aisp 0.1.34__py3-none-any.whl → 0.1.40__py3-none-any.whl

aisp/nsa/_negative_selection.py

@@ -0,0 +1,682 @@
+"""Negative Selection Algorithm."""
+
+from typing import Dict, Literal, Optional, Union
+from tqdm import tqdm
+
+import numpy as np
+import numpy.typing as npt
+
+from ._ns_core import (
+    check_detector_bnsa_validity,
+    bnsa_class_prediction,
+    check_detector_rnsa_validity,
+)
+from ..exceptions import MaxDiscardsReachedError
+from ..utils import slice_index_list_by_class
+from ..utils.distance import (
+    min_distance_to_class_vectors,
+    get_metric_code,
+    compute_metric_distance,
+)
+from ..utils.sanitizers import sanitize_seed, sanitize_choice, sanitize_param
+from ._base import BaseNSA, Detector
+
+
+class RNSA(BaseNSA):
+    """
+    The ``RNSA`` (Real-Valued Negative Selection Algorithm) class is for classification and
+    identification purposes. of anomalies through the self and not self method.
+
+    Parameters
+    ----------
+    * N (``int``): Number of detectors. Defaults to ``100``.
+    * r (``float``): Radius of the detector. Defaults to ``0.05``.
+    * r_s (``float``): rₛ Radius of the ``X`` own samples. Defaults to ``0.0001``.
+    * k (``int``): Number of neighbors near the randomly generated detectors to perform the
+        distance average calculation. Defaults to ``1``.
+    * metric (``str``): Way to calculate the distance between the detector and the sample:
+            + ``'Euclidean'`` ➜ The calculation of the distance is given by the expression:
+                √( (x₁ – x₂)² + (y₁ – y₂)² + ... + (yn – yn)²).
+            + ``'minkowski'`` ➜ The calculation of the distance is given by the expression:
+                ( |X₁ – Y₁|p + |X₂ – Y₂|p + ... + |Xn – Yn|p) ¹/ₚ.
+            + ``'manhattan'`` ➜ The calculation of the distance is given by the expression:
+                ( |x₁ – x₂| + |y₁ – y₂| + ... + |yn – yn|) .
+
+            Defaults to ``'euclidean'``.
+    * max_discards (``int``): This parameter indicates the maximum number of consecutive
+        detector discards, aimed at preventing a possible infinite loop in case a radius
+        is defined that cannot generate non-self detectors. Defaults to ``1000``.
+    * seed (``int``): Seed for the random generation of values in the detectors. Defaults to
+        ``None``.
+    * algorithm(``str``), Set the algorithm version:
+        * ``'default-NSA'``: Default algorithm with fixed radius.
+        * ``'V-detector'``: This algorithm is based on the article \
+            [Real-Valued Negative Selection Algorithm with Variable-Sized Detectors][2], \
+            by Ji, Z., Dasgupta, D. (2004), and uses a variable radius for anomaly \
+            detection in feature spaces.
+
+        Defaults to ``'default-NSA'``.
+
+    * ``**kwargs``:
+        - non_self_label (``str``): This variable stores the label that will be assigned \
+            when the data has only one output class, and the sample is classified as not \
+            belonging to that class. Defaults to ``'non-self'``.
+        - cell_bounds (``bool``): If set to ``True``, this option limits the generation \
+            of detectors to the space within the plane between 0 and 1. This means that \
+            any detector whose radius exceeds this limit is discarded, this variable is \
+            only used in the ``V-detector`` algorithm. Defaults to ``False``.
+        - p (``float``): This parameter stores the value of ``p`` used in the Minkowski \
+            distance. The default is ``2``, which represents normalized Euclidean distance.\
+            Different values of p lead to different variants of the [Minkowski Distance][1].
+
+    Notes
+    ----------
+    [1] https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.minkowski_distance.html
+
+    [2] https://doi.org/10.1007/978-3-540-24854-5_30
+
+    """
+
+    def __init__(
+        self,
+        N: int = 100,
+        r: float = 0.05,
+        r_s: float = 0.0001,
+        k: int = 1,
+        metric: Literal["manhattan", "minkowski", "euclidean"] = "euclidean",
+        max_discards: int = 1000,
+        seed: int = None,
+        algorithm: Literal["default-NSA", "V-detector"] = "default-NSA",
+        **kwargs: Dict[str, Union[bool, str, float]],
+    ):
+        self.metric = sanitize_choice(metric, ["manhattan", "minkowski"], "euclidean")
+        self.seed = sanitize_seed(seed)
+        if self.seed is not None:
+            np.random.seed(seed)
+        self.k: int = sanitize_param(k, 1, lambda x: x > 1)
+        self.N: int = sanitize_param(N, 100, lambda x: x >= 1)
+        self.r: float = sanitize_param(r, 0.05, lambda x: x > 0)
+        self.r_s: float = sanitize_param(r_s, 0.0001, lambda x: x > 0)
+        self.algorithm: str = sanitize_param(
+            algorithm, "default-NSA", lambda x: x == "V-detector"
+        )
+        self.max_discards: int = sanitize_param(max_discards, 1000, lambda x: x > 0)
+
+        # Retrieves the variables from kwargs.
+        self.p: float = kwargs.get("p", 2)
+        self.cell_bounds: bool = kwargs.get("cell_bounds", False)
+        self.non_self_label: str = kwargs.get("non_self_label", "non-self")
+
+        # Initializes the other class variables as None.
+        self._detectors: Union[dict, None] = None
+        self.classes: npt.NDArray = None
+
+    def fit(self, X: npt.NDArray, y: npt.NDArray, verbose: bool = True):
+        """
+        The function ``fit(...)``, performs the training according to ``X`` and ``y``, using the
+        method negative selection method(``NegativeSelect``).
+
+        Parameters
+        ----------
+        * X (``npt.NDArray``): Training array, containing the samples and their \
+            characteristics, [``N samples`` (rows)][``N features`` (columns)].
+        * y (``npt.NDArray``): Array of target classes of ``X`` with [``N samples`` (lines)].
+            verbose (``bool``): Feedback from detector generation to the user.
+
+        Raises
+        ----------
+        * TypeError: If X or y are not ndarrays or have incompatible shapes.
+        * MaxDiscardsReachedError: The maximum number of detector discards was reached during
+            maturation. Check the defined radius value and consider reducing it.
+
+        Returns
+        ----------
+        * (``self``): Returns the instance itself.
+        """
+        progress = None
+        super()._check_and_raise_exceptions_fit(X, y)
+
+        # Identifying the possible classes within the output array `y`.
+        self.classes = np.unique(y)
+        # Dictionary that will store detectors with classes as keys.
+        list_detectors_by_class = {}
+        # Separates the classes for training.
+        sample_index = self.__slice_index_list_by_class(y)
+        # Progress bar for generating all detectors.
+        if verbose:
+            progress = tqdm(
+                total=int(self.N * (len(self.classes))),
+                bar_format="{desc} ┇{bar}┇ {n}/{total} detectors",
+                postfix="\n",
+            )
+        for _class_ in self.classes:
+            # Initializes the empty set that will contain the valid detectors.
+            valid_detectors_set = []
+            discard_count = 0
+            x_class = X[sample_index[_class_]]
+            # Indicating which class the algorithm is currently processing for the progress bar.
+            if verbose:
+                progress.set_description_str(
+                    f"Generating the detectors for the {_class_} class:"
+                )
+            while len(valid_detectors_set) < self.N:
+                # Generates a candidate detector vector randomly with values between 0 and 1.
+                vector_x = np.random.random_sample(size=X.shape[1])
+                # Checks the validity of the detector for non-self with respect to the class samples
+                valid_detector = self.__checks_valid_detector(x_class, vector_x)
+
+                # If the detector is valid, add it to the list of valid detectors.
+                if valid_detector is not False:
+                    discard_count = 0
+                    radius = valid_detector[1] if self.algorithm == "V-detector" else None
+                    valid_detectors_set.append(Detector(vector_x, radius))
+                    if verbose:
+                        progress.update(1)
+                else:
+                    discard_count += 1
+                    if discard_count == self.max_discards:
+                        raise MaxDiscardsReachedError(_class_)
+
+            # Add detectors, with classes as keys in the dictionary.
+            list_detectors_by_class[_class_] = np.array(valid_detectors_set)
+        # Notify completion of detector generation for the classes.
+        if verbose:
+            progress.set_description(
+                f"\033[92m✔ Non-self detectors for classes ({', '.join(map(str, self.classes))}) "
+                f"successfully generated\033[0m"
+            )
+        # Saves the found detectors in the attribute for the non-self detectors of the trained model
+        self._detectors = list_detectors_by_class
+        return self
+
+    def predict(self, X: npt.NDArray) -> Optional[npt.NDArray]:
+        """
+        Function to perform the prediction of classes based on detectors
+        created after training.
+
+        Parameters
+        ----------
+        * X (``npt.NDArray``)
+            Array with input samples with [``N samples`` (Lines)] and
+            [``N characteristics``(Columns)]
+
+        Raises
+        ----------
+        * TypeError
+            If X is not an ndarray or list.
+        * FeatureDimensionMismatch
+            If the number of features in X does not match the expected number.
+
+        Returns
+        ----------
+        * C (``npt.NDArray``)
+            an ndarray of the form ``C`` [``N samples``], containing the predicted classes
+            for ``X``.
+        * ``None``
+            If there are no detectors for the prediction.
+        """
+        # If there are no detectors, Returns None.
+        if self._detectors is None:
+            return None
+
+        super()._check_and_raise_exceptions_predict(
+            X, len(self._detectors[self.classes[0]][0].position)
+        )
+
+        # Initializes an empty array that will store the predictions.
+        c = []
+        # For each sample row in X.
+        for line in X:
+            class_found: bool
+            _class_ = self.__compare_sample_to_detectors(line)
+            if _class_ is None:
+                class_found = False
+            else:
+                c.append(_class_)
+                class_found = True
+
+            # If there is only one class and the sample is not classified,
+            # set the output as non-self.
+            if not class_found and len(self.classes) == 1:
+                c.append(self.non_self_label)
+            # If the class is not identified with the detectors, assign the class with
+            # the greatest distance from the mean of its detectors.
+            elif not class_found:
+                average_distance: dict = {}
+                for _class_ in self.classes:
+                    detectores = list(
+                        map(lambda x: x.position, self._detectors[_class_])
+                    )
+                    average_distance[_class_] = np.average(
+                        [self.__distance(detector, line) for detector in detectores]
+                    )
+                c.append(max(average_distance, key=average_distance.get))
+        return np.array(c)
+
+    def __slice_index_list_by_class(self, y: npt.NDArray) -> dict:
+        """
+        The function ``__slice_index_list_by_class(...)``, separates the indices of the lines
+        according to the output class, to loop through the sample array, only in positions where
+        the output is the class being trained.
+
+        Parameters
+        ----------
+        * y (npt.NDArray)
+            Receives a ``y``[``N sample``] array with the output classes of the \
+            ``X`` sample array.
+
+        Returns
+        ----------
+        * dict: A dictionary with the list of array positions(``y``), with the classes as key.
+        """
+        return slice_index_list_by_class(self.classes, y)
+
+    def __checks_valid_detector(
+        self, x_class: npt.NDArray = None, vector_x: npt.NDArray = None
+    ) -> Union[bool, tuple[bool, float]]:
+        """
+        Function to check if the detector has a valid non-proper ``r`` radius for the class.
+
+        Parameters
+        ----------
+        * x_class (``npt.NDArray``)
+            Array ``x_class`` with the samples per class.
+        * vector_x (``npt.NDArray``)
+            Randomly generated vector x candidate detector with values between[0, 1].
+
+        Returns
+        ----------
+        * Validity (``bool``): Returns whether the detector is valid or not.
+        """
+        # If any of the input arrays have zero size, Returns false.
+        if np.size(x_class) == 0 or np.size(vector_x) == 0:
+            return False
+        # If self.k > 1, uses the k nearest neighbors (kNN); otherwise, checks the detector
+        # without considering kNN.
+        if self.k > 1:
+            knn_list = np.empty(shape=0)
+            for x in x_class:
+                # Calculates the distance between the two vectors and adds it to the kNN list if
+                # the distance is smaller than the largest distance in the list.
+                knn_list = self.__compare_knearest_neighbors_list(
+                    knn_list, self.__distance(x, vector_x)
+                )
+            # If the average of the distances in the kNN list is less than the radius, Returns true.
+            distance_mean = np.mean(knn_list)
+            if self.algorithm == "V-detector":
+                return self.__detector_is_valid_to_vdetector(distance_mean, vector_x)
+            if distance_mean > (self.r + self.r_s):
+                return True
+        else:
+            if self.algorithm == "V-detector":
+                distance = min_distance_to_class_vectors(
+                    x_class, vector_x, get_metric_code(self.metric), self.p
+                )
+                return self.__detector_is_valid_to_vdetector(distance, vector_x)
+
+            # Calculates the distance between the vectors; if not it is less than or equal to
+            # the radius plus the sample's radius, sets the validity of the detector to
+            # true.
+            threshold: float = self.r + self.r_s
+            if check_detector_rnsa_validity(
+                x_class, vector_x, threshold, get_metric_code(self.metric), self.p
+            ):
+                return True  # Detector is valid!
+
+        return False  # Detector is not valid!
+
+    def __compare_knearest_neighbors_list(
+        self, knn: npt.NDArray, distance: float
+    ) -> npt.NDArray:
+        """
+        Compares the k-nearest neighbor distance at position ``k-1`` in the list ``knn``,
+        if the distance of the new sample is less, replace it and sort in ascending order.
+
+
+        Parameters
+        ----------
+        * knn (``npt.NDArray``)
+            List of k-nearest neighbor distances.
+        * distance (``float``)
+            Distance to check.
+
+        Returns
+        ----------
+        * ``npt.NDArray``: Updated and sorted nearest neighbor list.
+        """
+        # If the number of distances in kNN is less than k, adds the distance.
+        if len(knn) < self.k:
+            knn = np.append(knn, distance)
+            knn.sort()
+            return knn
+
+        # Otherwise, add the distance if the new distance is smaller than the largest
+        # distance in the list.
+        if knn[self.k - 1] > distance:
+            knn[self.k - 1] = distance
+            knn.sort()
+
+        return knn
+
+    def __compare_sample_to_detectors(self, line: npt.NDArray):
+        """
+        Function to compare a sample with the detectors, verifying if the sample is proper.
+
+        Parameters
+        ----------
+        * line (``npt.NDArray``): vector with N-features
+
+        Returns
+        ----------
+        * Returns the predicted class with the detectors or None if the sample does not qualify
+        for any class.
+        """
+        # List to store the classes and the average distance between the detectors and the sample.
+        possible_classes = []
+        for _class_ in self.classes:
+            # Variable to indicate if the class was found with the detectors.
+            class_found: bool = True
+            sum_distance = 0
+            for detector in self._detectors[_class_]:
+                distance = self.__distance(detector.position, line)
+                sum_distance += distance
+                if self.algorithm == "V-detector":
+                    if distance <= detector.radius:
+                        class_found = False
+                        break
+                elif distance <= self.r:
+                    class_found = False
+                    break
+
+            # If the sample passes through all the detectors of a class, adds the class as a
+            # possible prediction.
+            if class_found:
+                possible_classes.append([_class_, sum_distance / self.N])
+        # If classified as belonging to only one class, Returns the class.
+        if len(possible_classes) == 1:
+            return possible_classes[0][0]
+        # If belonging to more than one class, Returns the class with the greatest average distance.
+        if len(possible_classes) > 1:
+            return max(possible_classes, key=lambda x: x[1])[0]
+
+        return None
+
+    def __distance(self, u: npt.NDArray, v: npt.NDArray) -> float:
+        """
+        Function to calculate the distance between two points by the chosen ``metric``.
+
+        Parameters
+        ----------
+        * u (``npt.NDArray``): Coordinates of the first point.
+        * v (``npt.NDArray``): Coordinates of the second point.
+
+        Returns
+        ----------
+        * Distance (``float``): between the two points.
+        """
+        return compute_metric_distance(u, v, get_metric_code(self.metric), self.p)
+
+    def __detector_is_valid_to_vdetector(
+        self, distance: float, vector_x: npt.NDArray
+    ) -> Union[bool, tuple[bool, float]]:
+        """
+        Check if the distance between the detector and the samples, minus the radius of the samples,
+        is greater than the minimum radius.
+
+        Parameters
+        ----------
+        * distance (``float``): minimum distance calculated between all samples.
+        * vector_x (``numpy.ndarray``): randomly generated candidate detector vector x with
+            values between 0 and 1.
+
+        Returns
+        ----------
+        * ``False`` if the calculated radius is smaller than the minimum distance or exceeds the
+            edge of the space, if this option is enabled.
+        * ``True`` and the distance minus the radius of the samples, if the radius is valid.`
+        """
+        new_detector_r = float(distance - self.r_s)
+        if self.r >= new_detector_r:
+            return False
+
+        # If _cell_bounds is True, considers the detector to be within the plane bounds.
+        if self.cell_bounds:
+            for p in vector_x:
+                if (p - new_detector_r) < 0 or (p + new_detector_r) > 1:
+                    return False
+
+        return (True, new_detector_r)
+
+
+class BNSA(BaseNSA):
+    """
+    The ``BNSA`` (Binary Negative Selection Algorithm) class is for classification and
+    identification purposes of anomalies through the self and not self method.
+
+    Parameters
+    ----------
+    * N (``int``): Number of detectors. Defaults to ``100``.
+    * aff_thresh (``float``): The variable represents the percentage of similarity
+        between the T cell and the own samples. The default value is 10% (0.1), while a value of
+        1.0 represents 100% similarity.
+    * max_discards (``int``): This parameter indicates the maximum number of detector discards in
+        sequence, which aims to avoid a possible infinite loop if a radius is defined that it is
+        not possible to generate non-self detectors. Defaults to ``1000``.
+    * seed (``int``): Seed for the random generation of values in the detectors. Defaults to
+        ``None``.
+    * no_label_sample_selection (``str``): Method for selecting labels for samples designated as
+        non-self by all detectors. Available method types:
+        - (``max_average_difference``): Selects the class with the highest average difference
+            among the detectors.
+        - (``max_nearest_difference``): Selects the class with the highest difference between
+            the nearest and farthest detector from the sample.
+    """
+
+    def __init__(
+        self,
+        N: int = 100,
+        aff_thresh: float = 0.1,
+        max_discards: int = 1000,
+        seed: int = None,
+        no_label_sample_selection: Literal[
+            "max_average_difference", "max_nearest_difference"
+        ] = "max_average_difference",
+    ):
+        super().__init__()
+
+        self.N: int = sanitize_param(N, 100, lambda x: x > 0)
+        self.aff_thresh: float = sanitize_param(aff_thresh, 0.1, lambda x: 0 < x < 1)
+        self.max_discards: float = sanitize_param(max_discards, 1000, lambda x: x > 0)
+
+        self.seed = sanitize_seed(seed)
+
+        if self.seed is not None:
+            np.random.seed(seed)
+
+        self.no_label_sample_selection: float = sanitize_param(
+            no_label_sample_selection,
+            "max_average_difference",
+            lambda x: x == "nearest_difference",
+        )
+
+        self.classes: npt.NDArray = None
+        self._detectors: Optional[dict] = None
+        self._detectors_stack: npt.NDArray = None
+
+    def fit(self, X: npt.NDArray, y: npt.NDArray, verbose: bool = True):
+        """
+        The function ``fit(...)``, performs the training according to ``X`` and ``y``, using the
+        method negative selection method(``NegativeSelect``).
+
+        Parameters
+        ----------
+        * X (``npt.NDArray``):
+            Training array, containing the samples and their characteristics,
+            [``N samples`` (rows)][``N features`` (columns)].
+        * y (``npt.NDArray``):
+            Array of target classes of ``X`` with [``N samples`` (lines)].
+            verbose (``bool``): Feedback from detector generation to the user.
+
+        Returns
+        ----------
+        * (``self``): Returns the instance itself.
+        """
+        super()._check_and_raise_exceptions_fit(X, y, "BNSA")
+
+        # Converts the entire array X to boolean
+        X = X.astype(np.bool_)
+
+        # Identifying the possible classes within the output array `y`.
+        self.classes = np.unique(y)
+        # Dictionary that will store detectors with classes as keys.
+        list_detectors_by_class = {}
+        # Separates the classes for training.
+        sample_index: dict = self.__slice_index_list_by_class(y)
+        # Progress bar for generating all detectors.
+        if verbose:
+            progress = tqdm(
+                total=int(self.N * (len(self.classes))),
+                bar_format="{desc} ┇{bar}┇ {n}/{total} detectors",
+                postfix="\n",
+            )
+
+        for _class_ in self.classes:
+            # Initializes the empty set that will contain the valid detectors.
+            valid_detectors_set: list = []
+            discard_count: int = 0
+            # Updating the progress bar with the current class the algorithm is processing.
+            if verbose:
+                progress.set_description_str(
+                    f"Generating the detectors for the {_class_} class:"
+                )
+            x_class = X[sample_index[_class_]]
+            while len(valid_detectors_set) < self.N:
+                # Generates a candidate detector vector randomly with values 0 and 1.
+                vector_x = np.random.randint(0, 2, size=X.shape[1]).astype(np.bool_)
+                # If the detector is valid, add it to the list of valid detectors.
+                if check_detector_bnsa_validity(x_class, vector_x, self.aff_thresh):
+                    discard_count = 0
+                    valid_detectors_set.append(vector_x)
+                    if verbose:
+                        progress.update(1)
+                else:
+                    discard_count += 1
+                    if discard_count == self.max_discards:
+                        raise MaxDiscardsReachedError(_class_)
+
+            # Add detectors to the dictionary with classes as keys.
+            list_detectors_by_class[_class_] = np.array(valid_detectors_set)
+
+        # Notify the completion of detector generation for the classes.
+        if verbose:
+            progress.set_description(
+                f"\033[92m✔ Non-self detectors for classes ({', '.join(map(str, self.classes))}) "
+                f"successfully generated\033[0m"
+            )
+        # Saves the found detectors in the attribute for the class detectors.
+        self._detectors = list_detectors_by_class
+        self._detectors_stack = np.array(
+            [np.stack(self._detectors[class_name]) for class_name in self.classes]
+        )
+        return self
+
+    def predict(self, X: npt.NDArray) -> Optional[npt.NDArray]:
+        """
+        Function to perform the prediction of classes based on detectors
+        created after training.
+
+        Parameters
+        ----------
+        * X (``npt.NDArray``): Array with input samples with [``N samples`` (Lines)] and
+            [``N characteristics``(Columns)]
+
+        Returns
+        ----------
+        * c (``npt.NDArray``): an ndarray of the form ``C`` [``N samples``],
+            containing the predicted classes for ``X``
+        * ``None``: If there are no detectors for the prediction.
+        """
+        # If there are no detectors, Returns None.
+        if self._detectors is None:
+            return None
+
+        super()._check_and_raise_exceptions_predict(
+            X, len(self._detectors[self.classes[0]][0]), "BNSA"
+        )
+
+        # Converts the entire array X to boolean.
+        if X.dtype != bool:
+            X = X.astype(bool)
+
+        # Initializes an empty array that will store the predictions.
+        c = []
+        # For each sample row in X.
+        for line in X:
+            class_found: bool = True
+            # Class prediction based on detectors
+            class_index = bnsa_class_prediction(
+                line, self._detectors_stack, self.aff_thresh
+            )
+            # If belonging to one or more classes, adds the class with the greatest
+            # average distance
+            if class_index > -1:
+                c.append(self.classes[class_index])
+                class_found = True
+            else:
+                class_found = False
+
+            # If there is only one class and the sample is not classified, sets the
+            # output as non-self.
+            if not class_found and len(self.classes) == 1:
+                c.append("non-self")
+            # If the class cannot be identified by the detectors
+            elif not class_found:
+                self.__assign_class_to_non_self_sample(line, c)
+
+        return np.array(c)
+
+    def __assign_class_to_non_self_sample(self, line: npt.NDArray, c: list):
+        """
+        This function determines the class of a sample when all detectors classify it
+        as "non-self". Classification is performed using the ``max_average_difference``
+        and ``max_nearest_difference`` methods.
+
+        Parameters
+        ----------
+        * line (list): Sample to be classified.
+        * c (list): List of predictions to be updated with the new classification.
+
+        Returns
+        ----------
+        * list: The list of predictions `c` updated with the class assigned to the sample.
+        """
+        class_differences: dict = {}
+        for _class_ in self.classes:
+            distances = np.sum(line != self._detectors[_class_]) / self.N
+            # Assign the label to the class with the greatest distance from
+            # the nearest detector.
+            if self.no_label_sample_selection == "nearest_difference":
+                class_differences[_class_] = distances.min()
+            # Or based on the greatest distance from the average distances of the detectors.
+            else:
+                class_differences[_class_] = distances.sum() / self.N
+
+        c.append(max(class_differences, key=class_differences.get))
+
+    def __slice_index_list_by_class(self, y: npt.NDArray) -> dict:
+        """
+        The function ``__slice_index_list_by_class(...)``, separates the indices of the lines
+        according to the output class, to loop through the sample array, only in positions where
+        the output is the class being trained.
+
+        Parameters
+        ----------
+        * y (``npt.NDArray``):
+            Receives a ``y``[``N sample``] array with the output classes of the ``X``
+            sample array.
+
+        Returns
+        ----------
+        * dict: A dictionary with the list of array positions(``y``), with the classes as key.
+        """
+        return slice_index_list_by_class(self.classes, y)