def _calc_all_intracls_dists(
cls,
N: np.ndarray,
y: np.ndarray,
dist_metric: str = "euclidean",
get_max_dist: bool = True,
cls_inds: t.Optional[np.ndarray] = None,
classes: t.Optional[np.ndarray] = None,
) -> np.ndarray:
"""Calculate all intraclass (internal to a class) distances."""
if cls_inds is None:
if classes is None:
classes = np.unique(y)
cls_inds = _utils.calc_cls_inds(y=y, classes=classes)
intracls_dists = np.array(
[
cls._calc_intracls_dists(
N[cur_class, :],
dist_metric=dist_metric,
get_max_dist=get_max_dist,
) for cur_class in cls_inds
],
dtype=object,
)
return intracls_dists
def _calc_pwise_norm_intercls_dist(
cls,
N: np.ndarray,
y: np.ndarray,
dist_metric: str = "euclidean",
classes: t.Optional[np.ndarray] = None,
cls_inds: t.Optional[np.ndarray] = None,
) -> t.List[np.ndarray]:
"""Calculate all pairwise normalized interclass distances."""
if cls_inds is None:
if classes is None:
classes = np.unique(y)
cls_inds = _utils.calc_cls_inds(y=y, classes=classes)
intercls_dists = [
cls._calc_normalized_intercls_dist(
N[cls_inds[id_cls_a, :], :],
N[cls_inds[id_cls_b, :], :],
dist_metric=dist_metric,
) for id_cls_a, id_cls_b in itertools.combinations(
np.arange(cls_inds.shape[0]), 2)
]
return intercls_dists
def precompute_complexity(cls,
y: t.Optional[np.ndarray] = None,
**kwargs) -> t.Dict[str, t.Any]:
"""Precompute some useful things to support feature-based measures.
Parameters
----------
y : :obj:`np.ndarray`, optional
Target attribute.
**kwargs
Additional arguments. May have previously precomputed before this
method from other precomputed methods, so they can help speed up
this precomputation.
Returns
-------
:obj:`dict`
With following precomputed items:
- ``ovo_comb`` (list): List of all class OVO combination,
i.e., all combinations of distinct class indices by pairs
([(0, 1), (0, 2) ...].)
- ``cls_inds`` (:obj:`np.ndarray`): Boolean array which
indicates whether each example belongs to each class. The
rows represents the distinct classes, and the instances
are represented by the columns.
- ``classes`` (:obj:`np.ndarray`): distinct classes in the
fitted target attribute.
- ``class_freqs`` (:obj:`np.ndarray`): The number of examples
in each class. The indices represent the classes.
"""
precomp_vals = {} # type: t.Dict[str, t.Any]
if (y is not None and not {"classes", "class_freqs"}.issubset(kwargs)):
sub_dic = MFEGeneral.precompute_general_class(y)
precomp_vals.update(sub_dic)
classes = kwargs.get("classes", precomp_vals.get("classes"))
if y is not None and "cls_inds" not in kwargs:
cls_inds = _utils.calc_cls_inds(y, classes)
precomp_vals["cls_inds"] = cls_inds
if y is not None and "ovo_comb" not in kwargs:
ovo_comb = cls._calc_ovo_comb(classes)
precomp_vals["ovo_comb"] = ovo_comb
return precomp_vals
def precompute_clustering_class(cls,
y: t.Optional[np.ndarray] = None,
**kwargs) -> t.Dict[str, t.Any]:
"""Precompute distinct classes and its frequencies from ``y``.
Parameters
----------
y : :obj:`np.ndarray`, optional
Instance cluster index (or target attribute).
**kwargs
Additional arguments. May have previously precomputed before
this method from other precomputed methods, so they can help
speed up this precomputation.
Returns
-------
:obj:`dict`
The following precomputed items are returned:
* ``classes`` (:obj:`np.ndarray`): distinct classes of
``y``, if ``y`` is not :obj:`NoneType`.
* ``class_freqs`` (:obj:`np.ndarray`): class frequencies of
``y``, if ``y`` is not :obj:`NoneType`.
* ``cls_inds`` (:obj:`np.ndarray`): Boolean array which
indicates whether each example belongs to each class. The
rows represents the distinct classes, and the instances
are represented by the columns.
"""
precomp_vals = {}
if y is not None and not {"classes", "class_freqs"}.issubset(kwargs):
classes, class_freqs = np.unique(y, return_counts=True)
precomp_vals["classes"] = classes
precomp_vals["class_freqs"] = class_freqs
classes = kwargs.get("classes", precomp_vals.get("classes"))
if y is not None and "cls_inds" not in kwargs:
cls_inds = _utils.calc_cls_inds(y, classes)
precomp_vals["cls_inds"] = cls_inds
return precomp_vals
def _get_class_representatives(
cls,
N: np.ndarray,
y: np.ndarray,
representative: t.Union[t.Sequence, np.ndarray, str] = "mean",
cls_inds: t.Optional[np.ndarray] = None,
classes: t.Optional[np.ndarray] = None) -> np.ndarray:
"""Get a representative instance for each distinct class.
If ``representative`` argument is a string, then it must be
some statistical method to be aplied in the attributes of
instances of the same class in ``N`` to construct the class
representative instance (currently supported only ``mean`` and
``median``). If ``representative`` is a sequence, then its
shape must be (number_of_classes, number_of_attributes) (i.e.,
there must have one class representative for each distinct class,
and every class representative must have the same dimension of
the instances in ``N``.)
"""
if classes is None:
classes = np.unique(y)
if isinstance(representative, str):
center_method = {
"mean": np.mean,
"median": np.median,
}.get(representative)
if center_method is None:
raise ValueError("'representative' must be 'mean' or "
"'median'. Got '{}'.".format(representative))
if cls_inds is None:
cls_inds = _utils.calc_cls_inds(y=y, classes=classes)
representative = [
center_method(N[cur_class, :], axis=0)
for cur_class in cls_inds
]
elif not isinstance(representative,
(collections.Sequence, np.ndarray)):
raise TypeError("'representative' type must be string "
"or a sequence or a numpy array. "
"Got '{}'.".format(type(representative)))
representative_arr = np.asarray(representative)
num_repr, repr_dim = representative_arr.shape
_, num_attr = N.shape
if num_repr != classes.size:
raise ValueError("There must exist one class representative "
"for every distinct class. (Expected '{}', "
"got '{}'".format(classes.size, num_repr))
if repr_dim != num_attr:
raise ValueError("The dimension of each class representative "
"must match the instances dimension. (Expected "
"'{}', got '{}'".format(classes.size, repr_dim))
return representative_arr
def ft_n4(cls,
N: np.ndarray,
y: np.ndarray,
cls_inds: t.Optional[np.ndarray] = None,
metric_n4: str = "minkowski",
p_n4: int = 2,
n_neighbors_n4: int = 1,
random_state: t.Optional[int] = None) -> np.ndarray:
"""Compute the non-linearity of the NN Classifier.
Parameters
----------
N : :obj:`np.ndarray`
Numerical fitted data.
y : :obj:`np.ndarray`
Target attribute.
cls_inds : :obj:`np.ndarray`, optional
Boolean array which indicates the examples of each class.
The rows represents each distinct class, and the columns
represents the instances.
metric_n4 : str, optional (default = "minkowski")
The distance metric used in the internal kNN classifier. See the
documentation of the ``sklearn.neighbors.DistanceMetric`` class
for a list of available metrics.
p_n4 : int, optional (default = 2)
Power parameter for the Minkowski metric. When p = 1, this is
equivalent to using manhattan_distance (l1), and
euclidean_distance (l2) for p = 2. For arbitrary p,
minkowski_distance (l_p) is used. Please, check the
``sklearn.neighbors.KNeighborsClassifier`` documentation for
more information.
random_state : int, optional
If given, set the random seed before computing the randomized
data interpolation.
Returns
-------
:obj:`np.ndarray`
Misclassifications of the NN classifier in the interpolated
dataset.
References
----------
.. [1] Ana C. Lorena, Luís P. F. Garcia, Jens Lehmann, Marcilio C. P.
Souto, and Tin K. Ho. How Complex is your classification problem?
A survey on measuring classification complexity (V2). (2019)
(Cited on page 9-11). Published in ACM Computing Surveys (CSUR),
Volume 52 Issue 5, October 2019, Article No. 107.
"""
if cls_inds is None:
classes = np.unique(y)
cls_inds = _utils.calc_cls_inds(y, classes)
# 0-1 feature scaling
N = sklearn.preprocessing.MinMaxScaler(
feature_range=(0, 1)).fit_transform(N)
if random_state is not None:
np.random.seed(random_state)
N_test = np.zeros(N.shape, dtype=N.dtype)
y_test = np.zeros(y.shape, dtype=y.dtype)
ind_cur = 0
for inds_cur_cls in cls_inds:
N_cur_cls = N[inds_cur_cls, :]
subset_size = N_cur_cls.shape[0]
# Currently it is allowed to a instance 'interpolate with itself',
# which holds the instance itself as result.
sample_a = N_cur_cls[np.random.choice(subset_size, subset_size), :]
sample_b = N_cur_cls[np.random.choice(subset_size, subset_size), :]
rand_delta = np.random.ranf(N_cur_cls.shape)
N_subset_interp = sample_a + (sample_b - sample_a) * rand_delta
ind_next = ind_cur + subset_size
N_test[ind_cur:ind_next, :] = N_subset_interp
y_test[ind_cur:ind_next] = y[inds_cur_cls]
ind_cur = ind_next
knn = sklearn.neighbors.KNeighborsClassifier(
n_neighbors=n_neighbors_n4, p=p_n4, metric=metric_n4).fit(N, y)
y_pred = knn.predict(N_test)
misclassifications = np.not_equal(y_test, y_pred).astype(int)
# The measure is computed in the literature using the mean. However, it
# is formulated here as a meta-feature. Therefore, the post-processing
# should be used to get the mean and other measures as well.
return misclassifications