TAN.py

import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
import pandas as pd
import networkx as nx
import matplotlib.pyplot as plt
import numpy as np
import pyAgrum as gum
import pyAgrum.skbn as skbn
from sklearn.model_selection import KFold
import matplotlib

matplotlib.use('SVG')

def NB_TAN_k_fold_with_steps(jumpSteps, selection_parameter, dataset, class_var):

  """
    Performs TAN classifier with k-fold cross-validation and visualization of the step-wise process.

    Args:
        jumpSteps (int): Number of steps to jump before recalculating weights and scores.
        selection_parameter (str): Parameter for feature selection. Valid options: "Score" or "MI" (Mutual Information).
        dataset (str): Path to the dataset file.
        class_variable (str): The class variable of the classification problem.

    Returns:
        figures_list (list): List of tuples containing the figure, scores, and learned Bayesian network for each step.
        
  """

  importr('utils')
  importr('bnclassify')

  figures_list = []

  df_cars = pd.read_csv(dataset)

  from rpy2.robjects import pandas2ri
  from rpy2.robjects.conversion import localconverter
  with localconverter(robjects.default_converter + pandas2ri.converter):
    r_from_pd_df = robjects.conversion.py2rpy(df_cars)
  
  robjects.globalenv['r_from_pd_df'] = r_from_pd_df

  robjects.r('r_from_pd_df <- as.data.frame(unclass(r_from_pd_df), stringsAsFactors = TRUE)')

  robjects.globalenv['class_variable'] = class_var

  tan = robjects.r("bnc('nb',class_variable, r_from_pd_df, smooth= 0.01)")

  robjects.globalenv["tan"] = tan

  modelo = robjects.r('modelstring(tan)')[0]
 
  class_var = robjects.r('class_var(tan)')[0]

  features = list(robjects.r('features(tan)'))

  #Graph construction

  g = nx.DiGraph()
  g.add_node(class_var)
  g.add_nodes_from(features)

  familia = robjects.r('families(tan)')

  edges_class_var = []

  rest_edges = []

  for x in familia:
    for i in range(1,len(x)):
      g.add_edge(x[i],x[0])
      if(x[i] == class_var):
        edges_class_var.append((x[i],x[0]))
      else:
        rest_edges.append((x[i],x[0]))

  distance = 0
  fixed_distance = 0.25
  fixed_distance_y = 0.5
  features_pos = {}

  for f in features:
    features_pos[f] = (distance,0.25)
    distance += fixed_distance

  df = pd.read_csv(dataset)
  df2 = pd.DataFrame()

  a

  g.clear()

  def obtain_weight_features(features_list):
    weight_list = []
    for f in features_list:
      order = "cmi('"+f+"','"+class_var+"', r_from_pd_df)"
      weight = float(robjects.r(order)[0])
      weight_list.append(weight)
    return weight_list

  weight_list = obtain_weight_features(features)

  weight_list_aux = weight_list

  features_list_aux = features

  bn = gum.BayesNet()

  df2[class_var] = df[class_var]
  bn.add(class_var)
  g.add_node(class_var)

  nodes_added = []

  distance = 0
  fixed_distance = 0.25
  fixed_distance_y = 0.5
  features_pos = {}

  steps_aux = jumpSteps
  i=0

  if len(df[class_var].unique()) == 2:
    y = df[class_var].map({df[class_var].unique()[1]: True, df[class_var].unique()[0]: False})
  else:
    y = df[class_var]
  
  while weight_list_aux:
      
      index_max = np.argmax(weight_list_aux)
      feature = features_list_aux.pop(index_max)
      weight_list_aux.pop(index_max)
      
      bn.add(feature)
      df2[feature] = df[feature]
      bn.addArc(class_var,feature)
      g.add_edge(class_var, feature)
      nodes_added.append(feature)

      features_pos[feature] = (distance,0.25)
      distance += fixed_distance

      i+=1

      if steps_aux == jumpSteps or not weight_list_aux:
        steps_aux = 0
        

        fig = plt.figure()

        nx.draw_networkx_nodes(g,pos=features_pos,nodelist = nodes_added, node_size = 1500, margins= 0.2)

        nx.draw_networkx_nodes(g,pos={class_var: ((distance-fixed_distance)/2,fixed_distance_y)},
                          nodelist=[class_var],node_color='#009900',node_size = 1500)
        
        features_pos[class_var] = ((distance-fixed_distance)/2,fixed_distance_y)

        
        nx.draw_networkx_labels(g, pos = features_pos, font_weight='bold', font_size = 5.5)

        nx.draw_networkx_edges(g,features_pos,arrows = True, edgelist=[(class_var,node) for node in nodes_added],node_size = 1500)

        kf = KFold(n_splits= 4,shuffle=True)

        scores = []
        
        for k in kf.split(df2):

          df_train = df2.iloc[k[0]]
          learner=gum.BNLearner(df_train)
          learner.useSmoothingPrior()
          bn2 = learner.learnParameters(bn.dag())

          bnc=skbn.BNClassifier()
          bnc.fromTrainedModel(bn2,targetAttribute=class_var)
          yTest = y.iloc[k[1]]

          scoreCSV1 = bnc.score(df.iloc[k[1]], y = yTest)
          scores.append(scoreCSV1)

        figures_list.append((fig, scores))
        
      else:
        steps_aux+=1

      


  with localconverter(robjects.default_converter + pandas2ri.converter):
    r_from_pd_df = robjects.conversion.py2rpy(df2)

  robjects.globalenv['r_from_pd_df'] = r_from_pd_df

  robjects.r('r_from_pd_df <- as.data.frame(unclass(r_from_pd_df), stringsAsFactors = TRUE)')

  tan = robjects.r("bnc('tan_cl', class_variable , r_from_pd_df, smooth = 1, dag_args = list(score = 'aic'))")

  robjects.globalenv["tan"] = tan

  class_var = robjects.r('class_var(tan)')[0]

  features = list(robjects.r('features(tan)'))

  g = nx.DiGraph()
  g.add_node(class_var)
  g.add_nodes_from(features)

  familia = robjects.r('families(tan)')

  edges_class_var = []

  rest_edges = []

  for x in familia:
    for i in range(1,len(x)):
      g.add_edge(x[i],x[0])
      if(x[i] == class_var):
        edges_class_var.append((x[i],x[0]))
      else:
        rest_edges.append((x[i],x[0]))

  def obtain_weight_edges(edges_list):
    weight_list = []
    for e in edges_list:
      order = "cmi('"+e[0]+"','"+e[1]+"', z = '"+class_var+"', r_from_pd_df)"
      weight = float(robjects.r(order)[0])
      weight_list.append(weight)
    return weight_list

  edges = [k for k in g.edges if k[0]!=class_var]

  edges_weights = obtain_weight_edges(edges)

  index_max = np.argmax(edges_weights)

  edges_weights_aux = edges_weights

  #Añadir las aristas del Naive Bayes y crear la red bayesiana
  g.clear_edges()

  bn = gum.BayesNet()

  for i in df2.columns:
      #print(i)
      bn.add(i)

  for i in df2.columns:
      if(i != class_var):
          g.add_edge(class_var,i)
          bn.addArc(class_var,i)

  edges_added = []

  steps_aux = jumpSteps

  if len(df[class_var].unique()) == 2:
    y = df[class_var].map({df[class_var].unique()[1]: True, df[class_var].unique()[0]: False})
  else:
    y = df[class_var]

  while edges_weights_aux:
      index_max = np.argmax(edges_weights_aux)
      edge = edges.pop(index_max)
      edges_weights_aux.pop(index_max)
      bn.addArc(*edge)
      g.add_edge(*edge)
      edges_added.append(edge)
      #Draw net

      if steps_aux == jumpSteps or not edges_weights_aux:

        steps_aux = 0

        fig = plt.figure()

        nx.draw_networkx_nodes(g,pos=features_pos,nodelist = features, node_size = 1500, margins= 0.2)

        nx.draw_networkx_nodes(g,pos={class_var: ((distance-fixed_distance)/2,fixed_distance_y)},
                          nodelist=[class_var],node_color='#009900',node_size = 1500)
        
        nx.draw_networkx_labels(g, pos = features_pos, font_weight='bold', font_size = 5.5)

        nx.draw_networkx_edges(g,features_pos,arrows = True, edgelist=edges_class_var,node_size = 1500)

        nx.draw_networkx_edges(g,features_pos,arrows = True, edgelist= edges_added, connectionstyle='arc3,rad=0.4',node_size = 1500)


        #plt.show()

        kf = KFold(n_splits= 4,shuffle=True)

        scores = []
        
        for k in kf.split(df2):

          df_train = df2.iloc[k[0]]

          learner=gum.BNLearner(df_train) 
          learner.useSmoothingPrior()
          bn2 = learner.learnParameters(bn.dag())

          bnc=skbn.BNClassifier()
          bnc.fromTrainedModel(bn2,targetAttribute=class_var)
          yTest = y.iloc[k[1]]

          scoreCSV1 = bnc.score(df.iloc[k[1]], y =yTest)
          #print("{0:.2f}% good predictions".format(100*scoreCSV1))
          scores.append(scoreCSV1)

        figures_list.append((fig, scores, bn2))
        #bnc.showROC_PR('datos/cars_discrete.csv')
      
      else:
        steps_aux+=1

  return figures_list