MetricResult:
        """Compute the Coverage metric.

        Parameters
        ----------
        real_features : np.ndarray
            Array of shape (n_samples, n_features) containing the real features.
        fake_features : np.ndarray
            Array of shape (n_samples, n_features) containing the fake features.

        Notes
        -----
        Intermediate computations can be stored in the `context` dictionary of the `kwargs` parameter.
        Usefull when calculating multiple metrics that share the same intermediate computations.

        Returns
        -------
        result: MetricResult
            Dataset-level and instance-level results for the coverage metric.
        """
        state = kwargs.get("context", {})
        if "real_nn_distances" not in state:
            state["real_nn_distances"] = compute_nearest_neighbour_distances(
                real_features,
                nearest_k=self.k,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        if "real_fake_distances" not in state:
            state["real_fake_distances"] = compute_pairwise_distance(
                real_features,
                fake_features,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        coverage = np.logical_or(
            state["real_fake_distances"].min(axis=1) < state["real_nn_distances"],
            np.isclose(state["real_fake_distances"].min(axis=1), state["real_nn_distances"]),
        )

        # matrix with (R, F) shape -> .any() -> matrix with (F,) shape
        # an array that indicates for each F in how many real manifolds it is contained in
        coverage_counts = np.logical_or(
            state["real_fake_distances"] < np.expand_dims(state["real_nn_distances"], axis=1),
            np.isclose(state["real_fake_distances"], np.expand_dims(state["real_nn_distances"], axis=1)),
        ).sum(axis=0)

        return MetricResult(
            dataset_level={"dtype": OutputsTypes.NUMERIC, "subtype": "float", "value": coverage.mean()},
            instance_level={"dtype": OutputsTypes.ARRAY, "subtype": "int", "value": coverage_counts.tolist()},
        )

pymdma.tabular.measures.synthesis_val.StatisticalSimScore

Computes a dataset-level statistical similarity score between real and synthetic data.

This metric assesses how closely the statistical properties of the synthetic dataset resemble those of the real dataset, providing a fidelity measure for synthetic data generation.

Objective: Fidelity

Parameters:

Name Type Description Default
col_map dict

A mapping of column names to their types and properties. This is used to determine how to compute similarity for each column.

 None
**kwargs dict

Additional keyword arguments passed to the parent class.

 {}
References

Yang et al., Structured evaluation of synthetic tabular data (2024). https://arxiv.org/abs/2403.10424

Returns:

Type Description
MetricResult

A MetricResult object containing the similarity scores and their statistics.

Examples:

>>> # Example 1: Evaluating statistical similarity for a dataset with discrete and continuous variables
>>> import numpy as np
>>> real_data = np.array([
...     [1, 2.5],
...     [1, 3.0],
...     [2, 3.5],
...     [2, 4.0]
... ])
>>> syn_data = np.array([
...     [1, 2.6],
...     [1, 3.1],
...     [2, 3.4],
...     [2, 4.2]
... ])
>>> col_map = {
...     "column1": {"type": {"tag": "discrete"}},
...     "column2": {"type": {"tag": "continuous"}},
... }
>>> sim_score = StatisticalSimScore(col_map=col_map)
>>> result: MetricResult = sim_score.compute(real_data, syn_data)
>>> dataset_level, _ = result.value # Output: similarity scores for each column

>>> # Example 2: Evaluating similarity with mismatched column types
>>> real_data = np.array([
...     [1, 2],
...     [2, 3],
...     [3, 4]
... ])
>>> syn_data = np.array([
...     [1, 2],
...     [2, 3],
...     [3, 5]
... ])
>>> col_map = {
...     "column1": {"type": {"tag": "discrete"}},
...     "column2": {"type": {"tag": "discrete"}},
... }
>>> sim_score = StatisticalSimScore(col_map=col_map)
>>> result: MetricResult = sim_score.compute(real_data, syn_data)
>>> dataset_level, _ = result.value  # Output: similarity scores for each column
>>> dataset_stats, _ = result.stats  # Output: mean and std of similarity scores
Source code in src/pymdma/tabular/measures/synthesis_val/data/similarity.py 
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
class StatisticalSimScore(Metric):
    """Computes a dataset-level statistical similarity score between real and
    synthetic data.

    This metric assesses how closely the statistical properties of the synthetic dataset
    resemble those of the real dataset, providing a fidelity measure for synthetic data generation.

    **Objective**: Fidelity

    Parameters
    ----------
    col_map : dict, optional, default=None
        A mapping of column names to their types and properties. This is used to determine
        how to compute similarity for each column.
    **kwargs : dict
        Additional keyword arguments passed to the parent class.

    References
    ----------
    Yang et al., Structured evaluation of synthetic tabular data (2024).
    https://arxiv.org/abs/2403.10424

    Returns
    -------
    MetricResult
        A MetricResult object containing the similarity scores and their statistics.

    Examples
    --------
    >>> # Example 1: Evaluating statistical similarity for a dataset with discrete and continuous variables
    >>> import numpy as np
    >>> real_data = np.array([
    ...     [1, 2.5],
    ...     [1, 3.0],
    ...     [2, 3.5],
    ...     [2, 4.0]
    ... ])
    >>> syn_data = np.array([
    ...     [1, 2.6],
    ...     [1, 3.1],
    ...     [2, 3.4],
    ...     [2, 4.2]
    ... ])
    >>> col_map = {
    ...     "column1": {"type": {"tag": "discrete"}},
    ...     "column2": {"type": {"tag": "continuous"}},
    ... }
    >>> sim_score = StatisticalSimScore(col_map=col_map)
    >>> result: MetricResult = sim_score.compute(real_data, syn_data)
    >>> dataset_level, _ = result.value # Output: similarity scores for each column

    >>> # Example 2: Evaluating similarity with mismatched column types
    >>> real_data = np.array([
    ...     [1, 2],
    ...     [2, 3],
    ...     [3, 4]
    ... ])
    >>> syn_data = np.array([
    ...     [1, 2],
    ...     [2, 3],
    ...     [3, 5]
    ... ])
    >>> col_map = {
    ...     "column1": {"type": {"tag": "discrete"}},
    ...     "column2": {"type": {"tag": "discrete"}},
    ... }
    >>> sim_score = StatisticalSimScore(col_map=col_map)
    >>> result: MetricResult = sim_score.compute(real_data, syn_data)
    >>> dataset_level, _ = result.value  # Output: similarity scores for each column
    >>> dataset_stats, _ = result.stats  # Output: mean and std of similarity scores
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = EvaluationLevel.DATASET
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        col_map: Optional[Dict[str, Dict[str, str]]] = None,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.col_map = col_map

    def _stat_sim_1d(
        self,
        real_col: np.ndarray,
        syn_col: np.ndarray,
        kind: str = "discrete",
        **kwargs,
    ):
        """Computes an 1D statistical similarity.

        Parameters
        ----------
        real_col : np.ndarray
            The real data for the specific attribute.
        syn_col : np.ndarray
            The synthetic data for the specific attribute.
        kind : str
            The type of data ('discrete' or 'continuous').
        **kwargs : dict
            Additional keyword arguments for computation.

        Returns
        -------
        float
            The computed similarity score for the attribute.
        """
        # variable type assignment
        kind_ = kind.lower() if kind.lower() in ["discrete", "continuous"] else "continuous"

        # mapper
        kind_mapper = {
            "discrete": _get_tv_similarity,
            "continuous": _get_ks_similarity,
        }

        # score
        sim_score = kind_mapper.get(kind_)(
            real_col,
            syn_col,
        )
        return sim_score

    def compute(self, real_data: np.ndarray, syn_data: np.ndarray, **kwargs) -> MetricResult:
        """Computes the statistical similarity score between real and synthetic
        datasets.

        Parameters
        ----------
        real_data : np.ndarray
            The real dataset for comparison.
        syn_data : np.ndarray
            The synthetic dataset to evaluate.
        **kwargs : dict
            Additional keyword arguments for computation.

        Returns
        -------
        MetricResult
            A MetricResult object containing the similarity scores and their statistics.
        """

        # checkpoint
        assert real_data.shape[1] == syn_data.shape[1], "Mismatched columns. Please fix before computing metrics."

        # column map
        col_map_exists = isinstance(self.col_map, dict)
        cols = self.col_map.keys() if col_map_exists else [f"att_{idx}" for idx in range(real_data.shape[1])]

        # similarity map
        sim_score = {}

        # column similarity
        for idx, col in enumerate(cols):
            # continuous OR discrete
            if col_map_exists:
                # dtype
                vtype = self.col_map.get(col).get("type").get("tag")
            else:
                # dtype
                vtype = "discrete" if is_categorical(real_data[:, idx]) else "continuous"

            # compute similarity
            sim_ = self._stat_sim_1d(
                real_data[:, idx],
                syn_data[:, idx],
                kind=vtype,
            )

            # assign
            sim_score[col] = sim_

        # global scores
        global_d = {
            "mean": np.mean(list(sim_score.values())).round(2),
            "std": np.std(list(sim_score.values())).round(2),
        }

        return MetricResult(
            dataset_level={
                "dtype": OutputsTypes.KEY_VAL,
                "subtype": "float",
                "value": sim_score,
                "stats": global_d,
            },
        )

pymdma.tabular.measures.synthesis_val.StatisticalDivergenceScore

Computes a statistical divergence score for each column, specifically the Jensen-Shannon (JS) and Kullback-Leibler (KL) divergence scores.

Objective: Fidelity

Parameters:

Name Type Description Default
column_names list of str

List of the names of the columns (features) in the dataset.

 None
score str

Specifies the divergence score to compute ('js' for Jensen-Shannon, 'kl' for Kullback-Leibler, 'all' for both). By default, it is set to 'kl'.

 'kl'
**kwargs dict

Additional keyword arguments passed to the parent class.

 {}
References

Fonseca and Bacao, Tabular and latent space synthetic data generation: a literature review (2023). https://doi.org/10.1186/s40537-023-00792-7

Returns:

Type Description
MetricResult

A MetricResult object containing the divergence scores and their statistics.

Examples:

>>> # Example 1: Evaluating statistical divergence for a dataset
>>> import numpy as np
>>> real_data = np.array([
...     [1, 2, 3],
...     [2, 3, 4],
...     [3, 4, 5]
... ])
>>> syn_data = np.array([
...     [1, 2, 2],
...     [2, 2, 3],
...     [3, 3, 4]
... ])
>>> col_map = {
...     "column1": {"type": {"tag": "continuous"}},
...     "column2": {"type": {"tag": "continuous"}},
...     "column3": {"type": {"tag": "continuous"}},
... }
>>> divergence_score = StatisticalDivergenceScore(col_map=col_map, score='kl')
>>> result: MetricResult = divergence_score.compute(real_data, syn_data)
>>> dataset_level, _ = result.value  # Output: divergence scores for each column

>>> # Example 2: Using JS divergence instead of KL
>>> divergence_score_js = StatisticalDivergenceScore(col_map=col_map, score='js')
>>> result_js: MetricResult = divergence_score_js.compute(real_data, syn_data)
>>> dataset_level_js, _ = result_js.value  # Output: JS divergence scores for each column
>>> dataset_stats_js, _ = result_js.stats  # Output: mean and std of JS divergence scores
Source code in src/pymdma/tabular/measures/synthesis_val/data/similarity.py 
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
class StatisticalDivergenceScore(Metric):
    """Computes a statistical divergence score for each column, specifically
    the Jensen-Shannon (JS) and Kullback-Leibler (KL) divergence scores.

    **Objective**: Fidelity

    Parameters
    ----------
    column_names : list of str, optional, default=None
        List of the names of the columns (features) in the dataset.
    score : str, optional, default='kl'
        Specifies the divergence score to compute ('js' for Jensen-Shannon, 'kl' for Kullback-Leibler, 'all' for both).
        By default, it is set to 'kl'.
    **kwargs : dict
        Additional keyword arguments passed to the parent class.

    References
    ----------
    Fonseca and Bacao,  Tabular and latent space synthetic data generation: a literature review (2023).
    https://doi.org/10.1186/s40537-023-00792-7

    Returns
    -------
    MetricResult
        A MetricResult object containing the divergence scores and their statistics.

    Examples
    --------
    >>> # Example 1: Evaluating statistical divergence for a dataset
    >>> import numpy as np
    >>> real_data = np.array([
    ...     [1, 2, 3],
    ...     [2, 3, 4],
    ...     [3, 4, 5]
    ... ])
    >>> syn_data = np.array([
    ...     [1, 2, 2],
    ...     [2, 2, 3],
    ...     [3, 3, 4]
    ... ])
    >>> col_map = {
    ...     "column1": {"type": {"tag": "continuous"}},
    ...     "column2": {"type": {"tag": "continuous"}},
    ...     "column3": {"type": {"tag": "continuous"}},
    ... }
    >>> divergence_score = StatisticalDivergenceScore(col_map=col_map, score='kl')
    >>> result: MetricResult = divergence_score.compute(real_data, syn_data)
    >>> dataset_level, _ = result.value  # Output: divergence scores for each column

    >>> # Example 2: Using JS divergence instead of KL
    >>> divergence_score_js = StatisticalDivergenceScore(col_map=col_map, score='js')
    >>> result_js: MetricResult = divergence_score_js.compute(real_data, syn_data)
    >>> dataset_level_js, _ = result_js.value  # Output: JS divergence scores for each column
    >>> dataset_stats_js, _ = result_js.stats  # Output: mean and std of JS divergence scores
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = EvaluationLevel.DATASET
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = False
    min_value: float = -np.inf
    max_value: float = np.inf

    def __init__(
        self,
        column_names: Optional[List[str]] = None,
        score: Literal["js", "kl", "all"] = "kl",
        **kwargs,
    ):
        """Initializes the StatisticalDivergenceScore metric to evaluate the
        divergence between real and synthetic datasets based on defined column
        characteristics.

        Parameters
        ----------
        column_names : list of str, optional, default=None
            List of the names of the columns (features) in the dataset.
        score : str, optional, default='kl'
            Specifies the divergence score to compute ('js' for Jensen-Shannon,
            'kl' for Kullback-Leibler). Default is 'kl'.
        **kwargs : dict
            Additional keyword arguments passed to the parent class.
        """

        super().__init__(**kwargs)

        # column names
        self.column_names = column_names

        # score type
        self.score = score

    def _diverg_score_1d(
        self,
        real_col: np.ndarray,
        syn_col: np.ndarray,
        score: Literal["js", "kl", "all"] = "all",
        **kwargs,
    ):
        """Computes a column-level statistical divergence.

        Parameters
        ----------
        real_col : np.ndarray
            The real data for the specific attribute.
        syn_col : np.ndarray
            The synthetic data for the specific attribute.
        score : str
            The type of divergence to compute ('js' or 'kl').
        **kwargs : dict
            Additional keyword arguments for computation.

        Returns
        -------
        dict
            A dictionary containing the computed divergence scores.
        """

        # score map
        score_map = {"js": _get_js_divergence, "kl": _get_kl_divergence}

        # score tags
        score_tag = ["js", "kl"] if score.lower() == "all" else [score]

        # get probability distributions
        real_pdf, syn_pdf, bins = _get_nn_pdf(real_col, syn_col)

        # compute divergence scores
        div_score = {tag: score_map.get(tag)(real_pdf, syn_pdf) for tag in score_tag if tag in score_map.keys()}

        return div_score

    def compute(self, real_data: np.ndarray, syn_data: np.ndarray, **kwargs) -> MetricResult:
        """Computes the statistical divergence score between real and synthetic
        datasets.

        Parameters
        ----------
        real_data : np.ndarray
            The real dataset for comparison.
        syn_data : np.ndarray
            The synthetic dataset to evaluate.
        **kwargs : dict
            Additional keyword arguments for computation.

        Returns
        -------
        MetricResult
            A MetricResult object containing the divergence scores and their statistics.
        """

        # checkpoint
        assert real_data.shape[1] == syn_data.shape[1], "Mismatched columns. Please fix before computing metrics."

        # columns
        cols = (
            self.column_names
            if isinstance(self.column_names, list)
            else [f"att_{idx}" for idx in range(real_data.shape[1])]
        )

        # divergence map
        div_score = {}

        # column-wise
        for idx, col in enumerate(cols):
            # compute scores
            sim_ = self._diverg_score_1d(
                real_data[:, idx],
                syn_data[:, idx],
                score=self.score,
            )

            # assign
            div_score[col] = list(sim_.values())

        # global scores
        # auxiliary score dataframe
        aux_df = pd.DataFrame.from_dict(div_score.values())

        # aggregates
        mean_g, std_g = aux_df.mean(0).to_dict(), aux_df.std(0).to_dict()

        # global
        glob_d = {
            f"{self.score}_mean": mean_g[0],
            f"{self.score}_std": std_g[0],
        }

        return MetricResult(
            dataset_level={
                "dtype": OutputsTypes.KEY_ARRAY,
                "subtype": "float",
                "value": div_score,
                "stats": glob_d,
            },
        )

pymdma.tabular.measures.synthesis_val.CoherenceScore

Computes the coherence score between the correlation matrices of the target and synthetic datasets. A higher coherence score indicates better fidelity between the datasets in terms of their correlation structures.

Objective: Fidelity

Parameters:

Name Type Description Default
weights ndarray

Weights for the correlations, allowing for weighted contributions to the coherence score. If None, uniform weights are applied.

 None
corr_type str

The type of correlation to compute ('pearson' by default). Other types like 'spearman' may be supported depending on the implementation.

 'pearson'
**kwargs dict

Additional keyword arguments passed to the parent class.

 {}
References

Yang et al., Structured evaluation of synthetic tabular data (2024). https://arxiv.org/abs/2403.10424

Returns:

Type Description
MetricResult

A MetricResult object containing the coherence score.

Examples:

>>> # Example 1: Evaluating coherence score for a dataset
>>> import numpy as np
>>> real_data = np.array([
...     [1, 2, 3],
...     [2, 3, 4],
...     [3, 4, 5]
... ])
>>> syn_data = np.array([
...     [1, 2, 3],
...     [1, 2, 3],
...     [3, 4, 5]
... ])
>>> coherence_score = CoherenceScore(corr_type='pearson')
>>> result: MetricResult = coherence_score.compute(real_data, syn_data)
>>> dataset_level, _ = result.value  # Output: coherence score

>>> # Example 2: Evaluating with custom weights
>>> weights = np.array([0.5, 1.0, 1.5])  # Example weights
>>> coherence_score_weighted = CoherenceScore(weights=weights, corr_type='spearman')
>>> result_weighted: MetricResult = coherence_score_weighted.compute(real_data, syn_data)
>>> dataset_level, _ = result_weighted.value  # Output: weighted coherence score
Source code in src/pymdma/tabular/measures/synthesis_val/data/similarity.py 
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
class CoherenceScore(Metric):
    """Computes the coherence score between the correlation matrices of the
    target and synthetic datasets. A higher coherence score indicates better
    fidelity between the datasets in terms of their correlation structures.

    **Objective**: Fidelity

    Parameters
    ----------
    weights : np.ndarray, optional, default=None
        Weights for the correlations, allowing for weighted contributions
        to the coherence score. If None, uniform weights are applied.
    corr_type : str, optional, default='pearson'
        The type of correlation to compute ('pearson' by default).
        Other types like 'spearman' may be supported depending on the implementation.
    **kwargs : dict
        Additional keyword arguments passed to the parent class.

    References
    ----------
    Yang et al., Structured evaluation of synthetic tabular data (2024).
    https://arxiv.org/abs/2403.10424

    Returns
    -------
    MetricResult
        A MetricResult object containing the coherence score.

    Examples
    --------
    >>> # Example 1: Evaluating coherence score for a dataset
    >>> import numpy as np
    >>> real_data = np.array([
    ...     [1, 2, 3],
    ...     [2, 3, 4],
    ...     [3, 4, 5]
    ... ])
    >>> syn_data = np.array([
    ...     [1, 2, 3],
    ...     [1, 2, 3],
    ...     [3, 4, 5]
    ... ])
    >>> coherence_score = CoherenceScore(corr_type='pearson')
    >>> result: MetricResult = coherence_score.compute(real_data, syn_data)
    >>> dataset_level, _ = result.value  # Output: coherence score

    >>> # Example 2: Evaluating with custom weights
    >>> weights = np.array([0.5, 1.0, 1.5])  # Example weights
    >>> coherence_score_weighted = CoherenceScore(weights=weights, corr_type='spearman')
    >>> result_weighted: MetricResult = coherence_score_weighted.compute(real_data, syn_data)
    >>> dataset_level, _ = result_weighted.value  # Output: weighted coherence score
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = EvaluationLevel.DATASET
    metric_group = MetricGroup.QUALITY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        weights: Optional[np.ndarray] = None,
        corr_type: Optional[str] = "pearson",
        **kwargs,
    ):
        """Initializes the CoherenceScore metric to evaluate the coherence
        between the correlation matrices of real and synthetic datasets.

        Parameters
        ----------
        weights : np.ndarray, optional, default=None
            Weights for the correlations, allowing for weighted contributions
            to the coherence score. If None, uniform weights are applied.
        corr_type : str, optional, default=None
            The type of correlation to compute ('pearson' by default).
        **kwargs : dict
            Additional keyword arguments passed to the parent class.
        """

        super().__init__(**kwargs)

        # correlation type
        self.corr = corr_type

        # weights array
        self.weights = weights

    def compute(
        self,
        real_data: np.ndarray,
        syn_data: np.ndarray,
        **kwargs,
    ) -> MetricResult:
        """Computes the coherence score between the correlation matrices of
        real and synthetic datasets.

        Parameters
        ----------
        real_data : np.ndarray
            The real dataset for comparison.
        syn_data : np.ndarray
            The synthetic dataset to evaluate.
        **kwargs : dict
            Additional keyword arguments for computation.

        Returns
        -------
        MetricResult
            A MetricResult object containing the coherence score.
        """

        # compute correlation matrices
        real_corr = pd.DataFrame(real_data).corr(self.corr).replace(np.nan, 1).to_numpy()
        syn_corr = pd.DataFrame(syn_data).corr(self.corr).replace(np.nan, 1).to_numpy()

        # number columns
        n_cols = len(real_corr)

        # compute similarity between real and syn matrices
        delta_corr = np.abs(
            np.nan_to_num(real_corr) - np.nan_to_num(syn_corr),
        )  # differences

        # weight matrix
        id_mask = np.abs(np.identity(n_cols) - 1)

        if self.weights is not None:
            w_mask = np.array([self.weights] * n_cols) * id_mask
        else:
            w_mask = np.ones((n_cols, n_cols)) * id_mask

        # norm
        w_mask /= sum(w_mask)

        # correlation similarity (weighted avg.)
        corr_sim = np.sum(delta_corr * w_mask) / np.sum(w_mask)
        # ((n_cols * (n_cols - 1)))

        # average correlation
        avg_corr_coh = np.mean(np.round(1 - corr_sim / 2, 3))

        return MetricResult(
            dataset_level={
                "dtype": OutputsTypes.NUMERIC,
                "subtype": "float",
                "value": avg_corr_coh,
            },
        )

Privacy

pymdma.tabular.measures.synthesis_val.Authenticity

Authenticity Metric for assessing the authenticity of the generated samples. A synthetic sample is considered authentic if it is signficantly distinct from any real sample.

Objective: Privacy

Parameters:

Name Type Description Default
metric str

The metric to use when calculating distance between instances. For the available metrics, see the documentation of sklearn.metrics.pairwise_distances.

 "euclidean"
**kwargs

Additional keyword arguments for compatibility.

 {}
Notes

The authenticity metric is computed by checking if any fake sample is closer to a real sample than the real sample is to any other real sample.

References

Alaa et al., How Faithful Is Your Synthetic Data? Sample-Level Metrics for Evaluating and Auditing Generative Models. (2022) https://doi.org/10.48550/arXiv.2102.08921.

Hypersphere estimation code was adapted from: generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models. https://github.com/clovaai/generative-evaluation-prdc

Examples:

>>> authenticity = Authenticity()
>>> real_features = np.random.rand(100, 100)
>>> fake_features = np.random.rand(100, 100)
>>> result: MetricResult = authenticity.compute(real_features, fake_features)
Source code in src/pymdma/general/measures/prdc.py 
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
class Authenticity(FeatureMetric):
    """Authenticity Metric for assessing the authenticity of the generated
    samples. A synthetic sample is considered authentic if it is signficantly
    distinct from any real sample.

    **Objective**: Privacy

    Parameters
    ----------
    metric : str, optional, default="euclidean"
        The metric to use when calculating distance between instances.
        For the available metrics, see the documentation of `sklearn.metrics.pairwise_distances`.
    **kwargs
        Additional keyword arguments for compatibility.

    Notes
    -----
    The authenticity metric is computed by checking if any fake sample is closer to a real sample than the real sample is to any other real sample.

    References
    ----------
    Alaa et al., How Faithful Is Your Synthetic Data? Sample-Level Metrics for Evaluating and Auditing Generative Models. (2022)
    https://doi.org/10.48550/arXiv.2102.08921.

    Hypersphere estimation code was adapted from:
    generative-evaluation-prdc, Reliable Fidelity and Diversity Metrics for Generative Models.
    https://github.com/clovaai/generative-evaluation-prdc

    Examples
    --------
    >>> authenticity = Authenticity()
    >>> real_features = np.random.rand(100, 100)
    >>> fake_features = np.random.rand(100, 100)
    >>> result: MetricResult = authenticity.compute(real_features, fake_features)
    """

    reference_type = ReferenceType.DATASET
    evaluation_level = [EvaluationLevel.INSTANCE, EvaluationLevel.DATASET]
    metric_group = MetricGroup.PRIVACY

    higher_is_better: bool = True
    min_value: float = 0.0
    max_value: float = 1.0

    def __init__(
        self,
        metric: str = "euclidean",
        n_workers: int = 4,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.metric = metric
        self.n_workers = n_workers

    def compute(self, real_features: np.ndarray, fake_features: np.ndarray, **kwargs) -> MetricResult:
        """Compute the Authenticity metric.

        Parameters
        ----------
        real_features : np.ndarray
            Array of shape (n_samples, n_features) containing the real features.
        fake_features : np.ndarray
            Array of shape (n_samples, n_features) containing the fake features.

        Notes
        -----
        Intermediate computations can be stored in the `context` dictionary of the `kwargs` parameter.
        Usefull when calculating multiple metrics that share the same intermediate computations.

        Returns
        -------
        result: MetricResult
            Dataset-level and instance-level results for the authenticity metric
        """
        state = kwargs.get("context", {})

        if "real_fake_distances" not in state:
            state["real_fake_distances"] = compute_pairwise_distance(
                real_features,
                fake_features,
                metric=self.metric,
                n_workers=self.n_workers,
            )

        # compute distance to closest real samples
        state["real_closest_real_distances"] = compute_nearest_neighbour_distances(
            real_features,
            nearest_k=1,
            metric=self.metric,
            n_workers=self.n_workers,
        )

        # check if any fake sample is closer to Ri than Ri is to any other Rj
        authenticity = np.logical_or(
            state["real_fake_distances"] < np.expand_dims(state["real_closest_real_distances"], axis=1),
            np.isclose(state["real_fake_distances"], np.expand_dims(state["real_closest_real_distances"], axis=1)),
        )

        # mask of the values that are considered authentic in the fake dataset
        authenticity_mask = ~authenticity.any(axis=0)

        return MetricResult(
            dataset_level={"dtype": OutputsTypes.NUMERIC, "subtype": "float", "value": authenticity_mask.mean()},
            instance_level={
                "dtype": OutputsTypes.ARRAY,
                "subtype": "int",
                "value": authenticity_mask.astype(int).tolist(),
            },
        )

Data-based

Utility