¡Cotiza Ahora!

Clasificación de clientes para identificar la deserción (CHURN)

A continuación muestro un ejemplo de Clasificación de Clientes para Identificar la Deserción de clientes lo que tambien se conoce como CHURN.

Estos ejemplos / casos / tareas son tomados del curso de Redes Neuronales Artificiales en los Negocios de la Maestria en Ciencia de Datos de la Universidad Ricardo Palma a cargo del Ing. Mirko Rodriguez.

Para este ejemplo usaremos:

  • Red Neuronal con Keras y TensorFlow
  • Tipo de Aprendizaje Supervisado
  • Algoritmo de Aprendizaje Adam

Todo el código mostrado a continuación a sido probado en Google Colab sobre Python 3.X

https://drive.google.com/file/d/1_bwB_FGFbcMP7qaCnzaSdnatmnd6ReJV/view?usp=sharing

Empezamos descargando la data en csv hosteado en fragote.com que utilizaremos:

1
2
3
4
5
6
%%bash
if [ ! -f "clientes_data.csv" ]; then
    wget www.fragote.com/data/clientes_data.csv
fi
 
ls -l

Agregamos la funciones necesarias para poder hacer nuestro modelo predictivo / análisis:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
# Funciones
 
import numpy as np
 
from sklearn.metrics import confusion_matrix
from sklearn.utils.multiclass import unique_labels
from sklearn.externals import joblib
 
import seaborn as sns
import matplotlib.pyplot as plt
 
def plot_confusion_matrix(y_true, y_pred,
                          normalize=False,
                          title=None):
    """
    Esta función imprime y traza la matriz de confusión.
     La normalización se puede aplicar configurando `normalize=True`.
    """
    if not title:
        if normalize:
            title = 'Matriz de Confusión Normalizada'
        else:
            title = 'Matriz de Confusión sin Normalizar'
 
    # Calculando la Matriz de Confusion
    cm = confusion_matrix(y_true, y_pred)
    # solo usar las etiquetas que se tienen en la data
    classes = unique_labels(y_true, y_pred)
    if normalize:
        cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
        print("Matriz de Confusión Normalizada")
    else:
        print('Matriz de Confusión sin Normalizar')
 
    print(cm)
 
    fig, ax = plt.subplots()
    im = ax.imshow(cm, interpolation='nearest', cmap=plt.cm.Blues)
    ax.figure.colorbar(im, ax=ax)
    ax.grid(linewidth=.0)
    # Queremos mostrar todos los puntos...
    ax.set(xticks=np.arange(cm.shape[1]),
           yticks=np.arange(cm.shape[0]),
           # ... etiquetando la lista de datos
           xticklabels=classes, yticklabels=classes,
           title=title,
           ylabel='True label',
           xlabel='Predicted label')
 
    # rotando las etiquedas de los puntos.
    plt.setp(ax.get_xticklabels(), rotation=45, ha="right",rotation_mode="anchor")
 
    # Loop over data dimensions and create text annotations.
    fmt = '.2f' if normalize else 'd'
    thresh = cm.max() / 2.
    for i in range(cm.shape[0]):
        for j in range(cm.shape[1]):
            ax.text(j, i, format(cm[i, j], fmt),
                    ha="center", va="center",
                    color="white" if cm[i, j] > thresh else "black")
    fig.tight_layout()
    plt.show()
    return ax
 
def saveFile(object_to_save, scaler_filename):
    joblib.dump(object_to_save, scaler_filename)
 
def loadFile(scaler_filename):
    return joblib.load(scaler_filename)
 
def plotHistogram(dataset_final):
    dataset_final.hist(figsize=(20,14), edgecolor="black", bins=40)
    plt.show()
 
def plotCorrelations(dataset_final):
    fig, ax = plt.subplots(figsize=(10,8))   # size in inches
    g = sns.heatmap(dataset_final.corr(), annot=True, cmap="YlGnBu", ax=ax)
    g.set_yticklabels(g.get_yticklabels(), rotation = 0)
    g.set_xticklabels(g.get_xticklabels(), rotation = 45)
    fig.tight_layout()
    plt.show()

Preprocesando los datos del archivo csv:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
# Importando librerías
import pandas as pd
import numpy as np
from sklearn.preprocessing import LabelEncoder
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
 
# Importando Datasets
dataset_csv = pd.read_csv('clientes_data.csv')
 
# Columnas de la data
print ("\nColumnas del DataSet: ")
print (dataset_csv.columns)
 
# Describir la data original
print ("\nDataset original:\n", dataset_csv.describe(include='all'))
 
# Dataset reducido
dataset = dataset_csv.iloc[:,3:14]
dataset_columns = dataset.columns
dataset_values = dataset.values
 
# Describir la data truncada
print ("\nDataset reducido: ")
print("\n",dataset.head())
 
# Revisamos los tipos de datos de las Columnas
print ("\nTipos de Columnas del Dataset: ")
print(dataset.dtypes)

Columnas del DataSet: 
Index(['NumeroFila', 'DNI', 'Apellido', 'ScoreCrediticio', 'Pais', 'Genero',
       'edad', 'Tenure', 'Balance', 'NumDeProducts', 'TieneTarjetaCredito',
       'EsMiembroActivo', 'SalarioEstimado', 'Abandono'],
      dtype='object')

Dataset original:
          NumeroFila           DNI  ... SalarioEstimado      Abandono
count   10000.00000  1.000000e+04  ...    10000.000000  10000.000000
unique          NaN           NaN  ...             NaN           NaN
top             NaN           NaN  ...             NaN           NaN
freq            NaN           NaN  ...             NaN           NaN
mean     5000.50000  1.569094e+07  ...   100090.239881      0.203700
std      2886.89568  7.193619e+04  ...    57510.492818      0.402769
min         1.00000  1.556570e+07  ...       11.580000      0.000000
25%      2500.75000  1.562853e+07  ...    51002.110000      0.000000
50%      5000.50000  1.569074e+07  ...   100193.915000      0.000000
75%      7500.25000  1.575323e+07  ...   149388.247500      0.000000
max     10000.00000  1.581569e+07  ...   199992.480000      1.000000

[11 rows x 14 columns]

Dataset reducido: 

    ScoreCrediticio    Pais  Genero  ...  EsMiembroActivo  SalarioEstimado  Abandono
0              619  France  Female  ...                1        101348.88         1
1              608   Spain  Female  ...                1        112542.58         0
2              502  France  Female  ...                0        113931.57         1
3              699  France  Female  ...                0         93826.63         0
4              850   Spain  Female  ...                1         79084.10         0

[5 rows x 11 columns]

Tipos de Columnas del Dataset: 
ScoreCrediticio          int64
Pais                    object
Genero                  object
edad                     int64
Tenure                   int64
Balance                float64
NumDeProducts            int64
TieneTarjetaCredito      int64
EsMiembroActivo          int64
SalarioEstimado        float64
Abandono                 int64
dtype: object

Codificamos las columnas de caracteres a números, Genero / Pais

1
2
3
4
5
6
#Codificando datos categóricos:
labelEncoder_X_1 = LabelEncoder()
dataset_values[:, 1] = labelEncoder_X_1.fit_transform(dataset_values[:, 1])
labelEncoder_X_2 = LabelEncoder()
dataset_values[:, 2] = labelEncoder_X_2.fit_transform(dataset_values[:, 2])
print ("\nDataset Categorizado: \n", dataset_values)
1
2
3
4
5
6
7
8
Dataset Categorizado:
 [[619 0 0 ... 1 101348.88 1]
 [608 2 0 ... 1 112542.58 0]
 [502 0 0 ... 0 113931.57 1]
 ...
 [709 0 0 ... 1 42085.58 1]
 [772 1 1 ... 0 92888.52 1]
 [792 0 0 ... 0 38190.78 0]]

Escalando y normalizando los features  (StandardScaler: (x-u)/s): mean = 0 and standard deviation = 1

1
2
3
4
5
6
7
8
9
10
# Escalamiento/Normalización de Features (StandardScaler: (x-u)/s)
stdScaler = StandardScaler()
dataset_values[:,0:10] = stdScaler.fit_transform(dataset_values[:,0:10])
 
 
# Dataset final normalizado
dataset_final = pd.DataFrame(dataset_values,columns=dataset_columns, dtype = np.float64)
print("\nDataset Final:")
print(dataset_final.describe(include='all'))
print("\n", dataset_final.head())
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Dataset Final:
       ScoreCrediticio          Pais  ...  SalarioEstimado      Abandono
count     1.000000e+04  1.000000e+04  ...     1.000000e+04  10000.000000
mean     -4.870326e-16  5.266676e-16  ...    -1.580958e-17      0.203700
std       1.000050e+00  1.000050e+00  ...     1.000050e+00      0.402769
min      -3.109504e+00 -9.018862e-01  ...    -1.740268e+00      0.000000
25%      -6.883586e-01 -9.018862e-01  ...    -8.535935e-01      0.000000
50%       1.522218e-02 -9.018862e-01  ...     1.802807e-03      0.000000
75%       6.981094e-01  3.065906e-01  ...     8.572431e-01      0.000000
max       2.063884e+00  1.515067e+00  ...     1.737200e+00      1.000000
 
[8 rows x 11 columns]
 
    ScoreCrediticio      Pais  ...  SalarioEstimado  Abandono
0        -0.326221 -0.901886  ...         0.021886       1.0
1        -0.440036  1.515067  ...         0.216534       0.0
2        -1.536794 -0.901886  ...         0.240687       1.0
3         0.501521 -0.901886  ...        -0.108918       0.0
4         2.063884  1.515067  ...        -0.365276       0.0
 
[5 rows x 11 columns]

Graficamos la Distribution de los datos y la matriz de Correlación:

1
2
3
4
5
6
# Distribuciones de la data y Correlaciones
print("\n Histogramas:")
plotHistogram(dataset_final)
 
print("\n Correlaciones:")
plotCorrelations(dataset_final)
1
Histogramas:
1
Correlaciones:

Dividiendo data en conjuntos de Entrenamiento y Prueba en 80% para entrenamientos y 20% para pruebas:

1
2
3
4
5
6
# Obteniendo valores a procesar
X = dataset_final.iloc[:, 0:10].values
y = dataset_final.iloc[:, 10].values
 
# Dividiendo el Dataset en sets de Training y Test
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size = 0.2, random_state = 0)

Despues que ya preparamos la data generamos la red neuronal:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
# Importando Keras y Tensorflow
from keras.models import Sequential
from keras.layers import Dense
from keras.initializers import RandomUniform
 
# Inicializando la Red Neuronal
neural_network = Sequential()
 
# kernel_initializer Define la forma como se asignará los Pesos iniciales Wi
initial_weights = RandomUniform(minval = -0.5, maxval = 0.5)
 
# Agregado la Capa de entrada y la primera capa oculta
# 10 Neuronas en la capa de entrada y 8 Neuronas en la primera capa oculta
neural_network.add(Dense(units = 8, kernel_initializer = initial_weights, activation = 'relu', input_dim = 10))
 
# Agregando capa oculta
neural_network.add(Dense(units = 5, kernel_initializer = initial_weights, activation = 'relu'))
 
# Agregando capa oculta
neural_network.add(Dense(units = 4, kernel_initializer = initial_weights, activation = 'relu'))
 
# Agregando capa de salida
neural_network.add(Dense(units = 1, kernel_initializer = initial_weights, activation = 'sigmoid'))

Hemos generado una capa de entrada con 10 neuronas, 8 neuronas para la primera capa oculta, 5 neuronas para la segunda capa oculta, 4 para le tercera capa oculta y la capa de salida con 1 neurona.

Vemos la arquitectura de la red:

1
2
# Imprimir Arquitectura de la Red
neural_network.summary()
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 8)                 88        
_________________________________________________________________
dense_2 (Dense)              (None, 5)                 45        
_________________________________________________________________
dense_3 (Dense)              (None, 4)                 24        
_________________________________________________________________
dense_4 (Dense)              (None, 1)                 5         
=================================================================
Total params: 162
Trainable params: 162
Non-trainable params: 0
_________________________________________________________________

Compilando y Entrenando el modelo utilizando 100 epocas:

1
2
3
4
5
6
7
8
# Compilando la Red Neuronal
# optimizer: Algoritmo de optimización | binary_crossentropy = 2 Classes
# loss: error
neural_network.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
 
 
# Entrenamiento
neural_network.fit(X_train, y_train, batch_size = 32, epochs = 100)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:1033: The name tf.assign_add is deprecated. Please use tf.compat.v1.assign_add instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:1020: The name tf.assign is deprecated. Please use tf.compat.v1.assign instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:3005: The name tf.Session is deprecated. Please use tf.compat.v1.Session instead.
 
Epoch 1/100
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:190: The name tf.get_default_session is deprecated. Please use tf.compat.v1.get_default_session instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:197: The name tf.ConfigProto is deprecated. Please use tf.compat.v1.ConfigProto instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:207: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:216: The name tf.is_variable_initialized is deprecated. Please use tf.compat.v1.is_variable_initialized instead.
 
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/keras/backend/tensorflow_backend.py:223: The name tf.variables_initializer is deprecated. Please use tf.compat.v1.variables_initializer instead.
 
8000/8000 [==============================] - 7s 832us/step - loss: 0.5679 - acc: 0.7696
Epoch 2/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.4454 - acc: 0.8071
Epoch 3/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.4205 - acc: 0.8233
Epoch 4/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.4033 - acc: 0.8350
Epoch 5/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3872 - acc: 0.8442
Epoch 6/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3743 - acc: 0.8485
Epoch 7/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3659 - acc: 0.8480
Epoch 8/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3612 - acc: 0.8498
Epoch 9/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3585 - acc: 0.8520
Epoch 10/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3564 - acc: 0.8524
Epoch 11/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3548 - acc: 0.8538
Epoch 12/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3537 - acc: 0.8548
Epoch 13/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.3531 - acc: 0.8533
Epoch 14/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3521 - acc: 0.8556
Epoch 15/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3513 - acc: 0.8536
Epoch 16/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3510 - acc: 0.8551
Epoch 17/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3497 - acc: 0.8561
Epoch 18/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3493 - acc: 0.8553
Epoch 19/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3491 - acc: 0.8562
Epoch 20/100
8000/8000 [==============================] - 1s 154us/step - loss: 0.3477 - acc: 0.8559
Epoch 21/100
8000/8000 [==============================] - 1s 178us/step - loss: 0.3474 - acc: 0.8562
Epoch 22/100
8000/8000 [==============================] - 1s 181us/step - loss: 0.3461 - acc: 0.8580
Epoch 23/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3454 - acc: 0.8558
Epoch 24/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3444 - acc: 0.8569
Epoch 25/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3428 - acc: 0.8582
Epoch 26/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3425 - acc: 0.8574
Epoch 27/100
8000/8000 [==============================] - 2s 194us/step - loss: 0.3410 - acc: 0.8591
Epoch 28/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3403 - acc: 0.8600
Epoch 29/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3396 - acc: 0.8613
Epoch 30/100
8000/8000 [==============================] - 2s 192us/step - loss: 0.3385 - acc: 0.8611
Epoch 31/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3378 - acc: 0.8609
Epoch 32/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3371 - acc: 0.8611
Epoch 33/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3366 - acc: 0.8608
Epoch 34/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3361 - acc: 0.8618
Epoch 35/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3361 - acc: 0.8621
Epoch 36/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3360 - acc: 0.8626
Epoch 37/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3353 - acc: 0.8599
Epoch 38/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3354 - acc: 0.8629
Epoch 39/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3353 - acc: 0.8622
Epoch 40/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3345 - acc: 0.8619
Epoch 41/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3343 - acc: 0.8631
Epoch 42/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3348 - acc: 0.8619
Epoch 43/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3340 - acc: 0.8606
Epoch 44/100
8000/8000 [==============================] - 1s 183us/step - loss: 0.3342 - acc: 0.8621
Epoch 45/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3335 - acc: 0.8631
Epoch 46/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3339 - acc: 0.8624
Epoch 47/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3333 - acc: 0.8630
Epoch 48/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3336 - acc: 0.8611
Epoch 49/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3331 - acc: 0.8630
Epoch 50/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3333 - acc: 0.8635
Epoch 51/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3331 - acc: 0.8629
Epoch 52/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3333 - acc: 0.8609
Epoch 53/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3327 - acc: 0.8625
Epoch 54/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3325 - acc: 0.8624
Epoch 55/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3327 - acc: 0.8624
Epoch 56/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3325 - acc: 0.8629
Epoch 57/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3330 - acc: 0.8619
Epoch 58/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3326 - acc: 0.8625
Epoch 59/100
8000/8000 [==============================] - 1s 170us/step - loss: 0.3327 - acc: 0.8624
Epoch 60/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3322 - acc: 0.8638
Epoch 61/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3323 - acc: 0.8635
Epoch 62/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3318 - acc: 0.8618
Epoch 63/100
8000/8000 [==============================] - 1s 187us/step - loss: 0.3318 - acc: 0.8640
Epoch 64/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3315 - acc: 0.8640
Epoch 65/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3317 - acc: 0.8650
Epoch 66/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3318 - acc: 0.8625
Epoch 67/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3319 - acc: 0.8629
Epoch 68/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3317 - acc: 0.8631
Epoch 69/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3318 - acc: 0.8649
Epoch 70/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3316 - acc: 0.8634
Epoch 71/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3313 - acc: 0.8636
Epoch 72/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3317 - acc: 0.8642
Epoch 73/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.3310 - acc: 0.8626
Epoch 74/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3317 - acc: 0.8650
Epoch 75/100
8000/8000 [==============================] - 1s 184us/step - loss: 0.3320 - acc: 0.8629
Epoch 76/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3315 - acc: 0.8644
Epoch 77/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3313 - acc: 0.8636
Epoch 78/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3316 - acc: 0.8635
Epoch 79/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3315 - acc: 0.8639
Epoch 80/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3311 - acc: 0.8625
Epoch 81/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3310 - acc: 0.8638
Epoch 82/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3308 - acc: 0.8638
Epoch 83/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3310 - acc: 0.8645
Epoch 84/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3307 - acc: 0.8645
Epoch 85/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3307 - acc: 0.8626
Epoch 86/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3304 - acc: 0.8624
Epoch 87/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3308 - acc: 0.8658
Epoch 88/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3306 - acc: 0.8651
Epoch 89/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3305 - acc: 0.8645
Epoch 90/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3300 - acc: 0.8650
Epoch 91/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3308 - acc: 0.8640
Epoch 92/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3305 - acc: 0.8634
Epoch 93/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3304 - acc: 0.8647
Epoch 94/100
8000/8000 [==============================] - 2s 192us/step - loss: 0.3301 - acc: 0.8647
Epoch 95/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3302 - acc: 0.8634
Epoch 96/100
8000/8000 [==============================] - 1s 178us/step - loss: 0.3307 - acc: 0.8644
Epoch 97/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3306 - acc: 0.8636
Epoch 98/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3308 - acc: 0.8635
Epoch 99/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.3303 - acc: 0.8632
Epoch 100/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3299 - acc: 0.8640
<keras.callbacks.History at 0x7f4c83f1f5c0>

Después de las 100 épocas se puede observar que el modelo tiene un Accuracy de 86.4%

Hacemos la predicción con nuestros 50 primeros datos de prueba tomando a los mayores de 0.5 como abandono y a los menores o iguales como no abandono.

1
2
3
4
5
6
7
8
9
10
11
12
# Haciendo predicción de los resultados del Test
y_pred = neural_network.predict(X_test)
y_pred_norm = (y_pred > 0.5)
 
y_pred_norm = y_pred_norm.astype(int)
y_test = y_test.astype(int)
 
# 50 primeros resultados a comparar
print("\nPredicciones (50 primeros):")
print("\n\tReal", "\t", "Predicción(N)","\t", "Predicción(O)")
for i in range(50):
    print(i, '\t', y_test[i], '\t ', y_pred_norm[i], '\t \t', y_pred[i])
Predicciones (50 primeros):

	Real 	 Predicción(N) 	 Predicción(O)
0 	 0 	  [0] 	 	 [0.25591797]
1 	 1 	  [0] 	 	 [0.31653506]
2 	 0 	  [0] 	 	 [0.16617256]
3 	 0 	  [0] 	 	 [0.07125753]
4 	 0 	  [0] 	 	 [0.06709316]
5 	 1 	  [1] 	 	 [0.9561175]
6 	 0 	  [0] 	 	 [0.01982319]
7 	 0 	  [0] 	 	 [0.0801284]
8 	 1 	  [0] 	 	 [0.16028535]
9 	 1 	  [1] 	 	 [0.76239324]
10 	 0 	  [0] 	 	 [0.03265542]
11 	 0 	  [0] 	 	 [0.21288738]
12 	 0 	  [0] 	 	 [0.29622465]
13 	 0 	  [0] 	 	 [0.28117135]
14 	 1 	  [1] 	 	 [0.6112473]
15 	 1 	  [0] 	 	 [0.3089583]
16 	 0 	  [0] 	 	 [0.08254933]
17 	 0 	  [0] 	 	 [0.07191664]
18 	 0 	  [0] 	 	 [0.08332911]
19 	 0 	  [0] 	 	 [0.10069495]
20 	 0 	  [1] 	 	 [0.5985567]
21 	 0 	  [0] 	 	 [0.00851977]
22 	 0 	  [0] 	 	 [0.04273018]
23 	 0 	  [0] 	 	 [0.0951266]
24 	 0 	  [0] 	 	 [0.00786319]
25 	 0 	  [0] 	 	 [0.16710573]
26 	 0 	  [0] 	 	 [0.12770015]
27 	 0 	  [0] 	 	 [0.03628084]
28 	 0 	  [0] 	 	 [0.35433203]
29 	 0 	  [0] 	 	 [0.35582292]
30 	 0 	  [0] 	 	 [0.02590874]
31 	 1 	  [0] 	 	 [0.4530319]
32 	 0 	  [0] 	 	 [0.10775042]
33 	 0 	  [0] 	 	 [0.02804729]
34 	 0 	  [0] 	 	 [0.32664955]
35 	 0 	  [0] 	 	 [0.02045524]
36 	 0 	  [0] 	 	 [0.01842964]
37 	 0 	  [0] 	 	 [0.02650151]
38 	 0 	  [0] 	 	 [0.02766502]
39 	 0 	  [0] 	 	 [0.1354188]
40 	 0 	  [0] 	 	 [0.12443781]
41 	 1 	  [0] 	 	 [0.42804298]
42 	 1 	  [0] 	 	 [0.46373382]
43 	 0 	  [0] 	 	 [0.04600218]
44 	 0 	  [1] 	 	 [0.6296924]
45 	 0 	  [0] 	 	 [0.05090821]
46 	 0 	  [0] 	 	 [0.3721879]
47 	 0 	  [0] 	 	 [0.12022552]
48 	 0 	  [1] 	 	 [0.5840349]
49 	 0 	  [0] 	 	 [0.00435194]

Aplicando la Matriz de Confusión:

1
2
3
4
# Aplicando la Matriz de Confusión
from sklearn.metrics import confusion_matrix
cm = confusion_matrix(y_test, y_pred_norm)
print ("\nMatriz de Confusión: \n", cm)
1
2
3
Matriz de Confusión:
 [[1499   96]
 [ 196  209]]

Graficando la Matriz de Confusión:

1
plot_confusion_matrix(y_test, y_pred_norm, normalize=False,title="Matriz de Confusión: Abandono de clientes")
Matriz de Confusión sin Normalizar
[[1499   96]
 [ 196  209]]

Agregamos una nueva arquitectura:

Entrada => 10
Oculta => 8/4
Salida => 1

1
2
3
4
5
6
7
8
9
10
# Inicializando la Red Neuronal
neural_network2 = Sequential()
# kernel_initializer Define la forma como se asignará los Pesos iniciales Wi
initial_weights = RandomUniform(minval = -0.5, maxval = 0.5)
neural_network2.add(Dense(units = 8, kernel_initializer = initial_weights, activation = 'relu', input_dim = 10))
neural_network2.add(Dense(units = 4, kernel_initializer = initial_weights, activation = 'relu'))
# Agregando capa de salida
neural_network2.add(Dense(units = 1, kernel_initializer = initial_weights, activation = 'sigmoid'))
# Imprimir Arquitectura de la Red
neural_network2.summary()
Model: "sequential_2"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_5 (Dense)              (None, 8)                 88        
_________________________________________________________________
dense_6 (Dense)              (None, 4)                 36        
_________________________________________________________________
dense_7 (Dense)              (None, 1)                 5         
=================================================================
Total params: 129
Trainable params: 129
Non-trainable params: 0
_________________________________________________________________

Entrenando la El Modelo 2

1
2
3
4
5
6
7
8
# Compilando la Red Neuronal
# optimizer: Algoritmo de optimización | binary_crossentropy = 2 Classes
# loss: error
neural_network2.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
 
 
# Entrenamiento
neural_network2.fit(X_train, y_train, batch_size = 32, epochs = 100)
Epoch 1/100
8000/8000 [==============================] - 2s 208us/step - loss: 0.6309 - acc: 0.7251
Epoch 2/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.4994 - acc: 0.7960
Epoch 3/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.4579 - acc: 0.7962
Epoch 4/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.4301 - acc: 0.8054
Epoch 5/100
8000/8000 [==============================] - 1s 152us/step - loss: 0.4111 - acc: 0.8207
Epoch 6/100
8000/8000 [==============================] - 1s 142us/step - loss: 0.3908 - acc: 0.8379
Epoch 7/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3766 - acc: 0.8441
Epoch 8/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3682 - acc: 0.8490
Epoch 9/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3637 - acc: 0.8508
Epoch 10/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3604 - acc: 0.8521
Epoch 11/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3585 - acc: 0.8535
Epoch 12/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3569 - acc: 0.8544
Epoch 13/100
8000/8000 [==============================] - 1s 158us/step - loss: 0.3557 - acc: 0.8538
Epoch 14/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3547 - acc: 0.8556
Epoch 15/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3541 - acc: 0.8555
Epoch 16/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3533 - acc: 0.8568
Epoch 17/100
8000/8000 [==============================] - 1s 173us/step - loss: 0.3527 - acc: 0.8568
Epoch 18/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3519 - acc: 0.8576
Epoch 19/100
8000/8000 [==============================] - 1s 146us/step - loss: 0.3516 - acc: 0.8581
Epoch 20/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3509 - acc: 0.8584
Epoch 21/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3508 - acc: 0.8575
Epoch 22/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3502 - acc: 0.8586
Epoch 23/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3500 - acc: 0.8585
Epoch 24/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3497 - acc: 0.8579
Epoch 25/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3493 - acc: 0.8591
Epoch 26/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3486 - acc: 0.8569
Epoch 27/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3489 - acc: 0.8579
Epoch 28/100
8000/8000 [==============================] - 1s 158us/step - loss: 0.3486 - acc: 0.8595
Epoch 29/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3483 - acc: 0.8582
Epoch 30/100
8000/8000 [==============================] - 1s 152us/step - loss: 0.3484 - acc: 0.8575
Epoch 31/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3476 - acc: 0.8590
Epoch 32/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3478 - acc: 0.8585
Epoch 33/100
8000/8000 [==============================] - 1s 150us/step - loss: 0.3474 - acc: 0.8578
Epoch 34/100
8000/8000 [==============================] - 1s 158us/step - loss: 0.3472 - acc: 0.8592
Epoch 35/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3468 - acc: 0.8595
Epoch 36/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3467 - acc: 0.8585
Epoch 37/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3467 - acc: 0.8590
Epoch 38/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3462 - acc: 0.8579
Epoch 39/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3462 - acc: 0.8586
Epoch 40/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3458 - acc: 0.8587
Epoch 41/100
8000/8000 [==============================] - 1s 154us/step - loss: 0.3458 - acc: 0.8598
Epoch 42/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3462 - acc: 0.8587
Epoch 43/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3454 - acc: 0.8578
Epoch 44/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3455 - acc: 0.8581
Epoch 45/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3452 - acc: 0.8582
Epoch 46/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3454 - acc: 0.8579
Epoch 47/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3453 - acc: 0.8584
Epoch 48/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3452 - acc: 0.8587
Epoch 49/100
8000/8000 [==============================] - 1s 154us/step - loss: 0.3450 - acc: 0.8582
Epoch 50/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3447 - acc: 0.8591
Epoch 51/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3448 - acc: 0.8586
Epoch 52/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3445 - acc: 0.8596
Epoch 53/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3443 - acc: 0.8579
Epoch 54/100
8000/8000 [==============================] - 1s 147us/step - loss: 0.3441 - acc: 0.8586
Epoch 55/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3439 - acc: 0.8590
Epoch 56/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3436 - acc: 0.8608
Epoch 57/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3439 - acc: 0.8596
Epoch 58/100
8000/8000 [==============================] - 1s 154us/step - loss: 0.3435 - acc: 0.8591
Epoch 59/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3433 - acc: 0.8594
Epoch 60/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3426 - acc: 0.8603
Epoch 61/100
8000/8000 [==============================] - 1s 148us/step - loss: 0.3429 - acc: 0.8595
Epoch 62/100
8000/8000 [==============================] - 1s 151us/step - loss: 0.3427 - acc: 0.8609
Epoch 63/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3426 - acc: 0.8606
Epoch 64/100
8000/8000 [==============================] - 1s 151us/step - loss: 0.3424 - acc: 0.8608
Epoch 65/100
8000/8000 [==============================] - 1s 151us/step - loss: 0.3420 - acc: 0.8608
Epoch 66/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3423 - acc: 0.8610
Epoch 67/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3417 - acc: 0.8598
Epoch 68/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3418 - acc: 0.8621
Epoch 69/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3416 - acc: 0.8603
Epoch 70/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3414 - acc: 0.8614
Epoch 71/100
8000/8000 [==============================] - 1s 150us/step - loss: 0.3413 - acc: 0.8619
Epoch 72/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3408 - acc: 0.8609
Epoch 73/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3411 - acc: 0.8625
Epoch 74/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3409 - acc: 0.8600
Epoch 75/100
8000/8000 [==============================] - 1s 141us/step - loss: 0.3405 - acc: 0.8604
Epoch 76/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3406 - acc: 0.8605
Epoch 77/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3405 - acc: 0.8616
Epoch 78/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3404 - acc: 0.8608
Epoch 79/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3399 - acc: 0.8618
Epoch 80/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3400 - acc: 0.8606
Epoch 81/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3399 - acc: 0.8600
Epoch 82/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3399 - acc: 0.8598
Epoch 83/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3397 - acc: 0.8604
Epoch 84/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3397 - acc: 0.8613
Epoch 85/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3394 - acc: 0.8616
Epoch 86/100
8000/8000 [==============================] - 1s 155us/step - loss: 0.3393 - acc: 0.8598
Epoch 87/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3394 - acc: 0.8611
Epoch 88/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3393 - acc: 0.8613
Epoch 89/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3392 - acc: 0.8604
Epoch 90/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3392 - acc: 0.8606
Epoch 91/100
8000/8000 [==============================] - 1s 153us/step - loss: 0.3391 - acc: 0.8608
Epoch 92/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3391 - acc: 0.8595
Epoch 93/100
8000/8000 [==============================] - 1s 158us/step - loss: 0.3389 - acc: 0.8619
Epoch 94/100
8000/8000 [==============================] - 1s 144us/step - loss: 0.3392 - acc: 0.8611
Epoch 95/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3392 - acc: 0.8601
Epoch 96/100
8000/8000 [==============================] - 1s 143us/step - loss: 0.3388 - acc: 0.8596
Epoch 97/100
8000/8000 [==============================] - 1s 146us/step - loss: 0.3390 - acc: 0.8615
Epoch 98/100
8000/8000 [==============================] - 1s 145us/step - loss: 0.3389 - acc: 0.8614
Epoch 99/100
8000/8000 [==============================] - 1s 151us/step - loss: 0.3387 - acc: 0.8592
Epoch 100/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3386 - acc: 0.8608
<keras.callbacks.History at 0x7f4c84735860>

Como podemos ver después de 100 épocas con la nueva arquitectura tenemos 86% de accuracy.

Revisamos la matriz de confusión:

1
2
3
4
5
6
7
8
# Haciendo predicción de los resultados del Test
y_pred = neural_network2.predict(X_test)
y_pred_norm = (y_pred > 0.5)
 
y_pred_norm = y_pred_norm.astype(int)
y_test = y_test.astype(int)
 
plot_confusion_matrix(y_test, y_pred_norm, normalize=False,title="Matriz de Confusión: Abandono de clientes")
Matriz de Confusión sin Normalizar
[[1520   75]
 [ 206  199]]

Agregamos una nueva arquitectura:

Entrada => 10
Oculta => 9/7/5
Salida => 1

1
2
3
4
5
6
7
8
9
10
11
12
# Inicializando la Red Neuronal
neural_network3 = Sequential()
# kernel_initializer Define la forma como se asignará los Pesos iniciales Wi
initial_weights = RandomUniform(minval = -0.5, maxval = 0.5)
neural_network3.add(Dense(units = 9, kernel_initializer = initial_weights, activation = 'relu', input_dim = 10))
neural_network3.add(Dense(units = 7, kernel_initializer = initial_weights, activation = 'relu'))
neural_network3.add(Dense(units = 5, kernel_initializer = initial_weights, activation = 'relu'))
 
# Agregando capa de salida
neural_network3.add(Dense(units = 1, kernel_initializer = initial_weights, activation = 'sigmoid'))
# Imprimir Arquitectura de la Red
neural_network3.summary()
Model: "sequential_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_8 (Dense)              (None, 9)                 99        
_________________________________________________________________
dense_9 (Dense)              (None, 7)                 70        
_________________________________________________________________
dense_10 (Dense)             (None, 5)                 40        
_________________________________________________________________
dense_11 (Dense)             (None, 1)                 6         
=================================================================
Total params: 215
Trainable params: 215
Non-trainable params: 0
_________________________________________________________________
1
2
3
4
5
6
7
8
# Compilando la Red Neuronal
# optimizer: Algoritmo de optimización | binary_crossentropy = 2 Classes
# loss: error
neural_network3.compile(optimizer = 'adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
 
 
# Entrenamiento
neural_network3.fit(X_train, y_train, batch_size = 32, epochs = 100)
Epoch 1/100
8000/8000 [==============================] - 2s 216us/step - loss: 0.5574 - acc: 0.7860
Epoch 2/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.4544 - acc: 0.7969
Epoch 3/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.4302 - acc: 0.8099
Epoch 4/100
8000/8000 [==============================] - 1s 183us/step - loss: 0.4141 - acc: 0.8233
Epoch 5/100
8000/8000 [==============================] - 1s 170us/step - loss: 0.3949 - acc: 0.8340
Epoch 6/100
8000/8000 [==============================] - 1s 183us/step - loss: 0.3783 - acc: 0.8467
Epoch 7/100
8000/8000 [==============================] - 1s 187us/step - loss: 0.3683 - acc: 0.8530
Epoch 8/100
8000/8000 [==============================] - 2s 190us/step - loss: 0.3644 - acc: 0.8540
Epoch 9/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3611 - acc: 0.8561
Epoch 10/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3602 - acc: 0.8553
Epoch 11/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3586 - acc: 0.8560
Epoch 12/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3570 - acc: 0.8566
Epoch 13/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3565 - acc: 0.8558
Epoch 14/100
8000/8000 [==============================] - 2s 189us/step - loss: 0.3553 - acc: 0.8562
Epoch 15/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3549 - acc: 0.8551
Epoch 16/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3536 - acc: 0.8561
Epoch 17/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3533 - acc: 0.8562
Epoch 18/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3523 - acc: 0.8560
Epoch 19/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3511 - acc: 0.8568
Epoch 20/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3512 - acc: 0.8566
Epoch 21/100
8000/8000 [==============================] - 1s 170us/step - loss: 0.3509 - acc: 0.8584
Epoch 22/100
8000/8000 [==============================] - 2s 188us/step - loss: 0.3497 - acc: 0.8572
Epoch 23/100
8000/8000 [==============================] - 1s 181us/step - loss: 0.3487 - acc: 0.8570
Epoch 24/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3490 - acc: 0.8572
Epoch 25/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3484 - acc: 0.8572
Epoch 26/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3477 - acc: 0.8589
Epoch 27/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3467 - acc: 0.8570
Epoch 28/100
8000/8000 [==============================] - 1s 184us/step - loss: 0.3466 - acc: 0.8578
Epoch 29/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3462 - acc: 0.8584
Epoch 30/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3459 - acc: 0.8576
Epoch 31/100
8000/8000 [==============================] - 1s 178us/step - loss: 0.3457 - acc: 0.8571
Epoch 32/100
8000/8000 [==============================] - 2s 193us/step - loss: 0.3450 - acc: 0.8589
Epoch 33/100
8000/8000 [==============================] - 1s 181us/step - loss: 0.3440 - acc: 0.8581
Epoch 34/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3438 - acc: 0.8585
Epoch 35/100
8000/8000 [==============================] - 1s 187us/step - loss: 0.3437 - acc: 0.8600
Epoch 36/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3431 - acc: 0.8580
Epoch 37/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3425 - acc: 0.8589
Epoch 38/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3419 - acc: 0.8596
Epoch 39/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3416 - acc: 0.8585
Epoch 40/100
8000/8000 [==============================] - 1s 162us/step - loss: 0.3410 - acc: 0.8596
Epoch 41/100
8000/8000 [==============================] - 1s 166us/step - loss: 0.3407 - acc: 0.8601
Epoch 42/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3410 - acc: 0.8589
Epoch 43/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3401 - acc: 0.8595
Epoch 44/100
8000/8000 [==============================] - 1s 178us/step - loss: 0.3396 - acc: 0.8596
Epoch 45/100
8000/8000 [==============================] - 1s 183us/step - loss: 0.3393 - acc: 0.8586
Epoch 46/100
8000/8000 [==============================] - 2s 192us/step - loss: 0.3388 - acc: 0.8613
Epoch 47/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3388 - acc: 0.8600
Epoch 48/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3381 - acc: 0.8606
Epoch 49/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3378 - acc: 0.8621
Epoch 50/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3374 - acc: 0.8610
Epoch 51/100
8000/8000 [==============================] - 1s 159us/step - loss: 0.3378 - acc: 0.8616
Epoch 52/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3367 - acc: 0.8619
Epoch 53/100
8000/8000 [==============================] - 1s 167us/step - loss: 0.3367 - acc: 0.8622
Epoch 54/100
8000/8000 [==============================] - 1s 184us/step - loss: 0.3361 - acc: 0.8624
Epoch 55/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3351 - acc: 0.8611
Epoch 56/100
8000/8000 [==============================] - 1s 187us/step - loss: 0.3355 - acc: 0.8625
Epoch 57/100
8000/8000 [==============================] - 1s 186us/step - loss: 0.3348 - acc: 0.8610
Epoch 58/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3348 - acc: 0.8626
Epoch 59/100
8000/8000 [==============================] - 1s 184us/step - loss: 0.3349 - acc: 0.8630
Epoch 60/100
8000/8000 [==============================] - 1s 174us/step - loss: 0.3346 - acc: 0.8625
Epoch 61/100
8000/8000 [==============================] - 2s 188us/step - loss: 0.3347 - acc: 0.8604
Epoch 62/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3344 - acc: 0.8634
Epoch 63/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3340 - acc: 0.8629
Epoch 64/100
8000/8000 [==============================] - 1s 172us/step - loss: 0.3335 - acc: 0.8632
Epoch 65/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3340 - acc: 0.8631
Epoch 66/100
8000/8000 [==============================] - 2s 199us/step - loss: 0.3333 - acc: 0.8620
Epoch 67/100
8000/8000 [==============================] - 1s 185us/step - loss: 0.3331 - acc: 0.8631
Epoch 68/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3325 - acc: 0.8636
Epoch 69/100
8000/8000 [==============================] - 1s 185us/step - loss: 0.3325 - acc: 0.8628
Epoch 70/100
8000/8000 [==============================] - 2s 189us/step - loss: 0.3325 - acc: 0.8635
Epoch 71/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3322 - acc: 0.8626
Epoch 72/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3323 - acc: 0.8645
Epoch 73/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3325 - acc: 0.8638
Epoch 74/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3321 - acc: 0.8642
Epoch 75/100
8000/8000 [==============================] - 1s 185us/step - loss: 0.3320 - acc: 0.8620
Epoch 76/100
8000/8000 [==============================] - 1s 181us/step - loss: 0.3318 - acc: 0.8625
Epoch 77/100
8000/8000 [==============================] - 2s 188us/step - loss: 0.3316 - acc: 0.8641
Epoch 78/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3319 - acc: 0.8639
Epoch 79/100
8000/8000 [==============================] - 1s 163us/step - loss: 0.3320 - acc: 0.8642
Epoch 80/100
8000/8000 [==============================] - 1s 170us/step - loss: 0.3314 - acc: 0.8631
Epoch 81/100
8000/8000 [==============================] - 1s 180us/step - loss: 0.3313 - acc: 0.8618
Epoch 82/100
8000/8000 [==============================] - 1s 176us/step - loss: 0.3317 - acc: 0.8635
Epoch 83/100
8000/8000 [==============================] - 1s 161us/step - loss: 0.3315 - acc: 0.8644
Epoch 84/100
8000/8000 [==============================] - 1s 181us/step - loss: 0.3315 - acc: 0.8635
Epoch 85/100
8000/8000 [==============================] - 1s 171us/step - loss: 0.3311 - acc: 0.8642
Epoch 86/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3307 - acc: 0.8644
Epoch 87/100
8000/8000 [==============================] - 1s 177us/step - loss: 0.3316 - acc: 0.8644
Epoch 88/100
8000/8000 [==============================] - 1s 179us/step - loss: 0.3309 - acc: 0.8656
Epoch 89/100
8000/8000 [==============================] - 1s 175us/step - loss: 0.3312 - acc: 0.8632
Epoch 90/100
8000/8000 [==============================] - 1s 156us/step - loss: 0.3309 - acc: 0.8639
Epoch 91/100
8000/8000 [==============================] - 1s 185us/step - loss: 0.3308 - acc: 0.8638
Epoch 92/100
8000/8000 [==============================] - 1s 160us/step - loss: 0.3310 - acc: 0.8641
Epoch 93/100
8000/8000 [==============================] - 1s 164us/step - loss: 0.3302 - acc: 0.8638
Epoch 94/100
8000/8000 [==============================] - 1s 182us/step - loss: 0.3312 - acc: 0.8639
Epoch 95/100
8000/8000 [==============================] - 1s 178us/step - loss: 0.3305 - acc: 0.8651
Epoch 96/100
8000/8000 [==============================] - 1s 183us/step - loss: 0.3311 - acc: 0.8644
Epoch 97/100
8000/8000 [==============================] - 1s 169us/step - loss: 0.3303 - acc: 0.8640
Epoch 98/100
8000/8000 [==============================] - 1s 165us/step - loss: 0.3307 - acc: 0.8631
Epoch 99/100
8000/8000 [==============================] - 1s 157us/step - loss: 0.3306 - acc: 0.8645
Epoch 100/100
8000/8000 [==============================] - 1s 168us/step - loss: 0.3307 - acc: 0.8638
<keras.callbacks.History at 0x7f4c84823a20>

Se puede observar que se obtiene un acccuracy de 86.3%

1
2
3
4
5
6
7
8
# Haciendo predicción de los resultados del Test
y_pred = neural_network3.predict(X_test)
y_pred_norm = (y_pred > 0.5)
 
y_pred_norm = y_pred_norm.astype(int)
y_test = y_test.astype(int)
 
plot_confusion_matrix(y_test, y_pred_norm, normalize=False,title="Matriz de Confusión: Abandono de clientes")
Matriz de Confusión sin Normalizar
[[1526   69]
 [ 214  191]]

Leave a Reply

Su dirección de correo no se hará público. Los campos requeridos están marcados *

es_PEES
en_USEN es_PEES