Clasificación sobre datos desequilibrados

Ver en TensorFlow.org Ejecutar en Google Colab Ver fuente en GitHub Descargar libreta

Este tutorial demuestra cómo clasificar un conjunto de datos altamente desequilibrado en el que la cantidad de ejemplos en una clase supera en gran medida a los ejemplos en otra. Trabajará con el conjunto de datos de detección de fraude de tarjetas de crédito alojado en Kaggle. El objetivo es detectar apenas 492 transacciones fraudulentas de 284.807 transacciones en total. Utilizará Keras para definir el modelo y los pesos de clase para ayudar al modelo a aprender de los datos desequilibrados. .

Este tutorial contiene código completo para:

  • Cargue un archivo CSV usando Pandas.
  • Cree conjuntos de entrenamiento, validación y prueba.
  • Defina y entrene un modelo usando Keras (incluido el establecimiento de pesos de clase).
  • Evalúe el modelo utilizando varias métricas (incluidas la precisión y la recuperación).
  • Pruebe técnicas comunes para manejar datos desequilibrados como:
    • Ponderación de clase
    • sobremuestreo

Configuración

import tensorflow as tf
from tensorflow import keras

import os
import tempfile

import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns

import sklearn
from sklearn.metrics import confusion_matrix
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
mpl.rcParams['figure.figsize'] = (12, 10)
colors = plt.rcParams['axes.prop_cycle'].by_key()['color']

Procesamiento y exploración de datos

Descargue el conjunto de datos de fraude de tarjetas de crédito de Kaggle

Pandas es una biblioteca de Python con muchas utilidades útiles para cargar y trabajar con datos estructurados. Se puede usar para descargar archivos CSV en un DataFrame de Pandas.

file = tf.keras.utils
raw_df = pd.read_csv('https://storage.googleapis.com/download.tensorflow.org/data/creditcard.csv')
raw_df.head()
raw_df[['Time', 'V1', 'V2', 'V3', 'V4', 'V5', 'V26', 'V27', 'V28', 'Amount', 'Class']].describe()

Examinar el desequilibrio de la etiqueta de clase

Veamos el desequilibrio del conjunto de datos:

neg, pos = np.bincount(raw_df['Class'])
total = neg + pos
print('Examples:\n    Total: {}\n    Positive: {} ({:.2f}% of total)\n'.format(
    total, pos, 100 * pos / total))
Examples:
    Total: 284807
    Positive: 492 (0.17% of total)

Esto muestra la pequeña fracción de muestras positivas.

Limpiar, dividir y normalizar los datos

Los datos sin procesar tienen algunos problemas. Primero, las columnas Time y Amount son demasiado variables para usarlas directamente. Suelte la columna Time (ya que no está claro lo que significa) y tome el registro de la columna Amount para reducir su rango.

cleaned_df = raw_df.copy()

# You don't want the `Time` column.
cleaned_df.pop('Time')

# The `Amount` column covers a huge range. Convert to log-space.
eps = 0.001 # 0 => 0.1¢
cleaned_df['Log Ammount'] = np.log(cleaned_df.pop('Amount')+eps)

Divida el conjunto de datos en conjuntos de entrenamiento, validación y prueba. El conjunto de validación se utiliza durante el ajuste del modelo para evaluar la pérdida y cualquier métrica; sin embargo, el modelo no se ajusta a estos datos. El conjunto de prueba no se usa en absoluto durante la fase de entrenamiento y solo se usa al final para evaluar qué tan bien se generaliza el modelo a nuevos datos. Esto es especialmente importante con conjuntos de datos desequilibrados donde el sobreajuste es una preocupación importante debido a la falta de datos de entrenamiento.

# Use a utility from sklearn to split and shuffle your dataset.
train_df, test_df = train_test_split(cleaned_df, test_size=0.2)
train_df, val_df = train_test_split(train_df, test_size=0.2)

# Form np arrays of labels and features.
train_labels = np.array(train_df.pop('Class'))
bool_train_labels = train_labels != 0
val_labels = np.array(val_df.pop('Class'))
test_labels = np.array(test_df.pop('Class'))

train_features = np.array(train_df)
val_features = np.array(val_df)
test_features = np.array(test_df)

Normalice las características de entrada utilizando el StandardScaler de sklearn. Esto establecerá la media en 0 y la desviación estándar en 1.

scaler = StandardScaler()
train_features = scaler.fit_transform(train_features)

val_features = scaler.transform(val_features)
test_features = scaler.transform(test_features)

train_features = np.clip(train_features, -5, 5)
val_features = np.clip(val_features, -5, 5)
test_features = np.clip(test_features, -5, 5)


print('Training labels shape:', train_labels.shape)
print('Validation labels shape:', val_labels.shape)
print('Test labels shape:', test_labels.shape)

print('Training features shape:', train_features.shape)
print('Validation features shape:', val_features.shape)
print('Test features shape:', test_features.shape)
Training labels shape: (182276,)
Validation labels shape: (45569,)
Test labels shape: (56962,)
Training features shape: (182276, 29)
Validation features shape: (45569, 29)
Test features shape: (56962, 29)

Mira la distribución de datos.

A continuación, compare las distribuciones de los ejemplos positivos y negativos sobre algunas características. Buenas preguntas para hacerse en este punto son:

  • ¿Tienen sentido estas distribuciones?
    • Si. Ha normalizado la entrada y estos se concentran principalmente en el rango +/- 2 .
  • ¿Puedes ver la diferencia entre las distribuciones?
    • Sí, los ejemplos positivos contienen una tasa mucho más alta de valores extremos.
pos_df = pd.DataFrame(train_features[ bool_train_labels], columns=train_df.columns)
neg_df = pd.DataFrame(train_features[~bool_train_labels], columns=train_df.columns)

sns.jointplot(x=pos_df['V5'], y=pos_df['V6'],
              kind='hex', xlim=(-5,5), ylim=(-5,5))
plt.suptitle("Positive distribution")

sns.jointplot(x=neg_df['V5'], y=neg_df['V6'],
              kind='hex', xlim=(-5,5), ylim=(-5,5))
_ = plt.suptitle("Negative distribution")

png

png

Definir el modelo y las métricas

Defina una función que cree una red neuronal simple con una capa oculta densamente conectada, una capa de abandono para reducir el sobreajuste y una capa sigmoidea de salida que devuelva la probabilidad de que una transacción sea fraudulenta:

METRICS = [
      keras.metrics.TruePositives(name='tp'),
      keras.metrics.FalsePositives(name='fp'),
      keras.metrics.TrueNegatives(name='tn'),
      keras.metrics.FalseNegatives(name='fn'), 
      keras.metrics.BinaryAccuracy(name='accuracy'),
      keras.metrics.Precision(name='precision'),
      keras.metrics.Recall(name='recall'),
      keras.metrics.AUC(name='auc'),
      keras.metrics.AUC(name='prc', curve='PR'), # precision-recall curve
]

def make_model(metrics=METRICS, output_bias=None):
  if output_bias is not None:
    output_bias = tf.keras.initializers.Constant(output_bias)
  model = keras.Sequential([
      keras.layers.Dense(
          16, activation='relu',
          input_shape=(train_features.shape[-1],)),
      keras.layers.Dropout(0.5),
      keras.layers.Dense(1, activation='sigmoid',
                         bias_initializer=output_bias),
  ])

  model.compile(
      optimizer=keras.optimizers.Adam(learning_rate=1e-3),
      loss=keras.losses.BinaryCrossentropy(),
      metrics=metrics)

  return model

Comprender las métricas útiles

Tenga en cuenta que hay algunas métricas definidas anteriormente que el modelo puede calcular y que serán útiles al evaluar el rendimiento.

  • Los falsos negativos y los falsos positivos son muestras que se clasificaron incorrectamente
  • Los verdaderos negativos y los verdaderos positivos son muestras que se clasificaron correctamente
  • La precisión es el porcentaje de ejemplos clasificados correctamente > \(\frac{\text{true samples} }{\text{total samples} }\)
  • La precisión es el porcentaje de positivos pronosticados que se clasificaron correctamente > \(\frac{\text{true positives} }{\text{true positives + false positives} }\)
  • Recordar es el porcentaje de positivos reales que se clasificaron correctamente > \(\frac{\text{true positives} }{\text{true positives + false negatives} }\)
  • AUC se refiere al área bajo la curva de una curva característica operativa del receptor (ROC-AUC). Esta métrica es igual a la probabilidad de que un clasificador clasifique una muestra aleatoria positiva más alta que una muestra aleatoria negativa.
  • AUPRC se refiere al área bajo la curva de la curva de recuperación de precisión. Esta métrica calcula pares de recuperación de precisión para diferentes umbrales de probabilidad.

Lee mas:

modelo de referencia

Construye el modelo

Ahora cree y entrene su modelo usando la función que se definió anteriormente. Tenga en cuenta que el modelo se ajusta utilizando un tamaño de lote mayor que el predeterminado de 2048, esto es importante para garantizar que cada lote tenga una probabilidad decente de contener algunas muestras positivas. Si el tamaño del lote fuera demasiado pequeño, probablemente no tendrían transacciones fraudulentas de las que aprender.

EPOCHS = 100
BATCH_SIZE = 2048

early_stopping = tf.keras.callbacks.EarlyStopping(
    monitor='val_prc', 
    verbose=1,
    patience=10,
    mode='max',
    restore_best_weights=True)
model = make_model()
model.summary()
Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 dense (Dense)               (None, 16)                480       
                                                                 
 dropout (Dropout)           (None, 16)                0         
                                                                 
 dense_1 (Dense)             (None, 1)                 17        
                                                                 
=================================================================
Total params: 497
Trainable params: 497
Non-trainable params: 0
_________________________________________________________________

Prueba ejecutar el modelo:

model.predict(train_features[:10])
array([[0.9466284 ],
       [0.7211031 ],
       [0.60527885],
       [0.8335568 ],
       [0.5909625 ],
       [0.6751574 ],
       [0.6623665 ],
       [0.81066036],
       [0.50712407],
       [0.8296292 ]], dtype=float32)

Opcional: Establezca el sesgo inicial correcto.

Estas conjeturas iniciales no son geniales. Sabe que el conjunto de datos está desequilibrado. Establezca el sesgo de la capa de salida para reflejar eso (Consulte: Una receta para entrenar redes neuronales: "iniciar bien" ). Esto puede ayudar con la convergencia inicial.

Con la inicialización del sesgo predeterminado, la pérdida debería ser aproximadamente math.log(2) = 0.69314

results = model.evaluate(train_features, train_labels, batch_size=BATCH_SIZE, verbose=0)
print("Loss: {:0.4f}".format(results[0]))
Loss: 1.2781

El sesgo correcto para establecer se puede derivar de:

\[ p_0 = pos/(pos + neg) = 1/(1+e^{-b_0}) \]

\[ b_0 = -log_e(1/p_0 - 1) \]

\[ b_0 = log_e(pos/neg)\]

initial_bias = np.log([pos/neg])
initial_bias
array([-6.35935934])

Establézcalo como el sesgo inicial y el modelo dará conjeturas iniciales mucho más razonables.

Debe estar cerca: pos/total = 0.0018

model = make_model(output_bias=initial_bias)
model.predict(train_features[:10])
array([[2.3598122e-05],
       [1.5476024e-03],
       [6.8338902e-04],
       [9.4873342e-04],
       [1.0742771e-03],
       [7.7475846e-04],
       [1.2199467e-03],
       [5.5399281e-04],
       [1.6213538e-03],
       [3.0470363e-04]], dtype=float32)

Con esta inicialización la pérdida inicial debería ser aproximadamente:

\[-p_0log(p_0)-(1-p_0)log(1-p_0) = 0.01317\]

results = model.evaluate(train_features, train_labels, batch_size=BATCH_SIZE, verbose=0)
print("Loss: {:0.4f}".format(results[0]))
Loss: 0.0200

Esta pérdida inicial es aproximadamente 50 veces menor que si hubiera sido con una inicialización ingenua.

De esta forma, el modelo no necesita pasar las primeras épocas simplemente aprendiendo que los ejemplos positivos son poco probables. Esto también facilita la lectura de gráficos de la pérdida durante el entrenamiento.

Controlar los pesos iniciales

Para que las distintas ejecuciones de entrenamiento sean más comparables, mantenga los pesos de este modelo inicial en un archivo de punto de control y cárguelos en cada modelo antes del entrenamiento:

initial_weights = os.path.join(tempfile.mkdtemp(), 'initial_weights')
model.save_weights(initial_weights)

Confirme que la solución de sesgo ayuda

Antes de continuar, confirme rápidamente que la cuidadosa inicialización del sesgo realmente ayudó.

Entrene el modelo durante 20 épocas, con y sin esta cuidadosa inicialización, y compare las pérdidas:

model = make_model()
model.load_weights(initial_weights)
model.layers[-1].bias.assign([0.0])
zero_bias_history = model.fit(
    train_features,
    train_labels,
    batch_size=BATCH_SIZE,
    epochs=20,
    validation_data=(val_features, val_labels), 
    verbose=0)
model = make_model()
model.load_weights(initial_weights)
careful_bias_history = model.fit(
    train_features,
    train_labels,
    batch_size=BATCH_SIZE,
    epochs=20,
    validation_data=(val_features, val_labels), 
    verbose=0)
def plot_loss(history, label, n):
  # Use a log scale on y-axis to show the wide range of values.
  plt.semilogy(history.epoch, history.history['loss'],
               color=colors[n], label='Train ' + label)
  plt.semilogy(history.epoch, history.history['val_loss'],
               color=colors[n], label='Val ' + label,
               linestyle="--")
  plt.xlabel('Epoch')
  plt.ylabel('Loss')
plot_loss(zero_bias_history, "Zero Bias", 0)
plot_loss(careful_bias_history, "Careful Bias", 1)

png

La figura anterior lo deja claro: en términos de pérdida de validación, en este problema, esta inicialización cuidadosa brinda una clara ventaja.

entrenar al modelo

model = make_model()
model.load_weights(initial_weights)
baseline_history = model.fit(
    train_features,
    train_labels,
    batch_size=BATCH_SIZE,
    epochs=EPOCHS,
    callbacks=[early_stopping],
    validation_data=(val_features, val_labels))
Epoch 1/100
90/90 [==============================] - 3s 15ms/step - loss: 0.0161 - tp: 64.0000 - fp: 9.0000 - tn: 227425.0000 - fn: 347.0000 - accuracy: 0.9984 - precision: 0.8767 - recall: 0.1557 - auc: 0.6148 - prc: 0.1692 - val_loss: 0.0115 - val_tp: 0.0000e+00 - val_fp: 0.0000e+00 - val_tn: 45483.0000 - val_fn: 86.0000 - val_accuracy: 0.9981 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00 - val_auc: 0.7205 - val_prc: 0.2571
Epoch 2/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0087 - tp: 49.0000 - fp: 11.0000 - tn: 181940.0000 - fn: 276.0000 - accuracy: 0.9984 - precision: 0.8167 - recall: 0.1508 - auc: 0.8085 - prc: 0.3735 - val_loss: 0.0054 - val_tp: 35.0000 - val_fp: 6.0000 - val_tn: 45477.0000 - val_fn: 51.0000 - val_accuracy: 0.9987 - val_precision: 0.8537 - val_recall: 0.4070 - val_auc: 0.9065 - val_prc: 0.6598
Epoch 3/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0061 - tp: 126.0000 - fp: 27.0000 - tn: 181924.0000 - fn: 199.0000 - accuracy: 0.9988 - precision: 0.8235 - recall: 0.3877 - auc: 0.8997 - prc: 0.6187 - val_loss: 0.0046 - val_tp: 55.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 31.0000 - val_accuracy: 0.9991 - val_precision: 0.8730 - val_recall: 0.6395 - val_auc: 0.9063 - val_prc: 0.6941
Epoch 4/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0056 - tp: 172.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 153.0000 - accuracy: 0.9990 - precision: 0.8473 - recall: 0.5292 - auc: 0.9068 - prc: 0.6448 - val_loss: 0.0044 - val_tp: 58.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 28.0000 - val_accuracy: 0.9992 - val_precision: 0.8788 - val_recall: 0.6744 - val_auc: 0.9064 - val_prc: 0.7114
Epoch 5/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0056 - tp: 167.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 158.0000 - accuracy: 0.9990 - precision: 0.8477 - recall: 0.5138 - auc: 0.9134 - prc: 0.6215 - val_loss: 0.0043 - val_tp: 60.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 26.0000 - val_accuracy: 0.9993 - val_precision: 0.8824 - val_recall: 0.6977 - val_auc: 0.9064 - val_prc: 0.7181
Epoch 6/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0050 - tp: 193.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 132.0000 - accuracy: 0.9991 - precision: 0.8733 - recall: 0.5938 - auc: 0.9198 - prc: 0.6760 - val_loss: 0.0042 - val_tp: 59.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 27.0000 - val_accuracy: 0.9992 - val_precision: 0.8806 - val_recall: 0.6860 - val_auc: 0.9064 - val_prc: 0.7370
Epoch 7/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0048 - tp: 183.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 142.0000 - accuracy: 0.9991 - precision: 0.8592 - recall: 0.5631 - auc: 0.9202 - prc: 0.6737 - val_loss: 0.0042 - val_tp: 60.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 26.0000 - val_accuracy: 0.9993 - val_precision: 0.8824 - val_recall: 0.6977 - val_auc: 0.9064 - val_prc: 0.7463
Epoch 8/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0050 - tp: 171.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 154.0000 - accuracy: 0.9990 - precision: 0.8465 - recall: 0.5262 - auc: 0.9156 - prc: 0.6574 - val_loss: 0.0041 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9065 - val_prc: 0.7480
Epoch 9/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0047 - tp: 196.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 129.0000 - accuracy: 0.9991 - precision: 0.8711 - recall: 0.6031 - auc: 0.9218 - prc: 0.6799 - val_loss: 0.0041 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9065 - val_prc: 0.7550
Epoch 10/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0050 - tp: 173.0000 - fp: 27.0000 - tn: 181924.0000 - fn: 152.0000 - accuracy: 0.9990 - precision: 0.8650 - recall: 0.5323 - auc: 0.9048 - prc: 0.6520 - val_loss: 0.0040 - val_tp: 63.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 23.0000 - val_accuracy: 0.9993 - val_precision: 0.8750 - val_recall: 0.7326 - val_auc: 0.9122 - val_prc: 0.7598
Epoch 11/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0048 - tp: 190.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 135.0000 - accuracy: 0.9991 - precision: 0.8597 - recall: 0.5846 - auc: 0.9172 - prc: 0.6779 - val_loss: 0.0040 - val_tp: 63.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 23.0000 - val_accuracy: 0.9993 - val_precision: 0.8750 - val_recall: 0.7326 - val_auc: 0.9065 - val_prc: 0.7595
Epoch 12/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0043 - tp: 192.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 133.0000 - accuracy: 0.9991 - precision: 0.8571 - recall: 0.5908 - auc: 0.9281 - prc: 0.7312 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8767 - val_recall: 0.7442 - val_auc: 0.9123 - val_prc: 0.7648
Epoch 13/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0042 - tp: 185.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 140.0000 - accuracy: 0.9991 - precision: 0.8565 - recall: 0.5692 - auc: 0.9328 - prc: 0.7222 - val_loss: 0.0040 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9123 - val_prc: 0.7615
Epoch 14/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0047 - tp: 183.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 142.0000 - accuracy: 0.9990 - precision: 0.8472 - recall: 0.5631 - auc: 0.9295 - prc: 0.6770 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9123 - val_prc: 0.7670
Epoch 15/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0043 - tp: 194.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8700 - recall: 0.5969 - auc: 0.9344 - prc: 0.7233 - val_loss: 0.0040 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9123 - val_prc: 0.7672
Epoch 16/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0041 - tp: 207.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8697 - recall: 0.6369 - auc: 0.9329 - prc: 0.7194 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8767 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7694
Epoch 17/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0042 - tp: 190.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 135.0000 - accuracy: 0.9991 - precision: 0.8716 - recall: 0.5846 - auc: 0.9345 - prc: 0.7265 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7705
Epoch 18/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0040 - tp: 194.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8622 - recall: 0.5969 - auc: 0.9344 - prc: 0.7199 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7725
Epoch 19/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0041 - tp: 205.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 120.0000 - accuracy: 0.9992 - precision: 0.8613 - recall: 0.6308 - auc: 0.9346 - prc: 0.7266 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7739
Epoch 20/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 207.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8809 - recall: 0.6369 - auc: 0.9421 - prc: 0.7634 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8784 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7729
Epoch 21/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 204.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8644 - recall: 0.6277 - auc: 0.9360 - prc: 0.7340 - val_loss: 0.0038 - val_tp: 62.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8732 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7756
Epoch 22/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0040 - tp: 207.0000 - fp: 26.0000 - tn: 181925.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8884 - recall: 0.6369 - auc: 0.9328 - prc: 0.7277 - val_loss: 0.0038 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7773
Epoch 23/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 191.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8527 - recall: 0.5877 - auc: 0.9375 - prc: 0.7280 - val_loss: 0.0038 - val_tp: 62.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8857 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7790
Epoch 24/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0039 - tp: 196.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 129.0000 - accuracy: 0.9991 - precision: 0.8596 - recall: 0.6031 - auc: 0.9375 - prc: 0.7466 - val_loss: 0.0038 - val_tp: 65.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8667 - val_recall: 0.7558 - val_auc: 0.9123 - val_prc: 0.7762
Epoch 25/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 204.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8681 - recall: 0.6277 - auc: 0.9467 - prc: 0.7480 - val_loss: 0.0038 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9123 - val_prc: 0.7789
Epoch 26/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0040 - tp: 194.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8661 - recall: 0.5969 - auc: 0.9360 - prc: 0.7292 - val_loss: 0.0038 - val_tp: 60.0000 - val_fp: 7.0000 - val_tn: 45476.0000 - val_fn: 26.0000 - val_accuracy: 0.9993 - val_precision: 0.8955 - val_recall: 0.6977 - val_auc: 0.9123 - val_prc: 0.7783
Epoch 27/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 208.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 117.0000 - accuracy: 0.9992 - precision: 0.8776 - recall: 0.6400 - auc: 0.9376 - prc: 0.7632 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8667 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7772
Epoch 28/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 202.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 123.0000 - accuracy: 0.9991 - precision: 0.8596 - recall: 0.6215 - auc: 0.9408 - prc: 0.7638 - val_loss: 0.0039 - val_tp: 63.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 23.0000 - val_accuracy: 0.9993 - val_precision: 0.8630 - val_recall: 0.7326 - val_auc: 0.9124 - val_prc: 0.7808
Epoch 29/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 214.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 111.0000 - accuracy: 0.9992 - precision: 0.8807 - recall: 0.6585 - auc: 0.9347 - prc: 0.7626 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8732 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7806
Epoch 30/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0039 - tp: 197.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 128.0000 - accuracy: 0.9991 - precision: 0.8640 - recall: 0.6062 - auc: 0.9346 - prc: 0.7489 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8667 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7804
Epoch 31/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 213.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 112.0000 - accuracy: 0.9992 - precision: 0.8659 - recall: 0.6554 - auc: 0.9407 - prc: 0.7615 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7809
Epoch 32/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 217.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 108.0000 - accuracy: 0.9993 - precision: 0.8857 - recall: 0.6677 - auc: 0.9407 - prc: 0.7626 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8732 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7821
Epoch 33/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 210.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 115.0000 - accuracy: 0.9992 - precision: 0.8787 - recall: 0.6462 - auc: 0.9392 - prc: 0.7642 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8732 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7826
Epoch 34/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 217.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 108.0000 - accuracy: 0.9993 - precision: 0.8857 - recall: 0.6677 - auc: 0.9423 - prc: 0.7759 - val_loss: 0.0038 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7830
Epoch 35/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0038 - tp: 209.0000 - fp: 35.0000 - tn: 181916.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8566 - recall: 0.6431 - auc: 0.9407 - prc: 0.7381 - val_loss: 0.0038 - val_tp: 61.0000 - val_fp: 7.0000 - val_tn: 45476.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8971 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7836
Epoch 36/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 204.0000 - fp: 27.0000 - tn: 181924.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8831 - recall: 0.6277 - auc: 0.9407 - prc: 0.7587 - val_loss: 0.0038 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7840
Epoch 37/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0038 - tp: 209.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8672 - recall: 0.6431 - auc: 0.9345 - prc: 0.7386 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 7.0000 - val_tn: 45476.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8971 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7849
Epoch 38/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0038 - tp: 198.0000 - fp: 33.0000 - tn: 181918.0000 - fn: 127.0000 - accuracy: 0.9991 - precision: 0.8571 - recall: 0.6092 - auc: 0.9454 - prc: 0.7488 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7844
Epoch 39/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 209.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8782 - recall: 0.6431 - auc: 0.9407 - prc: 0.7419 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7840
Epoch 40/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 198.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 127.0000 - accuracy: 0.9991 - precision: 0.8761 - recall: 0.6092 - auc: 0.9546 - prc: 0.7644 - val_loss: 0.0039 - val_tp: 65.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8667 - val_recall: 0.7558 - val_auc: 0.9124 - val_prc: 0.7835
Epoch 41/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0038 - tp: 209.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8745 - recall: 0.6431 - auc: 0.9377 - prc: 0.7587 - val_loss: 0.0039 - val_tp: 63.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 23.0000 - val_accuracy: 0.9993 - val_precision: 0.8630 - val_recall: 0.7326 - val_auc: 0.9124 - val_prc: 0.7827
Epoch 42/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 195.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 130.0000 - accuracy: 0.9991 - precision: 0.8667 - recall: 0.6000 - auc: 0.9345 - prc: 0.7436 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8649 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7834
Epoch 43/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 206.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8655 - recall: 0.6338 - auc: 0.9500 - prc: 0.7699 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7836
Epoch 44/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 208.0000 - fp: 25.0000 - tn: 181926.0000 - fn: 117.0000 - accuracy: 0.9992 - precision: 0.8927 - recall: 0.6400 - auc: 0.9438 - prc: 0.7625 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8611 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7841
Epoch 45/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 205.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 120.0000 - accuracy: 0.9992 - precision: 0.8686 - recall: 0.6308 - auc: 0.9422 - prc: 0.7519 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7847
Epoch 46/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0037 - tp: 206.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8766 - recall: 0.6338 - auc: 0.9423 - prc: 0.7529 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8611 - val_recall: 0.7209 - val_auc: 0.9124 - val_prc: 0.7843
Epoch 47/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0035 - tp: 219.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 106.0000 - accuracy: 0.9993 - precision: 0.8866 - recall: 0.6738 - auc: 0.9377 - prc: 0.7677 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45475.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7871
Epoch 48/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 206.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8729 - recall: 0.6338 - auc: 0.9393 - prc: 0.7676 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8649 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7854
Epoch 49/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 215.0000 - fp: 29.0000 - tn: 181922.0000 - fn: 110.0000 - accuracy: 0.9992 - precision: 0.8811 - recall: 0.6615 - auc: 0.9407 - prc: 0.7618 - val_loss: 0.0039 - val_tp: 62.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 24.0000 - val_accuracy: 0.9993 - val_precision: 0.8611 - val_recall: 0.7209 - val_auc: 0.9125 - val_prc: 0.7855
Epoch 50/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0035 - tp: 214.0000 - fp: 32.0000 - tn: 181919.0000 - fn: 111.0000 - accuracy: 0.9992 - precision: 0.8699 - recall: 0.6585 - auc: 0.9377 - prc: 0.7727 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8649 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7858
Epoch 51/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0034 - tp: 219.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 106.0000 - accuracy: 0.9993 - precision: 0.8795 - recall: 0.6738 - auc: 0.9393 - prc: 0.7889 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 7.0000 - val_tn: 45476.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8971 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7876
Epoch 52/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0034 - tp: 217.0000 - fp: 25.0000 - tn: 181926.0000 - fn: 108.0000 - accuracy: 0.9993 - precision: 0.8967 - recall: 0.6677 - auc: 0.9439 - prc: 0.7812 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9125 - val_prc: 0.7887
Epoch 53/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0035 - tp: 206.0000 - fp: 28.0000 - tn: 181923.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8803 - recall: 0.6338 - auc: 0.9362 - prc: 0.7734 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8649 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7873
Epoch 54/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 223.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 102.0000 - accuracy: 0.9993 - precision: 0.8814 - recall: 0.6862 - auc: 0.9438 - prc: 0.7677 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9125 - val_prc: 0.7877
Epoch 55/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0034 - tp: 220.0000 - fp: 26.0000 - tn: 181925.0000 - fn: 105.0000 - accuracy: 0.9993 - precision: 0.8943 - recall: 0.6769 - auc: 0.9439 - prc: 0.7866 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7886
Epoch 56/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 209.0000 - fp: 24.0000 - tn: 181927.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8970 - recall: 0.6431 - auc: 0.9392 - prc: 0.7613 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 7.0000 - val_tn: 45476.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8971 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7886
Epoch 57/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0033 - tp: 221.0000 - fp: 23.0000 - tn: 181928.0000 - fn: 104.0000 - accuracy: 0.9993 - precision: 0.9057 - recall: 0.6800 - auc: 0.9516 - prc: 0.7954 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7873
Epoch 58/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 208.0000 - fp: 27.0000 - tn: 181924.0000 - fn: 117.0000 - accuracy: 0.9992 - precision: 0.8851 - recall: 0.6400 - auc: 0.9485 - prc: 0.7746 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7875
Epoch 59/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0034 - tp: 216.0000 - fp: 30.0000 - tn: 181921.0000 - fn: 109.0000 - accuracy: 0.9992 - precision: 0.8780 - recall: 0.6646 - auc: 0.9531 - prc: 0.7928 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9125 - val_prc: 0.7883
Epoch 60/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0035 - tp: 211.0000 - fp: 31.0000 - tn: 181920.0000 - fn: 114.0000 - accuracy: 0.9992 - precision: 0.8719 - recall: 0.6492 - auc: 0.9469 - prc: 0.7808 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9125 - val_prc: 0.7882
Epoch 61/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0036 - tp: 201.0000 - fp: 24.0000 - tn: 181927.0000 - fn: 124.0000 - accuracy: 0.9992 - precision: 0.8933 - recall: 0.6185 - auc: 0.9424 - prc: 0.7720 - val_loss: 0.0039 - val_tp: 61.0000 - val_fp: 9.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8714 - val_recall: 0.7093 - val_auc: 0.9124 - val_prc: 0.7881
Epoch 62/100
81/90 [==========================>...] - ETA: 0s - loss: 0.0034 - tp: 196.0000 - fp: 21.0000 - tn: 165565.0000 - fn: 106.0000 - accuracy: 0.9992 - precision: 0.9032 - recall: 0.6490 - auc: 0.9413 - prc: 0.7849Restoring model weights from the end of the best epoch: 52.
90/90 [==============================] - 1s 7ms/step - loss: 0.0034 - tp: 211.0000 - fp: 25.0000 - tn: 181926.0000 - fn: 114.0000 - accuracy: 0.9992 - precision: 0.8941 - recall: 0.6492 - auc: 0.9423 - prc: 0.7828 - val_loss: 0.0039 - val_tp: 64.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8649 - val_recall: 0.7442 - val_auc: 0.9124 - val_prc: 0.7860
Epoch 62: early stopping

Consultar historial de entrenamiento

En esta sección, producirá gráficos de la precisión y la pérdida de su modelo en el conjunto de entrenamiento y validación. Estos son útiles para verificar el sobreajuste, sobre el cual puede obtener más información en el tutorial sobreajuste y ajuste insuficiente .

Además, puede producir estos gráficos para cualquiera de las métricas que creó anteriormente. Los falsos negativos se incluyen como ejemplo.

def plot_metrics(history):
  metrics = ['loss', 'prc', 'precision', 'recall']
  for n, metric in enumerate(metrics):
    name = metric.replace("_"," ").capitalize()
    plt.subplot(2,2,n+1)
    plt.plot(history.epoch, history.history[metric], color=colors[0], label='Train')
    plt.plot(history.epoch, history.history['val_'+metric],
             color=colors[0], linestyle="--", label='Val')
    plt.xlabel('Epoch')
    plt.ylabel(name)
    if metric == 'loss':
      plt.ylim([0, plt.ylim()[1]])
    elif metric == 'auc':
      plt.ylim([0.8,1])
    else:
      plt.ylim([0,1])

    plt.legend();
plot_metrics(baseline_history)

png

Evaluar métricas

Puede usar una matriz de confusión para resumir las etiquetas reales frente a las predichas, donde el eje X es la etiqueta predicha y el eje Y es la etiqueta real:

train_predictions_baseline = model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_baseline = model.predict(test_features, batch_size=BATCH_SIZE)
def plot_cm(labels, predictions, p=0.5):
  cm = confusion_matrix(labels, predictions > p)
  plt.figure(figsize=(5,5))
  sns.heatmap(cm, annot=True, fmt="d")
  plt.title('Confusion matrix @{:.2f}'.format(p))
  plt.ylabel('Actual label')
  plt.xlabel('Predicted label')

  print('Legitimate Transactions Detected (True Negatives): ', cm[0][0])
  print('Legitimate Transactions Incorrectly Detected (False Positives): ', cm[0][1])
  print('Fraudulent Transactions Missed (False Negatives): ', cm[1][0])
  print('Fraudulent Transactions Detected (True Positives): ', cm[1][1])
  print('Total Fraudulent Transactions: ', np.sum(cm[1]))

Evalúe su modelo en el conjunto de datos de prueba y muestre los resultados de las métricas que creó anteriormente:

baseline_results = model.evaluate(test_features, test_labels,
                                  batch_size=BATCH_SIZE, verbose=0)
for name, value in zip(model.metrics_names, baseline_results):
  print(name, ': ', value)
print()

plot_cm(test_labels, test_predictions_baseline)
loss :  0.0024895435199141502
tp :  59.0
fp :  7.0
tn :  56874.0
fn :  22.0
accuracy :  0.9994909167289734
precision :  0.8939393758773804
recall :  0.7283950448036194
auc :  0.9318439960479736
prc :  0.8204483985900879

Legitimate Transactions Detected (True Negatives):  56874
Legitimate Transactions Incorrectly Detected (False Positives):  7
Fraudulent Transactions Missed (False Negatives):  22
Fraudulent Transactions Detected (True Positives):  59
Total Fraudulent Transactions:  81

png

Si el modelo hubiera predicho todo a la perfección, sería una matriz diagonal donde los valores fuera de la diagonal principal, que indican predicciones incorrectas, serían cero. En este caso, la matriz muestra que tiene relativamente pocos falsos positivos, lo que significa que hubo relativamente pocas transacciones legítimas que se marcaron incorrectamente. Sin embargo, es probable que desee tener aún menos falsos negativos a pesar del costo de aumentar la cantidad de falsos positivos. Esta compensación puede ser preferible porque los falsos negativos permitirían que se realicen transacciones fraudulentas, mientras que los falsos positivos pueden hacer que se envíe un correo electrónico a un cliente para pedirle que verifique la actividad de su tarjeta.

Trazar la República de China

Ahora traza la ROC . Este gráfico es útil porque muestra, de un vistazo, el rango de rendimiento que el modelo puede alcanzar simplemente ajustando el umbral de salida.

def plot_roc(name, labels, predictions, **kwargs):
  fp, tp, _ = sklearn.metrics.roc_curve(labels, predictions)

  plt.plot(100*fp, 100*tp, label=name, linewidth=2, **kwargs)
  plt.xlabel('False positives [%]')
  plt.ylabel('True positives [%]')
  plt.xlim([-0.5,20])
  plt.ylim([80,100.5])
  plt.grid(True)
  ax = plt.gca()
  ax.set_aspect('equal')
plot_roc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_roc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')
plt.legend(loc='lower right');

png

Trazar el AUPRC

Ahora traza el AUPRC . Área bajo la curva de precisión-recuperación interpolada, obtenida al graficar (recuperación, precisión) puntos para diferentes valores del umbral de clasificación. Dependiendo de cómo se calcule, PR AUC puede ser equivalente a la precisión promedio del modelo.

def plot_prc(name, labels, predictions, **kwargs):
    precision, recall, _ = sklearn.metrics.precision_recall_curve(labels, predictions)

    plt.plot(precision, recall, label=name, linewidth=2, **kwargs)
    plt.xlabel('Recall')
    plt.ylabel('Precision')
    plt.grid(True)
    ax = plt.gca()
    ax.set_aspect('equal')
plot_prc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_prc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')
plt.legend(loc='lower right');

png

Parece que la precisión es relativamente alta, pero la recuperación y el área bajo la curva ROC (AUC) no son tan altas como le gustaría. Los clasificadores a menudo enfrentan desafíos cuando intentan maximizar tanto la precisión como la recuperación, lo cual es especialmente cierto cuando se trabaja con conjuntos de datos desequilibrados. Es importante considerar los costos de los diferentes tipos de errores en el contexto del problema que le interesa. En este ejemplo, un falso negativo (se pierde una transacción fraudulenta) puede tener un costo financiero, mientras que un falso positivo (una transacción se marca incorrectamente como fraudulenta) puede disminuir la felicidad del usuario.

Pesos de clase

Calcular pesos de clase

El objetivo es identificar transacciones fraudulentas, pero no tiene muchas de esas muestras positivas con las que trabajar, por lo que le gustaría que el clasificador sopese mucho los pocos ejemplos disponibles. Puede hacer esto pasando los pesos de Keras para cada clase a través de un parámetro. Esto hará que el modelo "preste más atención" a los ejemplos de una clase subrepresentada.

# Scaling by total/2 helps keep the loss to a similar magnitude.
# The sum of the weights of all examples stays the same.
weight_for_0 = (1 / neg) * (total / 2.0)
weight_for_1 = (1 / pos) * (total / 2.0)

class_weight = {0: weight_for_0, 1: weight_for_1}

print('Weight for class 0: {:.2f}'.format(weight_for_0))
print('Weight for class 1: {:.2f}'.format(weight_for_1))
Weight for class 0: 0.50
Weight for class 1: 289.44

Entrena a un modelo con pesos de clase

Ahora intente volver a entrenar y evaluar el modelo con pesos de clase para ver cómo eso afecta las predicciones.

weighted_model = make_model()
weighted_model.load_weights(initial_weights)

weighted_history = weighted_model.fit(
    train_features,
    train_labels,
    batch_size=BATCH_SIZE,
    epochs=EPOCHS,
    callbacks=[early_stopping],
    validation_data=(val_features, val_labels),
    # The class weights go here
    class_weight=class_weight)
Epoch 1/100
90/90 [==============================] - 3s 15ms/step - loss: 4.1298 - tp: 59.0000 - fp: 11.0000 - tn: 238821.0000 - fn: 347.0000 - accuracy: 0.9985 - precision: 0.8429 - recall: 0.1453 - auc: 0.6238 - prc: 0.1649 - val_loss: 0.0119 - val_tp: 0.0000e+00 - val_fp: 0.0000e+00 - val_tn: 45483.0000 - val_fn: 86.0000 - val_accuracy: 0.9981 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00 - val_auc: 0.7124 - val_prc: 0.0294
Epoch 2/100
90/90 [==============================] - 1s 7ms/step - loss: 1.8711 - tp: 69.0000 - fp: 54.0000 - tn: 181897.0000 - fn: 256.0000 - accuracy: 0.9983 - precision: 0.5610 - recall: 0.2123 - auc: 0.8178 - prc: 0.2117 - val_loss: 0.0060 - val_tp: 56.0000 - val_fp: 10.0000 - val_tn: 45473.0000 - val_fn: 30.0000 - val_accuracy: 0.9991 - val_precision: 0.8485 - val_recall: 0.6512 - val_auc: 0.9427 - val_prc: 0.6870
Epoch 3/100
90/90 [==============================] - 1s 7ms/step - loss: 0.8666 - tp: 187.0000 - fp: 198.0000 - tn: 181753.0000 - fn: 138.0000 - accuracy: 0.9982 - precision: 0.4857 - recall: 0.5754 - auc: 0.9075 - prc: 0.4912 - val_loss: 0.0077 - val_tp: 65.0000 - val_fp: 19.0000 - val_tn: 45464.0000 - val_fn: 21.0000 - val_accuracy: 0.9991 - val_precision: 0.7738 - val_recall: 0.7558 - val_auc: 0.9564 - val_prc: 0.6924
Epoch 4/100
90/90 [==============================] - 1s 7ms/step - loss: 0.6876 - tp: 218.0000 - fp: 530.0000 - tn: 181421.0000 - fn: 107.0000 - accuracy: 0.9965 - precision: 0.2914 - recall: 0.6708 - auc: 0.9152 - prc: 0.5102 - val_loss: 0.0109 - val_tp: 68.0000 - val_fp: 39.0000 - val_tn: 45444.0000 - val_fn: 18.0000 - val_accuracy: 0.9987 - val_precision: 0.6355 - val_recall: 0.7907 - val_auc: 0.9661 - val_prc: 0.6926
Epoch 5/100
90/90 [==============================] - 1s 7ms/step - loss: 0.5229 - tp: 240.0000 - fp: 1102.0000 - tn: 180849.0000 - fn: 85.0000 - accuracy: 0.9935 - precision: 0.1788 - recall: 0.7385 - auc: 0.9395 - prc: 0.5228 - val_loss: 0.0154 - val_tp: 70.0000 - val_fp: 79.0000 - val_tn: 45404.0000 - val_fn: 16.0000 - val_accuracy: 0.9979 - val_precision: 0.4698 - val_recall: 0.8140 - val_auc: 0.9657 - val_prc: 0.7023
Epoch 6/100
90/90 [==============================] - 1s 7ms/step - loss: 0.4753 - tp: 251.0000 - fp: 1839.0000 - tn: 180112.0000 - fn: 74.0000 - accuracy: 0.9895 - precision: 0.1201 - recall: 0.7723 - auc: 0.9336 - prc: 0.4297 - val_loss: 0.0213 - val_tp: 70.0000 - val_fp: 156.0000 - val_tn: 45327.0000 - val_fn: 16.0000 - val_accuracy: 0.9962 - val_precision: 0.3097 - val_recall: 0.8140 - val_auc: 0.9654 - val_prc: 0.6742
Epoch 7/100
90/90 [==============================] - 1s 7ms/step - loss: 0.3870 - tp: 270.0000 - fp: 2554.0000 - tn: 179397.0000 - fn: 55.0000 - accuracy: 0.9857 - precision: 0.0956 - recall: 0.8308 - auc: 0.9463 - prc: 0.3800 - val_loss: 0.0269 - val_tp: 70.0000 - val_fp: 264.0000 - val_tn: 45219.0000 - val_fn: 16.0000 - val_accuracy: 0.9939 - val_precision: 0.2096 - val_recall: 0.8140 - val_auc: 0.9651 - val_prc: 0.6116
Epoch 8/100
90/90 [==============================] - 1s 7ms/step - loss: 0.3942 - tp: 268.0000 - fp: 3219.0000 - tn: 178732.0000 - fn: 57.0000 - accuracy: 0.9820 - precision: 0.0769 - recall: 0.8246 - auc: 0.9434 - prc: 0.3273 - val_loss: 0.0337 - val_tp: 70.0000 - val_fp: 355.0000 - val_tn: 45128.0000 - val_fn: 16.0000 - val_accuracy: 0.9919 - val_precision: 0.1647 - val_recall: 0.8140 - val_auc: 0.9682 - val_prc: 0.5918
Epoch 9/100
90/90 [==============================] - 1s 7ms/step - loss: 0.3886 - tp: 271.0000 - fp: 3845.0000 - tn: 178106.0000 - fn: 54.0000 - accuracy: 0.9786 - precision: 0.0658 - recall: 0.8338 - auc: 0.9397 - prc: 0.2995 - val_loss: 0.0386 - val_tp: 70.0000 - val_fp: 406.0000 - val_tn: 45077.0000 - val_fn: 16.0000 - val_accuracy: 0.9907 - val_precision: 0.1471 - val_recall: 0.8140 - val_auc: 0.9756 - val_prc: 0.5889
Epoch 10/100
90/90 [==============================] - 1s 7ms/step - loss: 0.2951 - tp: 281.0000 - fp: 4348.0000 - tn: 177603.0000 - fn: 44.0000 - accuracy: 0.9759 - precision: 0.0607 - recall: 0.8646 - auc: 0.9623 - prc: 0.2826 - val_loss: 0.0441 - val_tp: 72.0000 - val_fp: 464.0000 - val_tn: 45019.0000 - val_fn: 14.0000 - val_accuracy: 0.9895 - val_precision: 0.1343 - val_recall: 0.8372 - val_auc: 0.9748 - val_prc: 0.5895
Epoch 11/100
90/90 [==============================] - 1s 7ms/step - loss: 0.2703 - tp: 280.0000 - fp: 4697.0000 - tn: 177254.0000 - fn: 45.0000 - accuracy: 0.9740 - precision: 0.0563 - recall: 0.8615 - auc: 0.9660 - prc: 0.2589 - val_loss: 0.0490 - val_tp: 72.0000 - val_fp: 552.0000 - val_tn: 44931.0000 - val_fn: 14.0000 - val_accuracy: 0.9876 - val_precision: 0.1154 - val_recall: 0.8372 - val_auc: 0.9762 - val_prc: 0.5902
Epoch 12/100
90/90 [==============================] - 1s 7ms/step - loss: 0.3358 - tp: 278.0000 - fp: 5262.0000 - tn: 176689.0000 - fn: 47.0000 - accuracy: 0.9709 - precision: 0.0502 - recall: 0.8554 - auc: 0.9468 - prc: 0.2368 - val_loss: 0.0534 - val_tp: 74.0000 - val_fp: 597.0000 - val_tn: 44886.0000 - val_fn: 12.0000 - val_accuracy: 0.9866 - val_precision: 0.1103 - val_recall: 0.8605 - val_auc: 0.9752 - val_prc: 0.5848
Epoch 13/100
90/90 [==============================] - 1s 7ms/step - loss: 0.2833 - tp: 286.0000 - fp: 5502.0000 - tn: 176449.0000 - fn: 39.0000 - accuracy: 0.9696 - precision: 0.0494 - recall: 0.8800 - auc: 0.9582 - prc: 0.2572 - val_loss: 0.0563 - val_tp: 74.0000 - val_fp: 616.0000 - val_tn: 44867.0000 - val_fn: 12.0000 - val_accuracy: 0.9862 - val_precision: 0.1072 - val_recall: 0.8605 - val_auc: 0.9748 - val_prc: 0.5678
Epoch 14/100
90/90 [==============================] - 1s 7ms/step - loss: 0.2969 - tp: 280.0000 - fp: 5630.0000 - tn: 176321.0000 - fn: 45.0000 - accuracy: 0.9689 - precision: 0.0474 - recall: 0.8615 - auc: 0.9594 - prc: 0.2374 - val_loss: 0.0597 - val_tp: 74.0000 - val_fp: 644.0000 - val_tn: 44839.0000 - val_fn: 12.0000 - val_accuracy: 0.9856 - val_precision: 0.1031 - val_recall: 0.8605 - val_auc: 0.9741 - val_prc: 0.5627
Epoch 15/100
90/90 [==============================] - ETA: 0s - loss: 0.3183 - tp: 280.0000 - fp: 5954.0000 - tn: 175997.0000 - fn: 45.0000 - accuracy: 0.9671 - precision: 0.0449 - recall: 0.8615 - auc: 0.9496 - prc: 0.2224Restoring model weights from the end of the best epoch: 5.
90/90 [==============================] - 1s 7ms/step - loss: 0.3183 - tp: 280.0000 - fp: 5954.0000 - tn: 175997.0000 - fn: 45.0000 - accuracy: 0.9671 - precision: 0.0449 - recall: 0.8615 - auc: 0.9496 - prc: 0.2224 - val_loss: 0.0621 - val_tp: 74.0000 - val_fp: 665.0000 - val_tn: 44818.0000 - val_fn: 12.0000 - val_accuracy: 0.9851 - val_precision: 0.1001 - val_recall: 0.8605 - val_auc: 0.9771 - val_prc: 0.5550
Epoch 15: early stopping

Consultar historial de entrenamiento

plot_metrics(weighted_history)

png

Evaluar métricas

train_predictions_weighted = weighted_model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_weighted = weighted_model.predict(test_features, batch_size=BATCH_SIZE)
weighted_results = weighted_model.evaluate(test_features, test_labels,
                                           batch_size=BATCH_SIZE, verbose=0)
for name, value in zip(weighted_model.metrics_names, weighted_results):
  print(name, ': ', value)
print()

plot_cm(test_labels, test_predictions_weighted)
loss :  0.014327289536595345
tp :  69.0
fp :  88.0
tn :  56793.0
fn :  12.0
accuracy :  0.9982444643974304
precision :  0.4394904375076294
recall :  0.8518518805503845
auc :  0.9410961866378784
prc :  0.7397712469100952

Legitimate Transactions Detected (True Negatives):  56793
Legitimate Transactions Incorrectly Detected (False Positives):  88
Fraudulent Transactions Missed (False Negatives):  12
Fraudulent Transactions Detected (True Positives):  69
Total Fraudulent Transactions:  81

png

Aquí puede ver que con los pesos de clase, la exactitud y la precisión son más bajas porque hay más falsos positivos, pero a la inversa, el recuerdo y el AUC son más altos porque el modelo también encontró más positivos verdaderos. A pesar de tener una menor precisión, este modelo tiene un mayor recuerdo (e identifica más transacciones fraudulentas). Por supuesto, ambos tipos de error tienen un costo (tampoco querrá molestar a los usuarios marcando demasiadas transacciones legítimas como fraudulentas). Considere cuidadosamente las compensaciones entre estos diferentes tipos de errores para su aplicación.

Trazar la República de China

plot_roc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_roc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')

plot_roc("Train Weighted", train_labels, train_predictions_weighted, color=colors[1])
plot_roc("Test Weighted", test_labels, test_predictions_weighted, color=colors[1], linestyle='--')


plt.legend(loc='lower right');

png

Trazar el AUPRC

plot_prc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_prc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')

plot_prc("Train Weighted", train_labels, train_predictions_weighted, color=colors[1])
plot_prc("Test Weighted", test_labels, test_predictions_weighted, color=colors[1], linestyle='--')


plt.legend(loc='lower right');

png

sobremuestreo

Sobremuestra la clase minoritaria

Un enfoque relacionado sería volver a muestrear el conjunto de datos sobremuestreando la clase minoritaria.

pos_features = train_features[bool_train_labels]
neg_features = train_features[~bool_train_labels]

pos_labels = train_labels[bool_train_labels]
neg_labels = train_labels[~bool_train_labels]

Usando NumPy

Puede equilibrar el conjunto de datos manualmente eligiendo el número correcto de índices aleatorios de los ejemplos positivos:

ids = np.arange(len(pos_features))
choices = np.random.choice(ids, len(neg_features))

res_pos_features = pos_features[choices]
res_pos_labels = pos_labels[choices]

res_pos_features.shape
(181951, 29)
resampled_features = np.concatenate([res_pos_features, neg_features], axis=0)
resampled_labels = np.concatenate([res_pos_labels, neg_labels], axis=0)

order = np.arange(len(resampled_labels))
np.random.shuffle(order)
resampled_features = resampled_features[order]
resampled_labels = resampled_labels[order]

resampled_features.shape
(363902, 29)

Usando tf.data

Si está utilizando tf.data la forma más fácil de producir ejemplos equilibrados es comenzar con un conjunto de datos positive y negative , y fusionarlos. Consulte la guía tf.data para obtener más ejemplos.

BUFFER_SIZE = 100000

def make_ds(features, labels):
  ds = tf.data.Dataset.from_tensor_slices((features, labels))#.cache()
  ds = ds.shuffle(BUFFER_SIZE).repeat()
  return ds

pos_ds = make_ds(pos_features, pos_labels)
neg_ds = make_ds(neg_features, neg_labels)

Cada conjunto de datos proporciona pares (feature, label) :

for features, label in pos_ds.take(1):
  print("Features:\n", features.numpy())
  print()
  print("Label: ", label.numpy())
Features:
 [ 0.56826828  1.24841849 -2.52251105  3.84165891  0.05052604 -0.7621795
 -1.43118352  0.43296139 -1.85102109 -2.50477555  3.20133397 -3.52460861
 -0.95133935 -5.         -1.93144512 -0.7302767  -2.46735228  0.21827555
 -1.45046438  0.21081234  0.39176826 -0.23558789 -0.03611637 -0.62063738
  0.3686766   0.23622961  1.2242418   0.75555829 -1.45589162]

Label:  1

Combine los dos usando tf.data.Dataset.sample_from_datasets :

resampled_ds = tf.data.Dataset.sample_from_datasets([pos_ds, neg_ds], weights=[0.5, 0.5])
resampled_ds = resampled_ds.batch(BATCH_SIZE).prefetch(2)
for features, label in resampled_ds.take(1):
  print(label.numpy().mean())
0.50732421875

Para usar este conjunto de datos, necesitará la cantidad de pasos por época.

La definición de "época" en este caso es menos clara. Digamos que es la cantidad de lotes necesarios para ver cada ejemplo negativo una vez:

resampled_steps_per_epoch = np.ceil(2.0*neg/BATCH_SIZE)
resampled_steps_per_epoch
278.0

Entrene en los datos sobremuestreados

Ahora intente entrenar el modelo con el conjunto de datos remuestreado en lugar de usar pesos de clase para ver cómo se comparan estos métodos.

resampled_model = make_model()
resampled_model.load_weights(initial_weights)

# Reset the bias to zero, since this dataset is balanced.
output_layer = resampled_model.layers[-1] 
output_layer.bias.assign([0])

val_ds = tf.data.Dataset.from_tensor_slices((val_features, val_labels)).cache()
val_ds = val_ds.batch(BATCH_SIZE).prefetch(2) 

resampled_history = resampled_model.fit(
    resampled_ds,
    epochs=EPOCHS,
    steps_per_epoch=resampled_steps_per_epoch,
    callbacks=[early_stopping],
    validation_data=val_ds)
Epoch 1/100
278/278 [==============================] - 10s 32ms/step - loss: 0.5508 - tp: 214194.0000 - fp: 51114.0000 - tn: 290615.0000 - fn: 70383.0000 - accuracy: 0.8060 - precision: 0.8073 - recall: 0.7527 - auc: 0.8600 - prc: 0.8879 - val_loss: 0.2279 - val_tp: 73.0000 - val_fp: 969.0000 - val_tn: 44514.0000 - val_fn: 13.0000 - val_accuracy: 0.9785 - val_precision: 0.0701 - val_recall: 0.8488 - val_auc: 0.9551 - val_prc: 0.7044
Epoch 2/100
278/278 [==============================] - 8s 28ms/step - loss: 0.2235 - tp: 253877.0000 - fp: 15743.0000 - tn: 268530.0000 - fn: 31194.0000 - accuracy: 0.9176 - precision: 0.9416 - recall: 0.8906 - auc: 0.9658 - prc: 0.9746 - val_loss: 0.1367 - val_tp: 73.0000 - val_fp: 777.0000 - val_tn: 44706.0000 - val_fn: 13.0000 - val_accuracy: 0.9827 - val_precision: 0.0859 - val_recall: 0.8488 - val_auc: 0.9596 - val_prc: 0.7072
Epoch 3/100
278/278 [==============================] - 8s 28ms/step - loss: 0.1785 - tp: 258572.0000 - fp: 9840.0000 - tn: 274878.0000 - fn: 26054.0000 - accuracy: 0.9370 - precision: 0.9633 - recall: 0.9085 - auc: 0.9773 - prc: 0.9827 - val_loss: 0.1023 - val_tp: 72.0000 - val_fp: 699.0000 - val_tn: 44784.0000 - val_fn: 14.0000 - val_accuracy: 0.9844 - val_precision: 0.0934 - val_recall: 0.8372 - val_auc: 0.9632 - val_prc: 0.7032
Epoch 4/100
278/278 [==============================] - 8s 29ms/step - loss: 0.1571 - tp: 260447.0000 - fp: 8085.0000 - tn: 276389.0000 - fn: 24423.0000 - accuracy: 0.9429 - precision: 0.9699 - recall: 0.9143 - auc: 0.9826 - prc: 0.9863 - val_loss: 0.0869 - val_tp: 74.0000 - val_fp: 701.0000 - val_tn: 44782.0000 - val_fn: 12.0000 - val_accuracy: 0.9844 - val_precision: 0.0955 - val_recall: 0.8605 - val_auc: 0.9633 - val_prc: 0.6972
Epoch 5/100
278/278 [==============================] - 8s 30ms/step - loss: 0.1440 - tp: 261457.0000 - fp: 7449.0000 - tn: 277093.0000 - fn: 23345.0000 - accuracy: 0.9459 - precision: 0.9723 - recall: 0.9180 - auc: 0.9855 - prc: 0.9883 - val_loss: 0.0774 - val_tp: 73.0000 - val_fp: 679.0000 - val_tn: 44804.0000 - val_fn: 13.0000 - val_accuracy: 0.9848 - val_precision: 0.0971 - val_recall: 0.8488 - val_auc: 0.9645 - val_prc: 0.6971
Epoch 6/100
278/278 [==============================] - 8s 28ms/step - loss: 0.1349 - tp: 262460.0000 - fp: 6942.0000 - tn: 277723.0000 - fn: 22219.0000 - accuracy: 0.9488 - precision: 0.9742 - recall: 0.9220 - auc: 0.9876 - prc: 0.9896 - val_loss: 0.0718 - val_tp: 74.0000 - val_fp: 624.0000 - val_tn: 44859.0000 - val_fn: 12.0000 - val_accuracy: 0.9860 - val_precision: 0.1060 - val_recall: 0.8605 - val_auc: 0.9645 - val_prc: 0.6891
Epoch 7/100
278/278 [==============================] - 8s 28ms/step - loss: 0.1264 - tp: 263166.0000 - fp: 6780.0000 - tn: 278253.0000 - fn: 21145.0000 - accuracy: 0.9510 - precision: 0.9749 - recall: 0.9256 - auc: 0.9895 - prc: 0.9909 - val_loss: 0.0672 - val_tp: 75.0000 - val_fp: 602.0000 - val_tn: 44881.0000 - val_fn: 11.0000 - val_accuracy: 0.9865 - val_precision: 0.1108 - val_recall: 0.8721 - val_auc: 0.9670 - val_prc: 0.6822
Epoch 8/100
278/278 [==============================] - 8s 30ms/step - loss: 0.1190 - tp: 264216.0000 - fp: 6569.0000 - tn: 278270.0000 - fn: 20289.0000 - accuracy: 0.9528 - precision: 0.9757 - recall: 0.9287 - auc: 0.9910 - prc: 0.9920 - val_loss: 0.0628 - val_tp: 74.0000 - val_fp: 570.0000 - val_tn: 44913.0000 - val_fn: 12.0000 - val_accuracy: 0.9872 - val_precision: 0.1149 - val_recall: 0.8605 - val_auc: 0.9671 - val_prc: 0.6830
Epoch 9/100
278/278 [==============================] - 9s 31ms/step - loss: 0.1125 - tp: 264562.0000 - fp: 6339.0000 - tn: 279137.0000 - fn: 19306.0000 - accuracy: 0.9550 - precision: 0.9766 - recall: 0.9320 - auc: 0.9924 - prc: 0.9930 - val_loss: 0.0576 - val_tp: 74.0000 - val_fp: 544.0000 - val_tn: 44939.0000 - val_fn: 12.0000 - val_accuracy: 0.9878 - val_precision: 0.1197 - val_recall: 0.8605 - val_auc: 0.9672 - val_prc: 0.6828
Epoch 10/100
278/278 [==============================] - 8s 30ms/step - loss: 0.1064 - tp: 266549.0000 - fp: 6112.0000 - tn: 278323.0000 - fn: 18360.0000 - accuracy: 0.9570 - precision: 0.9776 - recall: 0.9356 - auc: 0.9934 - prc: 0.9937 - val_loss: 0.0544 - val_tp: 74.0000 - val_fp: 541.0000 - val_tn: 44942.0000 - val_fn: 12.0000 - val_accuracy: 0.9879 - val_precision: 0.1203 - val_recall: 0.8605 - val_auc: 0.9638 - val_prc: 0.6827
Epoch 11/100
278/278 [==============================] - 8s 30ms/step - loss: 0.1005 - tp: 267048.0000 - fp: 6123.0000 - tn: 278896.0000 - fn: 17277.0000 - accuracy: 0.9589 - precision: 0.9776 - recall: 0.9392 - auc: 0.9943 - prc: 0.9944 - val_loss: 0.0493 - val_tp: 74.0000 - val_fp: 500.0000 - val_tn: 44983.0000 - val_fn: 12.0000 - val_accuracy: 0.9888 - val_precision: 0.1289 - val_recall: 0.8605 - val_auc: 0.9578 - val_prc: 0.6761
Epoch 12/100
277/278 [============================>.] - ETA: 0s - loss: 0.0950 - tp: 266855.0000 - fp: 6079.0000 - tn: 277677.0000 - fn: 16685.0000 - accuracy: 0.9599 - precision: 0.9777 - recall: 0.9412 - auc: 0.9950 - prc: 0.9949Restoring model weights from the end of the best epoch: 2.
278/278 [==============================] - 8s 29ms/step - loss: 0.0950 - tp: 267815.0000 - fp: 6094.0000 - tn: 278693.0000 - fn: 16742.0000 - accuracy: 0.9599 - precision: 0.9778 - recall: 0.9412 - auc: 0.9950 - prc: 0.9949 - val_loss: 0.0451 - val_tp: 74.0000 - val_fp: 468.0000 - val_tn: 45015.0000 - val_fn: 12.0000 - val_accuracy: 0.9895 - val_precision: 0.1365 - val_recall: 0.8605 - val_auc: 0.9581 - val_prc: 0.6683
Epoch 12: early stopping

Si el proceso de entrenamiento considerara todo el conjunto de datos en cada actualización de gradiente, este sobremuestreo sería básicamente idéntico a la ponderación de clase.

Pero al entrenar el modelo por lotes, como lo hizo aquí, los datos sobremuestreados proporcionan una señal de gradiente más suave: en lugar de que cada ejemplo positivo se muestre en un lote con un peso grande, se muestran en muchos lotes diferentes cada vez con un pequeño peso

Esta señal de gradiente más suave facilita el entrenamiento del modelo.

Consultar historial de entrenamiento

Tenga en cuenta que las distribuciones de métricas serán diferentes aquí, porque los datos de entrenamiento tienen una distribución totalmente diferente de los datos de validación y prueba.

plot_metrics(resampled_history)

png

volver a entrenar

Debido a que el entrenamiento es más fácil con los datos balanceados, el procedimiento de entrenamiento anterior puede sobreajustarse rápidamente.

Por lo tanto, divida las épocas para darle a tf.keras.callbacks.EarlyStopping un control más preciso sobre cuándo detener el entrenamiento.

resampled_model = make_model()
resampled_model.load_weights(initial_weights)

# Reset the bias to zero, since this dataset is balanced.
output_layer = resampled_model.layers[-1] 
output_layer.bias.assign([0])

resampled_history = resampled_model.fit(
    resampled_ds,
    # These are not real epochs
    steps_per_epoch=20,
    epochs=10*EPOCHS,
    callbacks=[early_stopping],
    validation_data=(val_ds))
Epoch 1/1000
20/20 [==============================] - 3s 73ms/step - loss: 2.0114 - tp: 3382.0000 - fp: 5181.0000 - tn: 60589.0000 - fn: 17377.0000 - accuracy: 0.7393 - precision: 0.3950 - recall: 0.1629 - auc: 0.6308 - prc: 0.3325 - val_loss: 0.4343 - val_tp: 7.0000 - val_fp: 5042.0000 - val_tn: 40441.0000 - val_fn: 79.0000 - val_accuracy: 0.8876 - val_precision: 0.0014 - val_recall: 0.0814 - val_auc: 0.2282 - val_prc: 0.0012
Epoch 2/1000
20/20 [==============================] - 1s 33ms/step - loss: 1.2163 - tp: 7466.0000 - fp: 5137.0000 - tn: 15257.0000 - fn: 13100.0000 - accuracy: 0.5548 - precision: 0.5924 - recall: 0.3630 - auc: 0.4763 - prc: 0.5716 - val_loss: 0.4539 - val_tp: 36.0000 - val_fp: 5893.0000 - val_tn: 39590.0000 - val_fn: 50.0000 - val_accuracy: 0.8696 - val_precision: 0.0061 - val_recall: 0.4186 - val_auc: 0.6494 - val_prc: 0.0054
Epoch 3/1000
20/20 [==============================] - 1s 33ms/step - loss: 0.7406 - tp: 12289.0000 - fp: 5509.0000 - tn: 14872.0000 - fn: 8290.0000 - accuracy: 0.6631 - precision: 0.6905 - recall: 0.5972 - auc: 0.6803 - prc: 0.7580 - val_loss: 0.4611 - val_tp: 75.0000 - val_fp: 6273.0000 - val_tn: 39210.0000 - val_fn: 11.0000 - val_accuracy: 0.8621 - val_precision: 0.0118 - val_recall: 0.8721 - val_auc: 0.9293 - val_prc: 0.4539
Epoch 4/1000
20/20 [==============================] - 1s 33ms/step - loss: 0.5071 - tp: 15891.0000 - fp: 5370.0000 - tn: 15013.0000 - fn: 4686.0000 - accuracy: 0.7545 - precision: 0.7474 - recall: 0.7723 - auc: 0.8298 - prc: 0.8757 - val_loss: 0.4451 - val_tp: 78.0000 - val_fp: 5505.0000 - val_tn: 39978.0000 - val_fn: 8.0000 - val_accuracy: 0.8790 - val_precision: 0.0140 - val_recall: 0.9070 - val_auc: 0.9443 - val_prc: 0.6777
Epoch 5/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.4284 - tp: 17046.0000 - fp: 5072.0000 - tn: 15496.0000 - fn: 3346.0000 - accuracy: 0.7945 - precision: 0.7707 - recall: 0.8359 - auc: 0.8827 - prc: 0.9151 - val_loss: 0.4140 - val_tp: 77.0000 - val_fp: 4338.0000 - val_tn: 41145.0000 - val_fn: 9.0000 - val_accuracy: 0.9046 - val_precision: 0.0174 - val_recall: 0.8953 - val_auc: 0.9463 - val_prc: 0.6903
Epoch 6/1000
20/20 [==============================] - 1s 33ms/step - loss: 0.3836 - tp: 17606.0000 - fp: 4362.0000 - tn: 16113.0000 - fn: 2879.0000 - accuracy: 0.8232 - precision: 0.8014 - recall: 0.8595 - auc: 0.9080 - prc: 0.9336 - val_loss: 0.3824 - val_tp: 77.0000 - val_fp: 3314.0000 - val_tn: 42169.0000 - val_fn: 9.0000 - val_accuracy: 0.9271 - val_precision: 0.0227 - val_recall: 0.8953 - val_auc: 0.9475 - val_prc: 0.6752
Epoch 7/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.3574 - tp: 17856.0000 - fp: 3894.0000 - tn: 16553.0000 - fn: 2657.0000 - accuracy: 0.8401 - precision: 0.8210 - recall: 0.8705 - auc: 0.9208 - prc: 0.9432 - val_loss: 0.3538 - val_tp: 76.0000 - val_fp: 2592.0000 - val_tn: 42891.0000 - val_fn: 10.0000 - val_accuracy: 0.9429 - val_precision: 0.0285 - val_recall: 0.8837 - val_auc: 0.9486 - val_prc: 0.6819
Epoch 8/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.3377 - tp: 17766.0000 - fp: 3483.0000 - tn: 17067.0000 - fn: 2644.0000 - accuracy: 0.8504 - precision: 0.8361 - recall: 0.8705 - auc: 0.9280 - prc: 0.9481 - val_loss: 0.3271 - val_tp: 76.0000 - val_fp: 2047.0000 - val_tn: 43436.0000 - val_fn: 10.0000 - val_accuracy: 0.9549 - val_precision: 0.0358 - val_recall: 0.8837 - val_auc: 0.9497 - val_prc: 0.6910
Epoch 9/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.3188 - tp: 17749.0000 - fp: 2855.0000 - tn: 17547.0000 - fn: 2809.0000 - accuracy: 0.8617 - precision: 0.8614 - recall: 0.8634 - auc: 0.9360 - prc: 0.9539 - val_loss: 0.3051 - val_tp: 74.0000 - val_fp: 1657.0000 - val_tn: 43826.0000 - val_fn: 12.0000 - val_accuracy: 0.9634 - val_precision: 0.0427 - val_recall: 0.8605 - val_auc: 0.9514 - val_prc: 0.7022
Epoch 10/1000
20/20 [==============================] - 1s 33ms/step - loss: 0.3046 - tp: 17772.0000 - fp: 2599.0000 - tn: 17841.0000 - fn: 2748.0000 - accuracy: 0.8695 - precision: 0.8724 - recall: 0.8661 - auc: 0.9402 - prc: 0.9570 - val_loss: 0.2860 - val_tp: 74.0000 - val_fp: 1398.0000 - val_tn: 44085.0000 - val_fn: 12.0000 - val_accuracy: 0.9691 - val_precision: 0.0503 - val_recall: 0.8605 - val_auc: 0.9527 - val_prc: 0.6997
Epoch 11/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2937 - tp: 17673.0000 - fp: 2352.0000 - tn: 18273.0000 - fn: 2662.0000 - accuracy: 0.8776 - precision: 0.8825 - recall: 0.8691 - auc: 0.9447 - prc: 0.9595 - val_loss: 0.2687 - val_tp: 73.0000 - val_fp: 1235.0000 - val_tn: 44248.0000 - val_fn: 13.0000 - val_accuracy: 0.9726 - val_precision: 0.0558 - val_recall: 0.8488 - val_auc: 0.9534 - val_prc: 0.7066
Epoch 12/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2813 - tp: 17721.0000 - fp: 2109.0000 - tn: 18523.0000 - fn: 2607.0000 - accuracy: 0.8849 - precision: 0.8936 - recall: 0.8718 - auc: 0.9485 - prc: 0.9621 - val_loss: 0.2524 - val_tp: 73.0000 - val_fp: 1098.0000 - val_tn: 44385.0000 - val_fn: 13.0000 - val_accuracy: 0.9756 - val_precision: 0.0623 - val_recall: 0.8488 - val_auc: 0.9539 - val_prc: 0.7094
Epoch 13/1000
20/20 [==============================] - 1s 36ms/step - loss: 0.2706 - tp: 18031.0000 - fp: 1869.0000 - tn: 18502.0000 - fn: 2558.0000 - accuracy: 0.8919 - precision: 0.9061 - recall: 0.8758 - auc: 0.9520 - prc: 0.9652 - val_loss: 0.2395 - val_tp: 73.0000 - val_fp: 1037.0000 - val_tn: 44446.0000 - val_fn: 13.0000 - val_accuracy: 0.9770 - val_precision: 0.0658 - val_recall: 0.8488 - val_auc: 0.9549 - val_prc: 0.7119
Epoch 14/1000
20/20 [==============================] - 1s 37ms/step - loss: 0.2665 - tp: 18087.0000 - fp: 1748.0000 - tn: 18567.0000 - fn: 2558.0000 - accuracy: 0.8949 - precision: 0.9119 - recall: 0.8761 - auc: 0.9525 - prc: 0.9661 - val_loss: 0.2283 - val_tp: 73.0000 - val_fp: 972.0000 - val_tn: 44511.0000 - val_fn: 13.0000 - val_accuracy: 0.9784 - val_precision: 0.0699 - val_recall: 0.8488 - val_auc: 0.9556 - val_prc: 0.7045
Epoch 15/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2589 - tp: 18064.0000 - fp: 1630.0000 - tn: 18830.0000 - fn: 2436.0000 - accuracy: 0.9007 - precision: 0.9172 - recall: 0.8812 - auc: 0.9560 - prc: 0.9676 - val_loss: 0.2180 - val_tp: 73.0000 - val_fp: 941.0000 - val_tn: 44542.0000 - val_fn: 13.0000 - val_accuracy: 0.9791 - val_precision: 0.0720 - val_recall: 0.8488 - val_auc: 0.9563 - val_prc: 0.7069
Epoch 16/1000
20/20 [==============================] - 1s 36ms/step - loss: 0.2495 - tp: 18132.0000 - fp: 1481.0000 - tn: 18926.0000 - fn: 2421.0000 - accuracy: 0.9047 - precision: 0.9245 - recall: 0.8822 - auc: 0.9587 - prc: 0.9695 - val_loss: 0.2079 - val_tp: 73.0000 - val_fp: 905.0000 - val_tn: 44578.0000 - val_fn: 13.0000 - val_accuracy: 0.9799 - val_precision: 0.0746 - val_recall: 0.8488 - val_auc: 0.9565 - val_prc: 0.7110
Epoch 17/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.2435 - tp: 18047.0000 - fp: 1378.0000 - tn: 19144.0000 - fn: 2391.0000 - accuracy: 0.9080 - precision: 0.9291 - recall: 0.8830 - auc: 0.9601 - prc: 0.9706 - val_loss: 0.1990 - val_tp: 73.0000 - val_fp: 882.0000 - val_tn: 44601.0000 - val_fn: 13.0000 - val_accuracy: 0.9804 - val_precision: 0.0764 - val_recall: 0.8488 - val_auc: 0.9568 - val_prc: 0.7118
Epoch 18/1000
20/20 [==============================] - 1s 37ms/step - loss: 0.2396 - tp: 18223.0000 - fp: 1289.0000 - tn: 19075.0000 - fn: 2373.0000 - accuracy: 0.9106 - precision: 0.9339 - recall: 0.8848 - auc: 0.9612 - prc: 0.9714 - val_loss: 0.1911 - val_tp: 73.0000 - val_fp: 870.0000 - val_tn: 44613.0000 - val_fn: 13.0000 - val_accuracy: 0.9806 - val_precision: 0.0774 - val_recall: 0.8488 - val_auc: 0.9573 - val_prc: 0.7148
Epoch 19/1000
20/20 [==============================] - 1s 36ms/step - loss: 0.2324 - tp: 18179.0000 - fp: 1205.0000 - tn: 19254.0000 - fn: 2322.0000 - accuracy: 0.9139 - precision: 0.9378 - recall: 0.8867 - auc: 0.9633 - prc: 0.9728 - val_loss: 0.1839 - val_tp: 73.0000 - val_fp: 857.0000 - val_tn: 44626.0000 - val_fn: 13.0000 - val_accuracy: 0.9809 - val_precision: 0.0785 - val_recall: 0.8488 - val_auc: 0.9576 - val_prc: 0.7165
Epoch 20/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2318 - tp: 18119.0000 - fp: 1224.0000 - tn: 19279.0000 - fn: 2338.0000 - accuracy: 0.9130 - precision: 0.9367 - recall: 0.8857 - auc: 0.9640 - prc: 0.9728 - val_loss: 0.1758 - val_tp: 73.0000 - val_fp: 823.0000 - val_tn: 44660.0000 - val_fn: 13.0000 - val_accuracy: 0.9817 - val_precision: 0.0815 - val_recall: 0.8488 - val_auc: 0.9573 - val_prc: 0.7185
Epoch 21/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.2233 - tp: 18041.0000 - fp: 1074.0000 - tn: 19514.0000 - fn: 2331.0000 - accuracy: 0.9169 - precision: 0.9438 - recall: 0.8856 - auc: 0.9660 - prc: 0.9745 - val_loss: 0.1690 - val_tp: 73.0000 - val_fp: 813.0000 - val_tn: 44670.0000 - val_fn: 13.0000 - val_accuracy: 0.9819 - val_precision: 0.0824 - val_recall: 0.8488 - val_auc: 0.9578 - val_prc: 0.7211
Epoch 22/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.2193 - tp: 18258.0000 - fp: 1013.0000 - tn: 19414.0000 - fn: 2275.0000 - accuracy: 0.9197 - precision: 0.9474 - recall: 0.8892 - auc: 0.9666 - prc: 0.9753 - val_loss: 0.1634 - val_tp: 73.0000 - val_fp: 817.0000 - val_tn: 44666.0000 - val_fn: 13.0000 - val_accuracy: 0.9818 - val_precision: 0.0820 - val_recall: 0.8488 - val_auc: 0.9580 - val_prc: 0.7123
Epoch 23/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2114 - tp: 18439.0000 - fp: 993.0000 - tn: 19417.0000 - fn: 2111.0000 - accuracy: 0.9242 - precision: 0.9489 - recall: 0.8973 - auc: 0.9696 - prc: 0.9774 - val_loss: 0.1577 - val_tp: 73.0000 - val_fp: 807.0000 - val_tn: 44676.0000 - val_fn: 13.0000 - val_accuracy: 0.9820 - val_precision: 0.0830 - val_recall: 0.8488 - val_auc: 0.9584 - val_prc: 0.7122
Epoch 24/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.2076 - tp: 18459.0000 - fp: 896.0000 - tn: 19582.0000 - fn: 2023.0000 - accuracy: 0.9287 - precision: 0.9537 - recall: 0.9012 - auc: 0.9694 - prc: 0.9776 - val_loss: 0.1528 - val_tp: 73.0000 - val_fp: 807.0000 - val_tn: 44676.0000 - val_fn: 13.0000 - val_accuracy: 0.9820 - val_precision: 0.0830 - val_recall: 0.8488 - val_auc: 0.9587 - val_prc: 0.7129
Epoch 25/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.2044 - tp: 18340.0000 - fp: 907.0000 - tn: 19664.0000 - fn: 2049.0000 - accuracy: 0.9278 - precision: 0.9529 - recall: 0.8995 - auc: 0.9707 - prc: 0.9783 - val_loss: 0.1483 - val_tp: 73.0000 - val_fp: 800.0000 - val_tn: 44683.0000 - val_fn: 13.0000 - val_accuracy: 0.9822 - val_precision: 0.0836 - val_recall: 0.8488 - val_auc: 0.9591 - val_prc: 0.7054
Epoch 26/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.1997 - tp: 18293.0000 - fp: 918.0000 - tn: 19749.0000 - fn: 2000.0000 - accuracy: 0.9288 - precision: 0.9522 - recall: 0.9014 - auc: 0.9722 - prc: 0.9788 - val_loss: 0.1433 - val_tp: 73.0000 - val_fp: 788.0000 - val_tn: 44695.0000 - val_fn: 13.0000 - val_accuracy: 0.9824 - val_precision: 0.0848 - val_recall: 0.8488 - val_auc: 0.9590 - val_prc: 0.7059
Epoch 27/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.1987 - tp: 18562.0000 - fp: 848.0000 - tn: 19530.0000 - fn: 2020.0000 - accuracy: 0.9300 - precision: 0.9563 - recall: 0.9019 - auc: 0.9720 - prc: 0.9791 - val_loss: 0.1394 - val_tp: 73.0000 - val_fp: 784.0000 - val_tn: 44699.0000 - val_fn: 13.0000 - val_accuracy: 0.9825 - val_precision: 0.0852 - val_recall: 0.8488 - val_auc: 0.9595 - val_prc: 0.7062
Epoch 28/1000
20/20 [==============================] - 1s 34ms/step - loss: 0.1944 - tp: 18320.0000 - fp: 828.0000 - tn: 19823.0000 - fn: 1989.0000 - accuracy: 0.9312 - precision: 0.9568 - recall: 0.9021 - auc: 0.9734 - prc: 0.9798 - val_loss: 0.1351 - val_tp: 73.0000 - val_fp: 766.0000 - val_tn: 44717.0000 - val_fn: 13.0000 - val_accuracy: 0.9829 - val_precision: 0.0870 - val_recall: 0.8488 - val_auc: 0.9598 - val_prc: 0.7079
Epoch 29/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.1933 - tp: 18455.0000 - fp: 827.0000 - tn: 19704.0000 - fn: 1974.0000 - accuracy: 0.9316 - precision: 0.9571 - recall: 0.9034 - auc: 0.9732 - prc: 0.9797 - val_loss: 0.1313 - val_tp: 73.0000 - val_fp: 766.0000 - val_tn: 44717.0000 - val_fn: 13.0000 - val_accuracy: 0.9829 - val_precision: 0.0870 - val_recall: 0.8488 - val_auc: 0.9599 - val_prc: 0.7094
Epoch 30/1000
20/20 [==============================] - 1s 35ms/step - loss: 0.1910 - tp: 18417.0000 - fp: 768.0000 - tn: 19858.0000 - fn: 1917.0000 - accuracy: 0.9344 - precision: 0.9600 - recall: 0.9057 - auc: 0.9740 - prc: 0.9802 - val_loss: 0.1282 - val_tp: 73.0000 - val_fp: 759.0000 - val_tn: 44724.0000 - val_fn: 13.0000 - val_accuracy: 0.9831 - val_precision: 0.0877 - val_recall: 0.8488 - val_auc: 0.9602 - val_prc: 0.7094
Epoch 31/1000
20/20 [==============================] - ETA: 0s - loss: 0.1866 - tp: 18494.0000 - fp: 756.0000 - tn: 19815.0000 - fn: 1895.0000 - accuracy: 0.9353 - precision: 0.9607 - recall: 0.9071 - auc: 0.9753 - prc: 0.9811Restoring model weights from the end of the best epoch: 21.
20/20 [==============================] - 1s 34ms/step - loss: 0.1866 - tp: 18494.0000 - fp: 756.0000 - tn: 19815.0000 - fn: 1895.0000 - accuracy: 0.9353 - precision: 0.9607 - recall: 0.9071 - auc: 0.9753 - prc: 0.9811 - val_loss: 0.1246 - val_tp: 73.0000 - val_fp: 742.0000 - val_tn: 44741.0000 - val_fn: 13.0000 - val_accuracy: 0.9834 - val_precision: 0.0896 - val_recall: 0.8488 - val_auc: 0.9597 - val_prc: 0.7095
Epoch 31: early stopping

Vuelva a comprobar el historial de entrenamiento

plot_metrics(resampled_history)

png

Evaluar métricas

train_predictions_resampled = resampled_model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_resampled = resampled_model.predict(test_features, batch_size=BATCH_SIZE)
resampled_results = resampled_model.evaluate(test_features, test_labels,
                                             batch_size=BATCH_SIZE, verbose=0)
for name, value in zip(resampled_model.metrics_names, resampled_results):
  print(name, ': ', value)
print()

plot_cm(test_labels, test_predictions_resampled)
loss :  0.16882120072841644
tp :  71.0
fp :  1032.0
tn :  55849.0
fn :  10.0
accuracy :  0.9817070960998535
precision :  0.06436990201473236
recall :  0.8765432238578796
auc :  0.9518552422523499
prc :  0.7423797845840454

Legitimate Transactions Detected (True Negatives):  55849
Legitimate Transactions Incorrectly Detected (False Positives):  1032
Fraudulent Transactions Missed (False Negatives):  10
Fraudulent Transactions Detected (True Positives):  71
Total Fraudulent Transactions:  81

png

Trazar la República de China

plot_roc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_roc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')

plot_roc("Train Weighted", train_labels, train_predictions_weighted, color=colors[1])
plot_roc("Test Weighted", test_labels, test_predictions_weighted, color=colors[1], linestyle='--')

plot_roc("Train Resampled", train_labels, train_predictions_resampled, color=colors[2])
plot_roc("Test Resampled", test_labels, test_predictions_resampled, color=colors[2], linestyle='--')
plt.legend(loc='lower right');

png

Trazar el AUPRC

plot_prc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_prc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')

plot_prc("Train Weighted", train_labels, train_predictions_weighted, color=colors[1])
plot_prc("Test Weighted", test_labels, test_predictions_weighted, color=colors[1], linestyle='--')

plot_prc("Train Resampled", train_labels, train_predictions_resampled, color=colors[2])
plot_prc("Test Resampled", test_labels, test_predictions_resampled, color=colors[2], linestyle='--')
plt.legend(loc='lower right');

png

Aplicando este tutorial a tu problema

La clasificación de datos desequilibrados es una tarea intrínsecamente difícil, ya que hay muy pocas muestras de las que aprender. Siempre debe comenzar con los datos primero y hacer todo lo posible para recopilar tantas muestras como sea posible y reflexionar sobre qué características pueden ser relevantes para que el modelo pueda aprovechar al máximo su clase minoritaria. En algún momento, su modelo puede tener dificultades para mejorar y generar los resultados que desea, por lo que es importante tener en cuenta el contexto de su problema y las compensaciones entre los diferentes tipos de errores.