pyAgrum-nightly 2.3.1.9.dev202512261765915415__cp310-abi3-macosx_10_15_x86_64.whl

pyagrum/lib/discreteTypeProcessor.py

@@ -0,0 +1,1102 @@
+############################################################################
+#   This file is part of the aGrUM/pyAgrum library.                        #
+#                                                                          #
+#   Copyright (c) 2005-2025 by                                             #
+#       - Pierre-Henri WUILLEMIN(_at_LIP6)                                 #
+#       - Christophe GONZALES(_at_AMU)                                     #
+#                                                                          #
+#   The aGrUM/pyAgrum library is free software; you can redistribute it    #
+#   and/or modify it under the terms of either :                           #
+#                                                                          #
+#    - the GNU Lesser General Public License as published by               #
+#      the Free Software Foundation, either version 3 of the License,      #
+#      or (at your option) any later version,                              #
+#    - the MIT license (MIT),                                              #
+#    - or both in dual license, as here.                                   #
+#                                                                          #
+#   (see https://agrum.gitlab.io/articles/dual-licenses-lgplv3mit.html)    #
+#                                                                          #
+#   This aGrUM/pyAgrum library is distributed in the hope that it will be  #
+#   useful, but WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,          #
+#   INCLUDING BUT NOT LIMITED TO THE WARRANTIES MERCHANTABILITY or FITNESS #
+#   FOR A PARTICULAR PURPOSE  AND NONINFRINGEMENT. IN NO EVENT SHALL THE   #
+#   AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER #
+#   LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,        #
+#   ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR  #
+#   OTHER DEALINGS IN THE SOFTWARE.                                        #
+#                                                                          #
+#   See LICENCES for more details.                                         #
+#                                                                          #
+#   SPDX-FileCopyrightText: Copyright 2005-2025                            #
+#       - Pierre-Henri WUILLEMIN(_at_LIP6)                                 #
+#       - Christophe GONZALES(_at_AMU)                                     #
+#   SPDX-License-Identifier: LGPL-3.0-or-later OR MIT                      #
+#                                                                          #
+#   Contact  : info_at_agrum_dot_org                                       #
+#   homepage : http://agrum.gitlab.io                                      #
+#   gitlab   : https://gitlab.com/agrumery/agrum                           #
+#                                                                          #
+############################################################################
+
+"""
+This module contains the DiscreteTypeProcessor class used to attribute a type from the values of a database before learning a (discrete) Graphical Model.
+Particularly, the DiscreteTypeProcessor class is used to discretize some continuous variables in a database.
+
+The discretization is done using the following methods: 'quantile', 'uniform', 'kmeans', 'NML', 'CAIM' and 'MDLP'.
+Some methods need specialized treatment of arguments for classification (y separated from X).
+"""
+
+import math
+
+import numpy
+import pandas
+import scipy
+import sklearn
+import sklearn.preprocessing as skp
+
+from typing import Any
+
+import pyagrum as gum
+
+
+def check_int(v: Any) -> bool:
+  """
+  Test if v is an int or a str representing an int
+  """
+  if isinstance(v, (bool, numpy.bool_)):
+    return False
+  if isinstance(v, (int, numpy.integer)):
+    return True
+
+  if isinstance(v, str):
+    if v[0] in ("-", "+"):
+      return v[1:].isdigit()
+    return v.isdigit()
+  return False
+
+
+def check_float(v: Any) -> bool:
+  """
+  Test if v is a float or a str representing a float.
+
+  Parameters
+  ----------
+  v : Any
+
+  Returns
+  -------
+  bool:
+    True if v is a float or a str representing a float
+  """
+  if isinstance(v, (bool, numpy.bool_)):
+    return False
+
+  try:
+    float(v)
+    return True
+  except ValueError:
+    return False
+
+
+class DiscreteTypeProcessor:
+  """
+  Represents a tool to process the type of the variables in a database in order to obtain  a way to learn a pyAgrum's discrete Graphical Model.
+  Particularly, the DiscreteTypeProcessor class is used to discretize some continuous variables in a database.
+
+  Warnings
+  --------
+  - The data are represented by tabular data (X and possibly y) where the columns are the variables and the rows are the samples. Generally, X can be replaced by a the name of a csv file.
+  - In the case of a classification, y is the class variable and X are the features. y has not to be binary.
+
+  Parameters
+  ----------
+      defaultDiscretizationMethod: str
+            sets the default method of discretization for this discretizer. Possible values are: `quantile`, `uniform`,
+            `kmeans`, `NML`, `CAIM` and `MDLP`. This method will be used if the user has not specified another method
+            for that specific variable using the setDiscretizationParameters method.
+      defaultNumberOfBins: str or int
+          sets the number of bins if the method used is `quantile`, `kmeans`, `uniform`. In this case this parameter can also
+          be set to the string `elbowMethod` so that the best number of bins is found automatically.
+          If the method used is NML, this parameter sets the the maximum number of bins up to which the NML algorithm
+          searches for the optimal number of bins. In this case this parameter must be an int
+          If any other discretization method is used, this parameter is ignored.
+      discretizationThreshold: int or float
+          When using default parameters a variable will be treated as continuous only if it has more unique values than
+          this number (if the number is an int greater than 1).
+          If the number is a float between 0 and 1, we will test if the proportion of unique values is bigger than this
+          number. For example if you have entered 0.95, the variable will be treated as continuous only if more than 95%
+          of its values are unique.
+
+  """
+
+  def __init__(
+    self,
+    defaultDiscretizationMethod="quantile",
+    defaultNumberOfBins=10,
+    discretizationThreshold=25,
+  ):
+    """
+    Initializes the DiscreteTypeProcessor object.
+
+    Parameters
+    ----------
+        defaultDiscretizationMethod: str
+            sets the default method of discretization for this discretizer. Possible values are: 'quantile', 'uniform',
+            'kmeans', 'NML', 'CAIM' and 'MDLP'. This method will be used if the user has not specified another method
+            for that specific variable using the setDiscretizationParameters method.
+        defaultNumberOfBins: str or int
+            sets the number of bins if the method used is quantile, kmeans, uniform. In this case this parameter can
+            also be set to the string 'elbowMethod' so that the best number of bins is found automatically.
+            If the method used is NML, this parameter sets the the maximum number of bins up to which the NML algorithm
+            searches for the optimal number of bins. In this case this parameter must be an int
+            If any other discretization method is used, this parameter is ignored.
+        discretizationThreshold: int or float
+            When using default parameters a variable will be treated as continuous only if it has more unique values than
+            this number (if the number is an int greater than 1).
+            If the number is a float between 0 and 1, we will test if the proportion of unique values is bigger than
+            this number. For example if you have entered 0.95, the variable will be treated as continuous only if more
+            than 95% of its values are unique.
+    """
+    self.discretizationParametersDictionary = {}
+    self.numberOfContinuous = 0
+    self.totalNumberOfBins = 0
+    self.defaultMethod = None
+    self.defaultParamDiscretizationMethod = None
+    self.setDiscretizationParameters(None, defaultDiscretizationMethod, defaultNumberOfBins)
+
+    self.discretizationThreshold = discretizationThreshold
+
+  def clear(self, clearDiscretizationParameters=False):
+    """
+    Sets the number of continuous variables and the total number of bins created by this discretizer to 0. If
+    clearDiscretizationParameters is True, also clears the the parameters for discretization the user has set for each
+    variable.
+
+    Parameters
+    ----------
+      clearDiscretizationParameters: bool
+        if True, this method also clears the parameters the user has set for each variable and resets them to the default.
+    """
+    self.numberOfContinuous = 0
+    self.totalNumberOfBins = 0
+    if clearDiscretizationParameters:
+      self.discretizationParametersDictionary = {}
+
+  @gum.deprecated_arg(newA="parameters", oldA="paramDiscretizationMethod", version="2.0.0")
+  def setDiscretizationParameters(self, variableName: str, method: str, parameters: Any = None):
+    """
+    Sets the discretization parameters for a variable. If variableName is None, sets the default parameters.
+
+    Parameters
+    ----------
+      variableName: str
+          the name of the variable you want to set the discretization parameters of. Set to None to set the new
+          default.
+      method: str
+          The method of discretization used for this variable. Use "NoDiscretization" if you do not want to discretize this
+          variable. Possible values are: 'NoDiscretization', 'quantile', 'uniform', 'kmeans', 'NML', 'CAIM', 'MDLP' and 'expert'
+      parameters:
+          Each method of discretization has a parameter that can be set.
+          - 'quantile': the number of bins
+          - 'kmeans', 'uniform': the number of bins. The parameter can also be set to the string 'elbowMethod' so that the best
+          number of bins is found automatically.
+          - 'NML': this parameter sets the the maximum number of bins up to which the NML algorithm searches for the optimal
+          number of bins.
+          - 'MDLP', 'CAIM': this parameter is ignored
+          - 'expert': this parameter is the set of ticks proposed by the expert. The discretized variable will set the flag
+          'empirical' which means that if the values found in the data are not in the proposed intervals, they did not raise
+          any exception but  are nevertheless accepted (as belonging to the smallest or biggest interval).
+          - 'NoDiscretization': this parameter is a superset of the values for the variable found in the database (or None).
+    """
+    if parameters is None:
+      parameters = self.defaultParamDiscretizationMethod
+
+    match method:
+      case "quantile" | "NML":
+        if type(parameters) is not int:
+          raise ValueError(
+            "The parameter for the quantile/NML method must be an integer. You have entered: " + str(parameters)
+          )
+      case "kmeans" | "uniform":
+        if type(parameters) is not int and str(parameters) != "elbowMethod":
+          raise ValueError(
+            "The parameter for the kmeans/uniform method must be an integer or the string 'elbowMethod'. You have entered: "
+            + str(parameters)
+          )
+      case "expert":
+        if not (isinstance(parameters, list) and all(map(check_float, parameters))):
+          raise ValueError(
+            "The parameter for the expert method must be a list of float. You have entered: " + str(parameters)
+          )
+      case "NoDiscretization":
+        if parameters is not None and not (isinstance(parameters, str)):
+          raise ValueError(
+            "The parameter for the NoDiscretization method must be a string (fastVar syntax) or None. You have "
+            "entered: " + str(parameters)
+          )
+      case "CAIM" | "MDLP":
+        pass
+      case _:
+        raise ValueError(
+          "This discretization method is not recognized! Possible values are kmeans, uniform, quantile, NML, "
+          "CAIM, MDLP, NoDiscretization or expert. You have entered " + str(method)
+        )
+
+    if variableName is None:
+      self.defaultMethod = method
+      self.defaultParamDiscretizationMethod = parameters
+    else:
+      self.discretizationParametersDictionary[variableName] = {}
+      self.discretizationParametersDictionary[variableName]["method"] = method
+      self.discretizationParametersDictionary[variableName]["param"] = parameters
+
+  def audit(self, X, y=None):
+    """
+    Audits the passed values of X and y. Guess which columns in X are already discrete and which need to
+    be discretized, as well as the discretization algorithm that will be used to discretize them The parameters which
+    are suggested will be used when creating the variables. To change this the user can manually set discretization
+    parameters for each variable using the setDiscretizationParameters function.
+
+    Parameters
+    ----------
+      X: {array-like, pandas or polars dataframe} of shape (n_samples, n_features) or str (filename)
+          training data
+      y: {array-like, pandas or polars dataframe} of shape (n_samples,) or str (classname)
+          Target values
+    Returns
+    -------
+      Dict
+        for each variable, the proposition of audit
+    """
+    if isinstance(X, str):
+      Xp = pandas.read_csv(X)
+    elif hasattr(X, "to_pandas"):  # for instance, polars dataframe
+      Xp = X.to_pandas()
+    else:
+      Xp = X
+
+    if isinstance(y, str):
+      yp = Xp[y]
+    elif y is not None and hasattr(y, "to_pandas"):  # for instance, polars dataframe
+      yp = y.to_pandas()
+    else:
+      yp = y
+
+    return self._audit(Xp, yp)
+
+  def _audit(self, X, y=None):
+    """
+    Audits the passed values of X and y. Tells us which columns in X we think are already discrete and which need to
+    be discretized, as well as the discretization algorithm that will be used to discretize them The parameters which
+    are suggested will be used when creating the variables. To change this the user can manually set discretization
+    parameters for each variable using the setDiscretizationParameters function.
+
+        Parameters
+    ----------
+      X: {array-like, sparse matrix} of shape (n_samples, n_features)
+          training data
+      y: array-like of shape (n_samples,)
+          Target values
+    Returns
+    -------
+      Dict
+        for each variable, the proposition of audit
+    """
+
+    auditDict = {}
+
+    if isinstance(X, pandas.DataFrame):
+      variableNames = X.columns.tolist()
+    elif type(X) is pandas.core.series.Series:
+      variableNames = [X.name]
+    else:
+      variableNames = None
+
+    if y is not None:
+      X, y = sklearn.utils.check_X_y(X, y, dtype=None, accept_sparse=True)
+    else:
+      X = sklearn.utils.check_array(X, dtype=None)
+
+    d = X.shape[1]
+
+    if variableNames is None:
+      variableNames = ["x" + str(i) for i in range(d)]
+
+    possibleValues = {}  # counting the possible values gor this line
+
+    for i in range(d):
+      possibleValues[i] = numpy.unique(X[:, i])
+    possibleValues[d] = numpy.unique(y)
+
+    for i in range(d):
+      variable = variableNames[i]
+      auditDict[variable] = {}
+      try:
+        sklearn.utils.check_array(X[:, i], dtype="float", ensure_2d=False)
+        isNumeric = True
+      except ValueError:
+        isNumeric = False
+      if variable in self.discretizationParametersDictionary.keys():
+        auditDict[variable] = self.discretizationParametersDictionary[variable]
+        if self.discretizationParametersDictionary[variable]["method"] != "NoDiscretization" and not isNumeric:
+          raise ValueError("The variable " + variable + " is not numeric and cannot be discretized!")
+      else:
+        if len(possibleValues[i]) > self.discretizationThreshold and isNumeric:
+          auditDict[variable]["method"] = self.defaultMethod
+          auditDict[variable]["nbBins"] = self.defaultParamDiscretizationMethod
+        else:
+          auditDict[variable]["method"] = "NoDiscretization"
+          auditDict[variable]["values"] = possibleValues[i]
+
+      if auditDict[variable]["method"] == "NoDiscretization":
+        auditDict[variable]["type"] = "Discrete"
+      else:
+        auditDict[variable]["type"] = "Continuous"
+        auditDict[variable]["minInData"] = min(possibleValues[i])
+        auditDict[variable]["maxInData"] = max(possibleValues[i])
+
+    return auditDict
+
+  @staticmethod
+  def _discretizationElbowMethodRotation(discretizationStrategy, X):
+    """
+    Calculates the sum of squared errors as a function of the number of clusters using the discretization strategy
+    that is passed as a parameter. Returns the bins that are optimal for minimizing the variation and the number of
+    bins at the same time. Uses the elbow method to find this optimal point. To find the "elbow" we rotate the curve
+    and look for its minimum.
+
+    Parameters
+    ----------
+        discretizationStrategy: str
+            The method of discretization that will be used. Possible values are: 'quantile' , 'kmeans' and 'uniform'
+        X: one dimensional ndarray
+            Contains the data that should be discretized
+    Returns
+    -------
+        List[float]
+          the edges of the bins the algorithm has chosen.
+    """
+
+    if discretizationStrategy not in {"kmeans", "quantile", "uniform"}:
+      raise ValueError("cannot use elbow method with this type of discretization")
+    variationArray = numpy.zeros(14)
+    Xsorted = X[X.argsort(axis=None)]
+    binEdgeMatrix = [[]] * 14
+    for k in range(2, 16):
+      discretizer = skp.KBinsDiscretizer(
+        k,
+        strategy=discretizationStrategy,
+        quantile_method="averaged_inverted_cdf",
+        subsample=None,
+      )
+      discretizer.fit(Xsorted)
+      binEdges = discretizer.bin_edges_[0]
+      centresArray = (binEdges[1:] + binEdges[:-1]) / 2
+      i = 0
+      sumOfSquaredErrors = 0
+      for x in Xsorted:
+        if x > binEdges[i + 1]:
+          i = i + 1
+        sumOfSquaredErrors += (x - centresArray[i]) ** 2
+      variationArray[k - 2] = sumOfSquaredErrors
+      binEdgeMatrix[k - 2] = binEdges.to_list()
+
+    # we calculate the slope of the line that connects the first and last point on our graph
+    slope = (variationArray[13] - variationArray[0]) / 13
+
+    # we calculate the slope of the line perpendicular to it
+    otherSlope = -1 / slope
+
+    # we change the coordinate system to the one with the two lines previously mentioned as its axes
+    coordinateChangeMatrix = numpy.array([[1, slope], [1 / otherSlope, 1]])
+
+    # we linearly transform the coordinates of every point in our curve
+    transformedCoordinates = numpy.zeros((2, 14))
+    for i in range(14):
+      transformedCoordinates[:, i] = numpy.matmul(
+        coordinateChangeMatrix,
+        numpy.array([[i], [variationArray[i] - variationArray[0]]]),
+      ).reshape(2)
+
+    # we search for the minimum in our newly obtained curve
+    minkIndex = 0
+    for k in range(14):
+      if transformedCoordinates[1][minkIndex] > transformedCoordinates[1][k]:
+        minkIndex = k
+    # when we have found the minimum, we apply the inverse linear transformation to recover the optimal value of k
+    minimumVector = numpy.matmul(
+      numpy.linalg.inv(coordinateChangeMatrix),
+      transformedCoordinates[:, minkIndex].reshape(2, 1),
+    )
+
+    # we return the list of bin edges found using said optimal number of k
+    return binEdgeMatrix[int(round(minimumVector[0]))]
+
+  def _discretizationMDLP(self, X, y, possibleValuesX, possibleValuesY):
+    """
+    Uses the MDLP algorithm described in Fayyad, 1995 to discretize the values of x.
+
+    Parameters
+    ----------
+        X: ndarray with shape (n,1) where n is the number of samples
+            Column-vector that contains all the data that needs to be discretized
+        y: ndarray with shape (n,1) where n is the number of samples
+            Column-vector that contains the class for each sample. This vector will not be discretized, but the class-value of each sample is needed to properly apply the algorithm
+        possibleValuesX: one dimensional ndarray
+            Contains all the possible values that x can take sorted in increasing order. There shouldn't be any doubles inside
+        possibleValuesY: one dimensional ndarray
+            Contains the possible values of y.
+    Returns
+    -------
+        List[float]
+         a list of the edges of the bins that are chosen by this algorithm
+    """
+    xAndY = numpy.concatenate((X, y), axis=1)
+    xAndY = xAndY[xAndY[:, 0].argsort()]
+    B = (possibleValuesX[1:] + possibleValuesX[:-1]) / 2
+    [class0, _] = possibleValuesY
+
+    binEdgesIndex = []
+    nbElementsByIntervalClass0 = numpy.zeros(len(B) + 1)
+    nbElementsByIntervalClass1 = numpy.zeros(len(B) + 1)
+    currentIntervalIndex = 0
+    for X in xAndY:
+      if currentIntervalIndex < len(B) and X[0] > B[currentIntervalIndex]:
+        currentIntervalIndex += 1
+      if X[1] == class0:
+        nbElementsByIntervalClass0[currentIntervalIndex] += 1
+      else:
+        nbElementsByIntervalClass1[currentIntervalIndex] += 1
+
+    Class0ByLargeInterval = [nbElementsByIntervalClass0.sum()]
+    Class1ByLargeInterval = [nbElementsByIntervalClass1.sum()]
+
+    totalCountByLargeInterval = [Class0ByLargeInterval[0] + Class1ByLargeInterval[0]]
+
+    probabilityClass0 = Class0ByLargeInterval[0] / totalCountByLargeInterval[0]
+    probabilityClass1 = Class1ByLargeInterval[0] / totalCountByLargeInterval[0]
+    shannonEntropyByLargeInterval = [
+      (-1) * (probabilityClass0 * math.log2(probabilityClass0) + probabilityClass1 * math.log2(probabilityClass1))
+    ]
+
+    continueDividingInterval = [True]
+
+    currentValues = {}
+    minimalValues = {}
+
+    while any(continueDividingInterval):
+      minimalValues["classInformationEntropy"] = math.inf
+      for param in {
+        "boundaryIndex",
+        "leftSubintervalClass0",
+        "leftSubintervalClass1",
+        "leftSubintervalShannonEntropy",
+        "rightSubintervalClass0",
+        "rightSubintervalClass1",
+        "rightSubintervalShannonEntropy",
+      }:
+        (currentValues[param], minimalValues[param]) = (0, 0)
+
+      position = 0
+      while currentValues["boundaryIndex"] < len(B):
+        while not continueDividingInterval[position]:
+          position = position + 1
+          currentValues["boundaryIndex"] = binEdgesIndex[position - 1] + 1
+
+        if position < len(binEdgesIndex) and currentValues["boundaryIndex"] == binEdgesIndex[position]:
+          # this function decides whether to accept the cut point in this interval and updates the relevant lists if
+          # the value is accepted.
+          self._divideIntervalMDLP(
+            minimalValues,
+            shannonEntropyByLargeInterval,
+            Class0ByLargeInterval,
+            Class1ByLargeInterval,
+            continueDividingInterval,
+            totalCountByLargeInterval,
+            position,
+            binEdgesIndex,
+          )
+          position += 1
+          while position < len(continueDividingInterval) and not continueDividingInterval[position]:
+            position += 1
+          if position == len(continueDividingInterval):
+            break
+          else:
+            currentValues["boundaryIndex"] = binEdgesIndex[position - 1] + 1
+          (
+            currentValues["leftSubintervalClass0"],
+            currentValues["leftSubintervalClass1"],
+          ) = (0, 0)
+          minimalValues["classInformationEntropy"] = math.inf
+          continue
+
+        currentValues["leftSubintervalClass0"] += nbElementsByIntervalClass0[currentValues["boundaryIndex"]]
+        currentValues["leftSubintervalClass1"] += nbElementsByIntervalClass1[currentValues["boundaryIndex"]]
+
+        totalCountLeftInterval = currentValues["leftSubintervalClass0"] + currentValues["leftSubintervalClass1"]
+
+        probabilityClass0 = currentValues["leftSubintervalClass0"] / totalCountLeftInterval
+        probabilityClass1 = currentValues["leftSubintervalClass1"] / totalCountLeftInterval
+        currentValues["leftSubintervalShannonEntropy"] = 0
+        if probabilityClass0 > 0:
+          currentValues["leftSubintervalShannonEntropy"] += (-1) * probabilityClass0 * math.log2(probabilityClass0)
+        if probabilityClass1 > 0:
+          currentValues["leftSubintervalShannonEntropy"] += (-1) * probabilityClass1 * math.log2(probabilityClass1)
+
+        classInformationEntropy = (totalCountLeftInterval / totalCountByLargeInterval[position]) * currentValues[
+          "leftSubintervalShannonEntropy"
+        ]
+
+        currentValues["rightSubintervalClass0"] = (
+          Class0ByLargeInterval[position] - currentValues["leftSubintervalClass0"]
+        )
+        currentValues["rightSubintervalClass1"] = (
+          Class1ByLargeInterval[position] - currentValues["leftSubintervalClass1"]
+        )
+        NRightInterval = currentValues["rightSubintervalClass0"] + currentValues["rightSubintervalClass1"]
+
+        probabilityClass0 = currentValues["rightSubintervalClass0"] / NRightInterval
+        probabilityClass1 = currentValues["rightSubintervalClass1"] / NRightInterval
+        currentValues["rightSubintervalShannonEntropy"] = 0
+        if probabilityClass0 > 0:
+          currentValues["rightSubintervalShannonEntropy"] += (-1) * probabilityClass0 * math.log2(probabilityClass0)
+        if probabilityClass1 > 0:
+          currentValues["rightSubintervalShannonEntropy"] += (-1) * probabilityClass1 * math.log2(probabilityClass1)
+
+        classInformationEntropy += (NRightInterval / totalCountByLargeInterval[position]) * currentValues[
+          "rightSubintervalShannonEntropy"
+        ]
+
+        if classInformationEntropy < minimalValues["classInformationEntropy"]:
+          minimalValues = currentValues.copy()
+          minimalValues["classInformationEntropy"] = classInformationEntropy
+
+        currentValues["boundaryIndex"] += 1
+
+      if continueDividingInterval[-1]:
+        self._divideIntervalMDLP(
+          minimalValues,
+          shannonEntropyByLargeInterval,
+          Class0ByLargeInterval,
+          Class1ByLargeInterval,
+          continueDividingInterval,
+          totalCountByLargeInterval,
+          position,
+          binEdgesIndex,
+        )
+    binEdges = [xAndY[0][0]]
+    for index in binEdgesIndex:
+      binEdges.append(B[index])
+    binEdges.append(xAndY[-1][0])
+
+    return binEdges
+
+  @staticmethod
+  def _discretizationCAIM(X, y, possibleValuesX, possibleValuesY):
+    """
+    Applies the CAIM algorithm to discretize the values of x
+
+    Parameters
+    ----------
+        X: ndarray with shape (n,1) where n is the number of samples
+            Column-vector that contains all the data that needs to be discretized
+        y: ndarray with shape (n,1) where n is the number of samples
+            Column-vector that contains the class for each sample. This vector will not be discretized, but the class-value of each sample is needed to properly apply the algorithm
+        possibleValuesX: one dimensional ndarray
+            Contains all the possible values that x can take sorted in increasing order. There shouldn't be any doubles inside
+        possibleValuesY: one dimensional ndarray
+            Contains the possible values of y.
+    Returns
+    -------
+        list[float]
+          a list of the edges of the bins that are chosen by this algorithm
+    """
+    xAndY = numpy.concatenate((X, y), axis=1)
+    xAndY = xAndY[xAndY[:, 0].argsort()]
+    B = (possibleValuesX[1:] + possibleValuesX[:-1]) / 2
+    [class0, _] = possibleValuesY
+
+    binEdgesIndex = []
+    nbElementsByIntervalClass0 = numpy.zeros(len(B) + 1)
+    nbElementsByIntervalClass1 = numpy.zeros(len(B) + 1)
+    currentIntervalIndex = 0
+    for X in xAndY:
+      if currentIntervalIndex < len(B) and X[0] > B[currentIntervalIndex]:
+        currentIntervalIndex += 1
+      if X[1] == class0:
+        nbElementsByIntervalClass0[currentIntervalIndex] += 1
+      else:
+        nbElementsByIntervalClass1[currentIntervalIndex] += 1
+
+    Class0ByLargeInterval = [nbElementsByIntervalClass0.sum()]
+    Class1ByLargeInterval = [nbElementsByIntervalClass1.sum()]
+
+    k = 0
+    globalCAIM = 0.0
+    oldCaim = 0.0
+    while True:
+      caimMax = 0
+      maxPosition = 0
+      maxBoundaryIndex = 0
+      position = 0
+      currentSumClass0 = 0
+      currentSumClass1 = 0
+      maxLeftIntervalClass0 = currentSumClass0
+      maxLeftIntervalClass1 = currentSumClass1
+      maxRightIntervalClass0 = maxLeftIntervalClass0
+      maxRightIntervalClass1 = maxLeftIntervalClass1
+
+      for boundaryIndex in range(len(B)):
+        if position < len(binEdgesIndex) and boundaryIndex == binEdgesIndex[position]:
+          position += 1
+          if Class0ByLargeInterval[position] > Class1ByLargeInterval[position]:
+            oldCaim = globalCAIM * len(Class0ByLargeInterval) - math.pow(Class0ByLargeInterval[position], 2) / (
+              Class0ByLargeInterval[position] + Class1ByLargeInterval[position]
+            )
+          else:
+            oldCaim = globalCAIM * len(Class0ByLargeInterval) - math.pow(Class1ByLargeInterval[position], 2) / (
+              Class0ByLargeInterval[position] + Class1ByLargeInterval[position]
+            )
+          currentSumClass0 = 0
+          currentSumClass1 = 0
+          continue
+
+        currentSumClass0 += nbElementsByIntervalClass0[boundaryIndex]
+        currentSumClass1 += nbElementsByIntervalClass1[boundaryIndex]
+        caim = oldCaim
+
+        if currentSumClass0 > currentSumClass1:
+          caim = caim + math.pow(currentSumClass0, 2) / (currentSumClass0 + currentSumClass1)
+        else:
+          caim = caim + math.pow(currentSumClass1, 2) / (currentSumClass0 + currentSumClass1)
+
+        intervalClass0 = Class0ByLargeInterval[position] - currentSumClass0
+        intervalClass1 = Class1ByLargeInterval[position] - currentSumClass1
+
+        if intervalClass0 > intervalClass1:
+          caim = caim + math.pow(intervalClass0, 2) / (intervalClass0 + intervalClass1)
+        else:
+          caim = caim + math.pow(intervalClass1, 2) / (intervalClass0 + intervalClass1)
+
+        caim = caim / (len(Class0ByLargeInterval) + 1)
+
+        if caim > caimMax:
+          maxLeftIntervalClass0 = currentSumClass0
+          maxLeftIntervalClass1 = currentSumClass1
+          maxRightIntervalClass0 = intervalClass0
+          maxRightIntervalClass1 = intervalClass1
+          caimMax = caim
+          maxBoundaryIndex = boundaryIndex
+          maxPosition = position
+
+      if caimMax > globalCAIM:
+        globalCAIM = caimMax
+        binEdgesIndex.insert(maxPosition, maxBoundaryIndex)
+        Class0ByLargeInterval.insert(maxPosition + 1, maxRightIntervalClass0)
+        Class1ByLargeInterval.insert(maxPosition + 1, maxRightIntervalClass1)
+        Class0ByLargeInterval[maxPosition] = maxLeftIntervalClass0
+        Class1ByLargeInterval[maxPosition] = maxLeftIntervalClass1
+        k = k + 1
+        if Class0ByLargeInterval[0] > Class1ByLargeInterval[0]:
+          oldCaim = globalCAIM * len(Class0ByLargeInterval) - math.pow(Class0ByLargeInterval[0], 2) / (
+            Class0ByLargeInterval[0] + Class1ByLargeInterval[0]
+          )
+        else:
+          oldCaim = globalCAIM * len(Class0ByLargeInterval) - math.pow(Class1ByLargeInterval[0], 2) / (
+            Class0ByLargeInterval[0] + Class1ByLargeInterval[0]
+          )
+
+      else:
+        break
+
+    binEdges = [xAndY[0][0]]
+    for index in binEdgesIndex:
+      binEdges.append(B[index])
+    binEdges.append(xAndY[-1][0])
+
+    return binEdges
+
+  @staticmethod
+  def _discretizationNML(X, possibleValuesX, kMax=10, epsilon=None):
+    """
+    Uses the discretization algorithm described in "MDL Histogram Density Estimator", Kontkaken and Myllymaki, 2007 to
+    discretize.
+
+    Parameters
+    ----------
+        X: one dimensional ndarray
+            array that that contains all the data that needs to be discretized
+        possibleValuesX: one dimensional ndarray
+            Contains all the possible values that x can take sorted in increasing order. There shouldn't be any doubles
+            inside.
+        kMax: int
+            the maximum number of bins before the algorithm stops itself.
+        epsilon: float or None
+            the value of epsilon used in the algorithm. Should be as small as possible. If None is passed the value is
+            automatically calculated.
+    Returns
+    -------
+        List[float]
+          a list of the edges of the bins that are chosen by this algorithm
+    """
+    Xsorted = X[X.argsort(axis=None)]
+    if epsilon is None:
+      epsilon = numpy.amin(possibleValuesX[1:] - possibleValuesX[:-1]) / 2
+    epsilon = epsilon / 2
+    candidateCutPoints = numpy.zeros(2 * len(possibleValuesX))
+    for i in range(len(possibleValuesX)):
+      candidateCutPoints[2 * i] = possibleValuesX[i] - epsilon
+      candidateCutPoints[2 * i + 1] = possibleValuesX[i] + epsilon
+
+    E = len(candidateCutPoints)
+
+    binCount = numpy.zeros(E)
+    counter = 0
+    j = 0
+    for x in Xsorted:
+      while x > candidateCutPoints[j]:
+        binCount[j] = counter
+        j = j + 1
+      counter = counter + 1
+    binCount[j] = counter
+    n = binCount[-1]
+    Rkminus1 = numpy.ones(E)
+    Rk = numpy.zeros(E)
+    for i in range(1, E):
+      ne = int(binCount[i])
+      total_amount = 0
+      for h1 in range(ne + 1):
+        h2 = ne - h1
+        total_amount += math.pow(h1 / ne, h1) * math.pow(h2 / ne, h2) * scipy.special.comb(ne, h1)
+      Rk[i] = total_amount
+
+    k = 2
+
+    Bkminus1 = numpy.zeros(E)
+    for e in range(1, E):
+      ne = binCount[e]
+      Bkminus1[e] = -ne * (math.log(2 * epsilon * ne) - math.log(n * (candidateCutPoints[e] - candidateCutPoints[0])))
+
+    Bk = numpy.zeros(E)
+    cutpoints = [candidateCutPoints[0]]
+    Bvalues = [Bkminus1[-1]]
+    minimumeprime = 0
+    while k <= kMax:
+      for e in range(k, E):
+        minimum = math.inf
+        minimumeprime = 0
+        for eprime in range(k - 1, e):
+          if binCount[e] > binCount[eprime]:
+            temp = Bkminus1[eprime] - (binCount[e] - binCount[eprime]) * (
+              math.log(2 * epsilon * (binCount[e] - binCount[eprime]))
+              - math.log(n * (candidateCutPoints[e] - candidateCutPoints[eprime]))
+            )
+          else:
+            temp = Bkminus1[eprime]
+          temp = temp + math.log(Rk[e] / Rkminus1[eprime])
+          if minimum > temp:
+            minimum = temp
+            minimumeprime = eprime
+        minimum = minimum + math.log((E - k) / (k - 1))
+        Bk[e] = minimum
+      cutpoints.append(candidateCutPoints[minimumeprime])
+      Bvalues.append(Bk[-1])
+
+      k = k + 1
+      temp = Rk
+      Rk = Rk + numpy.multiply(binCount, Rkminus1) / (k - 2)
+      Rkminus1 = temp
+      Bkminus1 = Bk
+    minimum = math.inf
+    minimumIndex = 0
+    for k in range(1, len(Bvalues)):
+      if Bvalues[k] < minimum:
+        minimum = Bvalues[k]
+        minimumIndex = k
+    cutpoints = sorted(set(cutpoints[: minimumIndex + 1]))
+    cutpoints.append(candidateCutPoints[-1])
+
+    return cutpoints
+
+  def _createVariable(self, variableName, X, y=None, possibleValuesY=None):
+    """
+    Creates a variable for the column passed in as a parameter
+
+    Parameters
+    ----------
+        variableName:
+            the name of the created variable
+        X: ndarray shape(n,1)
+            A column vector containing n samples of a feature. The column for which the variable will be created
+        y: ndarray shape(n,1)
+            A column vector containing the corresponding for each element in X.
+        possibleValuesY: ndarray
+            An ndarray containing all the unique values of y
+    Returns
+    -------
+        pyagrum.DiscreteVariable
+            the created variable
+    """
+
+    if y is not None:
+      X, y = sklearn.utils.check_X_y(X, y, dtype=None, accept_sparse=True, ensure_2d=False)
+    X = sklearn.utils.check_array(X, dtype=None, ensure_2d=False)
+    try:
+      Xtransformed = sklearn.utils.check_array(X, dtype="float", ensure_2d=False)
+      isNumeric = True
+    except ValueError:
+      Xtransformed = X
+      isNumeric = False
+
+    foundValuesX = set(numpy.unique(X))
+    n = len(X)
+
+    if (
+      variableName not in self.discretizationParametersDictionary
+    ):  # The user has not manually set the discretization parameters for this variable
+      if (
+        isNumeric
+        and 1 <= self.discretizationThreshold < len(foundValuesX)
+        or (self.discretizationThreshold < 1 and len(foundValuesX) / len(X) > self.discretizationThreshold)
+      ):
+        self.discretizationParametersDictionary[variableName] = {}
+        self.discretizationParametersDictionary[variableName]["method"] = self.defaultMethod
+        self.discretizationParametersDictionary[variableName]["param"] = self.defaultParamDiscretizationMethod
+      else:
+        self.discretizationParametersDictionary[variableName] = {}
+        self.discretizationParametersDictionary[variableName]["method"] = "NoDiscretization"
+      usingDefaultParameters = True
+    else:  # The user has manually set the discretization parameters for this variable
+      usingDefaultParameters = False
+      if self.discretizationParametersDictionary[variableName]["method"] != "NoDiscretization" and not isNumeric:
+        raise ValueError(f"The variable {variableName} is not numeric and cannot be discretized!")
+
+    if self.discretizationParametersDictionary[variableName]["method"] == "NoDiscretization":
+      is_int_var = True
+
+      varSyntax = ""
+      if "param" in self.discretizationParametersDictionary[variableName]:
+        varSyntax = self.discretizationParametersDictionary[variableName]["param"]
+        if varSyntax is None:
+          varSyntax = ""
+
+      if varSyntax != "":
+        var = gum.fastVariable(variableName + varSyntax)
+        possibleValuesX = set(var.labels())
+        f = {str(x) for x in foundValuesX}
+        if not f.issubset(possibleValuesX):
+          raise ValueError(
+            f"The values passed in possibleValues ({sorted(possibleValuesX)}) do not match database values ("
+            f"{sorted(f)})"
+          )
+        return var
+
+      possibleValuesX = sorted(foundValuesX)
+      is_int_var = all(map(check_int, possibleValuesX))
+      if is_int_var:
+        possibleValuesX = [int(x) for x in possibleValuesX]
+        max_v = int(possibleValuesX[-1])  # sorted
+        min_v = int(possibleValuesX[0])
+
+        if len(possibleValuesX) == max_v - min_v + 1:  # no hole in the list of int
+          return gum.RangeVariable(variableName, variableName, min_v, max_v)
+        else:
+          return gum.IntegerVariable(variableName, variableName, possibleValuesX)
+
+      is_float_var = all(map(check_float, possibleValuesX))
+      if is_float_var:
+        possibleValuesX = [float(x) for x in possibleValuesX]
+        return gum.NumericalDiscreteVariable(variableName, variableName, possibleValuesX)
+      else:
+        return gum.LabelizedVariable(variableName, variableName, [str(v) for v in possibleValuesX])
+    else:
+      self.numberOfContinuous += 1
+      if self.discretizationParametersDictionary[variableName]["method"] == "expert":
+        binEdges = self.discretizationParametersDictionary[variableName]["param"]
+      elif self.discretizationParametersDictionary[variableName]["method"] == "CAIM":
+        if y is None:
+          raise ValueError(
+            "The CAIM discretization method requires a list of the associated classes for each data vector since it "
+            "is a supervised discretization method. You should pass it as y."
+          )
+        if possibleValuesY is None:
+          possibleValuesY = numpy.unique(y)
+        binEdges = self._discretizationCAIM(
+          Xtransformed.reshape(n, 1),
+          y.reshape(n, 1),
+          numpy.unique(Xtransformed),
+          possibleValuesY,
+        )
+      elif self.discretizationParametersDictionary[variableName]["method"] == "MDLP":
+        if y is None:
+          raise ValueError(
+            "The MDLP discretization method requires a list of the associated classes for each data vector since it "
+            "is a supervised discretization method. You should pass it as y."
+          )
+        if possibleValuesY is None:
+          possibleValuesY = numpy.unique(y)
+        binEdges = self._discretizationMDLP(
+          Xtransformed.reshape(n, 1),
+          y.reshape(n, 1),
+          numpy.unique(Xtransformed),
+          possibleValuesY,
+        )
+      elif self.discretizationParametersDictionary[variableName]["method"] == "NML":
+        binEdges = self._discretizationNML(
+          Xtransformed.flatten(),
+          numpy.unique(Xtransformed),
+          kMax=self.discretizationParametersDictionary[variableName]["param"],
+        )
+      else:
+        if self.discretizationParametersDictionary[variableName]["param"] == "elbowMethod":
+          binEdges = self._discretizationElbowMethodRotation(
+            self.discretizationParametersDictionary[variableName]["method"],
+            Xtransformed.flatten(),
+          )
+        else:
+          discre = skp.KBinsDiscretizer(
+            self.discretizationParametersDictionary[variableName]["param"],
+            strategy=self.discretizationParametersDictionary[variableName]["method"],
+            quantile_method="averaged_inverted_cdf",
+            subsample=None,
+          )
+          discre.fit(X.reshape(-1, 1))
+          binEdges = discre.bin_edges_[0].tolist()
+
+      if len(binEdges) == 2:
+        raise ValueError(
+          "Due to an error the discretization method "
+          + str(self.discretizationParametersDictionary[variableName]["method"])
+          + " using "
+          + str(self.discretizationParametersDictionary[variableName]["param"])
+          + " bins for the variable "
+          + str(variableName)
+          + "gave only 1 bin. Try increasing the number of bins used by this variable using "
+          "setDiscretizationParameters to avoid this error"
+        )
+
+      self.totalNumberOfBins += len(binEdges) - 1
+      var = gum.DiscretizedVariable(variableName, variableName, binEdges)
+      var.setEmpirical(True)
+
+    if usingDefaultParameters:
+      self.discretizationParametersDictionary.pop(variableName)
+
+    return var
+
+  @staticmethod
+  def _divideIntervalMDLP(
+    minimalValues,
+    shannonEntropyByLargeInterval,
+    Class0ByLargeInterval,
+    Class1ByLargeInterval,
+    continueDividingInterval,
+    totalCountByLargeInterval,
+    position,
+    binEdgesIndex,
+  ):
+    shannonEntropy = shannonEntropyByLargeInterval[position]
+
+    gain = shannonEntropy - minimalValues["classInformationEntropy"]
+
+    # all the 2's here should be replaced by the number of classes present in the interval. However we know that if
+    # the number of classes in the interval is equal to 1, then the shannon entropy will be 0 so the product of the 2
+    # will be 0.
+    deltaS = math.log2(7) - (
+      2 * shannonEntropy
+      - 2 * minimalValues["leftSubintervalShannonEntropy"]
+      - 2 * minimalValues["rightSubintervalShannonEntropy"]
+    )
+
+    if (
+      gain > (math.log2(totalCountByLargeInterval[position] - 1) + deltaS) / totalCountByLargeInterval[position]
+      or len(Class0ByLargeInterval) == 1
+    ):
+      binEdgesIndex.insert(position, minimalValues["boundaryIndex"])
+
+      Class0ByLargeInterval.insert(position + 1, minimalValues["rightSubintervalClass0"])
+      Class1ByLargeInterval.insert(position + 1, minimalValues["rightSubintervalClass1"])
+      continueDividingInterval.insert(position + 1, True)
+      totalCountByLargeInterval.insert(
+        position + 1,
+        minimalValues["rightSubintervalClass0"] + minimalValues["rightSubintervalClass1"],
+      )
+      shannonEntropyByLargeInterval.insert(position + 1, minimalValues["rightSubintervalShannonEntropy"])
+
+      Class0ByLargeInterval[position] = minimalValues["leftSubintervalClass0"]
+      Class1ByLargeInterval[position] = minimalValues["leftSubintervalClass1"]
+      totalCountByLargeInterval[position] = (
+        minimalValues["leftSubintervalClass0"] + minimalValues["leftSubintervalClass1"]
+      )
+      shannonEntropyByLargeInterval[position] = minimalValues["leftSubintervalShannonEntropy"]
+
+      # if the class information entropy is 0, then we have perfectly cut the interval so that a class only appears
+      # on one side, so we do not need to cut any further.
+      if minimalValues["leftSubintervalShannonEntropy"] == 0:
+        continueDividingInterval[position] = False
+      if minimalValues["rightSubintervalShannonEntropy"] == 0:
+        continueDividingInterval[position + 1] = False
+
+        # if there are no tensor boundary points left in this interval, we can't divide it any further
+      if position > 0 and minimalValues["boundaryIndex"] - 1 == binEdgesIndex[position - 1]:
+        continueDividingInterval[position] = False
+
+      if minimalValues["boundaryIndex"] == 0:
+        continueDividingInterval[position] = False
+
+      if position < len(binEdgesIndex) - 1 and binEdgesIndex[position] + 1 == binEdgesIndex[position + 1]:
+        continueDividingInterval[position + 1] = False
+      position += 1
+    else:
+      continueDividingInterval[position] = False
+
+  def discretizedTemplate(self, X, y=None, *, possibleValuesY=None, template=None):
+    """
+    return a graphical model discretized using the suggestion of the Discretized for date source X (and for target y).
+    This graphical model only contains the discretized variables.
+    For instance, it can be used as a template for a BNLearner.
+
+    Parameters
+    ----------
+        X: {array-like, sparse matrix, pandas or polars dataframe} of shape (n_samples, n_features)) or str (filename)
+            training data
+        y: array-like, pandas or polars dataframe of shape (n_samples,) or str (classname)
+            Target values
+        possibleValuesY: ndarray
+            An ndarray containing all the unique values of y
+        template: a graphical model such as pyagrum.BayesNet, pyagrum.MRF, etc...
+            the template that will contain the discretized variables.
+            If None, a new Bayesian network is created.
+
+    Returns
+    -------
+        pyagrum.BayesNet or other graphical model:
+          the discretized graphical model (only (discretized) random variables are created in the model)
+
+    Example
+    -------
+    >>> discretizer = DiscreteTypeProcessor(
+    ...   defaultDiscretizationMethod="uniform", defaultParamDiscretizationMethod=7, discretizationThreshold=10
+    ... )
+    >>> learner = gum.BNLearner(data, discretizer.discretizedTemplate(data))
+    """
+    if template is None:
+      template = gum.BayesNet()
+
+    if isinstance(X, str):
+      Xp = pandas.read_csv(X)
+    elif hasattr(X, "to_pandas"):
+      Xp = X.to_pandas()
+    else:
+      Xp = X
+
+    if isinstance(y, str):
+      yp = Xp[y]
+    elif y is not None and hasattr(y, "to_pandas"):
+      yp = y.to_pandas()
+    else:
+      yp = y
+
+    for name in Xp:
+      template.add(self._createVariable(name, Xp[name], yp, possibleValuesY))
+    return template