Classification on imbalanced data

This tutorial demonstrates how to classify a highly imbalanced dataset in which the number of examples in one class greatly outnumbers the examples in another. You will work with the Credit Card Fraud Detection dataset hosted on Kaggle. The aim is to detect a mere 492 fraudulent transactions from 284,807 transactions in total. You will use Keras to define the model and class weights to help the model learn from the imbalanced data. .

This tutorial contains complete code to:

• Load a CSV file using Pandas.
• Create train, validation, and test sets.
• Define and train a model using Keras (including setting class weights).
• Evaluate the model using various metrics (including precision and recall).
• Try common techniques for dealing with imbalanced data like:
  • Class weighting
  • Oversampling

Setup

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']

Data processing and exploration

Download the Kaggle Credit Card Fraud data set

Pandas is a Python library with many helpful utilities for loading and working with structured data. It can be used to download CSVs into a Pandas DataFrame.

Note: This dataset has been collected and analysed during a research collaboration of Worldline and the Machine Learning Group of ULB (Université Libre de Bruxelles) on big data mining and fraud detection. More details on current and past projects on related topics are available here and the page of the DefeatFraud project

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()

Examine the class label imbalance

Let's look at the dataset imbalance:

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)

This shows the small fraction of positive samples.

Clean, split and normalize the data

The raw data has a few issues. First the Time and Amount columns are too variable to use directly. Drop the Time column (since it's not clear what it means) and take the log of the Amount column to reduce its range.

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 Amount'] = np.log(cleaned_df.pop('Amount')+eps)

Split the dataset into train, validation, and test sets. The validation set is used during the model fitting to evaluate the loss and any metrics, however the model is not fit with this data. The test set is completely unused during the training phase and is only used at the end to evaluate how well the model generalizes to new data. This is especially important with imbalanced datasets where overfitting is a significant concern from the lack of training data.

# 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)

Normalize the input features using the sklearn StandardScaler. This will set the mean to 0 and standard deviation to 1.

Note: The StandardScaler is only fit using the train_features to be sure the model is not peeking at the validation or test sets.

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)

Caution: If you want to deploy a model, it's critical that you preserve the preprocessing calculations. The easiest way to implement them as layers, and attach them to your model before export.

Look at the data distribution

Next compare the distributions of the positive and negative examples over a few features. Good questions to ask yourself at this point are:

• Do these distributions make sense?
  • Yes. You've normalized the input and these are mostly concentrated in the +/- 2 range.
• Can you see the difference between the distributions?
  • Yes the positive examples contain a much higher rate of extreme values.

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")

Define the model and metrics

Define a function that creates a simple neural network with a densly connected hidden layer, a dropout layer to reduce overfitting, and an output sigmoid layer that returns the probability of a transaction being fraudulent:

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

Understanding useful metrics

Notice that there are a few metrics defined above that can be computed by the model that will be helpful when evaluating the performance.

• False negatives and false positives are samples that were incorrectly classified
• True negatives and true positives are samples that were correctly classified
• Accuracy is the percentage of examples correctly classified > \(\frac{\text{true samples} }{\text{total samples} }\)
• Precision is the percentage of predicted positives that were correctly classified > \(\frac{\text{true positives} }{\text{true positives + false positives} }\)
• Recall is the percentage of actual positives that were correctly classified > \(\frac{\text{true positives} }{\text{true positives + false negatives} }\)
• AUC refers to the Area Under the Curve of a Receiver Operating Characteristic curve (ROC-AUC). This metric is equal to the probability that a classifier will rank a random positive sample higher than a random negative sample.
• AUPRC refers to Area Under the Curve of the Precision-Recall Curve. This metric computes precision-recall pairs for different probability thresholds.

Note: Accuracy is not a helpful metric for this task. You can have 99.8%+ accuracy on this task by predicting False all the time.

Read more:

• True vs. False and Positive vs. Negative
• Accuracy
• Precision and Recall
• ROC-AUC
• Relationship between Precision-Recall and ROC Curves

Baseline model

Build the model

Now create and train your model using the function that was defined earlier. Notice that the model is fit using a larger than default batch size of 2048, this is important to ensure that each batch has a decent chance of containing a few positive samples. If the batch size was too small, they would likely have no fraudulent transactions to learn from.

Note: this model will not handle the class imbalance well. You will improve it later in this tutorial.

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
_________________________________________________________________

Test run the model:

model.predict(train_features[:10])

1/1 [==============================] - 0s 344ms/step
array([[0.6088169 ],
       [0.4941456 ],
       [0.6518074 ],
       [0.6477403 ],
       [0.3832439 ],
       [0.4007728 ],
       [0.32497266],
       [0.3713796 ],
       [0.42784396],
       [0.2801294 ]], dtype=float32)

Optional: Set the correct initial bias.

These initial guesses are not great. You know the dataset is imbalanced. Set the output layer's bias to reflect that (See: A Recipe for Training Neural Networks: "init well"). This can help with initial convergence.

With the default bias initialization the loss should be about 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: 0.6775

The correct bias to set can be derived from:

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

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

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

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

array([-6.35935934])

Set that as the initial bias, and the model will give much more reasonable initial guesses.

It should be near: pos/total = 0.0018

model = make_model(output_bias=initial_bias)
model.predict(train_features[:10])

1/1 [==============================] - 0s 40ms/step
array([[0.00102003],
       [0.00334534],
       [0.00060373],
       [0.00167802],
       [0.0011223 ],
       [0.0013187 ],
       [0.00594859],
       [0.00176013],
       [0.0008566 ],
       [0.5312835 ]], dtype=float32)

With this initialization the initial loss should be approximately:

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

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

Loss: 0.0209

This initial loss is about 50 times less than if would have been with naive initialization.

This way the model doesn't need to spend the first few epochs just learning that positive examples are unlikely. This also makes it easier to read plots of the loss during training.

Checkpoint the initial weights

To make the various training runs more comparable, keep this initial model's weights in a checkpoint file, and load them into each model before training:

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

Confirm that the bias fix helps

Before moving on, confirm quick that the careful bias initialization actually helped.

Train the model for 20 epochs, with and without this careful initialization, and compare the losses:

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)

The above figure makes it clear: In terms of validation loss, on this problem, this careful initialization gives a clear advantage.

Train the model

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 14ms/step - loss: 0.0179 - tp: 62.0000 - fp: 150.0000 - tn: 227299.0000 - fn: 334.0000 - accuracy: 0.9979 - precision: 0.2925 - recall: 0.1566 - auc: 0.6586 - prc: 0.1123 - val_loss: 0.0092 - val_tp: 0.0000e+00 - val_fp: 0.0000e+00 - val_tn: 45496.0000 - val_fn: 73.0000 - val_accuracy: 0.9984 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00 - val_auc: 0.7654 - val_prc: 0.3283
Epoch 2/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0095 - tp: 77.0000 - fp: 51.0000 - tn: 181902.0000 - fn: 246.0000 - accuracy: 0.9984 - precision: 0.6016 - recall: 0.2384 - auc: 0.8065 - prc: 0.3097 - val_loss: 0.0056 - val_tp: 26.0000 - val_fp: 8.0000 - val_tn: 45488.0000 - val_fn: 47.0000 - val_accuracy: 0.9988 - val_precision: 0.7647 - val_recall: 0.3562 - val_auc: 0.9107 - val_prc: 0.6655
Epoch 3/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0070 - tp: 136.0000 - fp: 33.0000 - tn: 181920.0000 - fn: 187.0000 - accuracy: 0.9988 - precision: 0.8047 - recall: 0.4211 - auc: 0.8727 - prc: 0.5329 - val_loss: 0.0047 - val_tp: 37.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 36.0000 - val_accuracy: 0.9990 - val_precision: 0.8043 - val_recall: 0.5068 - val_auc: 0.9244 - val_prc: 0.6576
Epoch 4/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0061 - tp: 157.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 166.0000 - accuracy: 0.9990 - precision: 0.8626 - recall: 0.4861 - auc: 0.8871 - prc: 0.6036 - val_loss: 0.0043 - val_tp: 42.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 31.0000 - val_accuracy: 0.9991 - val_precision: 0.8077 - val_recall: 0.5753 - val_auc: 0.9244 - val_prc: 0.6553
Epoch 5/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0058 - tp: 172.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 151.0000 - accuracy: 0.9990 - precision: 0.8731 - recall: 0.5325 - auc: 0.8942 - prc: 0.6344 - val_loss: 0.0041 - val_tp: 45.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 28.0000 - val_accuracy: 0.9992 - val_precision: 0.8182 - val_recall: 0.6164 - val_auc: 0.9312 - val_prc: 0.6639
Epoch 6/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0055 - tp: 172.0000 - fp: 22.0000 - tn: 181931.0000 - fn: 151.0000 - accuracy: 0.9991 - precision: 0.8866 - recall: 0.5325 - auc: 0.8976 - prc: 0.6494 - val_loss: 0.0039 - val_tp: 52.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8387 - val_recall: 0.7123 - val_auc: 0.9380 - val_prc: 0.6824
Epoch 7/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0057 - tp: 180.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 143.0000 - accuracy: 0.9991 - precision: 0.8654 - recall: 0.5573 - auc: 0.8946 - prc: 0.6345 - val_loss: 0.0037 - val_tp: 52.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8387 - val_recall: 0.7123 - val_auc: 0.9448 - val_prc: 0.6873
Epoch 8/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0053 - tp: 181.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 142.0000 - accuracy: 0.9991 - precision: 0.8744 - recall: 0.5604 - auc: 0.9027 - prc: 0.6565 - val_loss: 0.0036 - val_tp: 52.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 21.0000 - val_accuracy: 0.9993 - val_precision: 0.8525 - val_recall: 0.7123 - val_auc: 0.9448 - val_prc: 0.6884
Epoch 9/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0053 - tp: 180.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 143.0000 - accuracy: 0.9990 - precision: 0.8531 - recall: 0.5573 - auc: 0.8980 - prc: 0.6595 - val_loss: 0.0035 - val_tp: 53.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 20.0000 - val_accuracy: 0.9993 - val_precision: 0.8413 - val_recall: 0.7260 - val_auc: 0.9449 - val_prc: 0.7112
Epoch 10/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0049 - tp: 189.0000 - fp: 32.0000 - tn: 181921.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8552 - recall: 0.5851 - auc: 0.9123 - prc: 0.6817 - val_loss: 0.0034 - val_tp: 53.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 20.0000 - val_accuracy: 0.9994 - val_precision: 0.8548 - val_recall: 0.7260 - val_auc: 0.9448 - val_prc: 0.7094
Epoch 11/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0048 - tp: 189.0000 - fp: 32.0000 - tn: 181921.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8552 - recall: 0.5851 - auc: 0.9215 - prc: 0.6940 - val_loss: 0.0034 - val_tp: 54.0000 - val_fp: 10.0000 - val_tn: 45486.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8438 - val_recall: 0.7397 - val_auc: 0.9448 - val_prc: 0.7151
Epoch 12/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0050 - tp: 183.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 140.0000 - accuracy: 0.9991 - precision: 0.8551 - recall: 0.5666 - auc: 0.9091 - prc: 0.6663 - val_loss: 0.0032 - val_tp: 52.0000 - val_fp: 8.0000 - val_tn: 45488.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.8667 - val_recall: 0.7123 - val_auc: 0.9448 - val_prc: 0.7642
Epoch 13/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0047 - tp: 187.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 136.0000 - accuracy: 0.9991 - precision: 0.8578 - recall: 0.5789 - auc: 0.9278 - prc: 0.6926 - val_loss: 0.0032 - val_tp: 54.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8571 - val_recall: 0.7397 - val_auc: 0.9449 - val_prc: 0.7562
Epoch 14/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 198.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 125.0000 - accuracy: 0.9991 - precision: 0.8684 - recall: 0.6130 - auc: 0.9170 - prc: 0.7135 - val_loss: 0.0031 - val_tp: 52.0000 - val_fp: 8.0000 - val_tn: 45488.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.8667 - val_recall: 0.7123 - val_auc: 0.9449 - val_prc: 0.7759
Epoch 15/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 192.0000 - fp: 27.0000 - tn: 181926.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8767 - recall: 0.5944 - auc: 0.9216 - prc: 0.7232 - val_loss: 0.0031 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9448 - val_prc: 0.7478
Epoch 16/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0046 - tp: 192.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8727 - recall: 0.5944 - auc: 0.9200 - prc: 0.6946 - val_loss: 0.0030 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9448 - val_prc: 0.7806
Epoch 17/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 196.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 127.0000 - accuracy: 0.9992 - precision: 0.8869 - recall: 0.6068 - auc: 0.9231 - prc: 0.7186 - val_loss: 0.0030 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9449 - val_prc: 0.7846
Epoch 18/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0046 - tp: 189.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8832 - recall: 0.5851 - auc: 0.9107 - prc: 0.6852 - val_loss: 0.0030 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9448 - val_prc: 0.7923
Epoch 19/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 195.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 128.0000 - accuracy: 0.9991 - precision: 0.8628 - recall: 0.6037 - auc: 0.9247 - prc: 0.6924 - val_loss: 0.0029 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9448 - val_prc: 0.7919
Epoch 20/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 202.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8745 - recall: 0.6254 - auc: 0.9262 - prc: 0.7067 - val_loss: 0.0028 - val_tp: 56.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 17.0000 - val_accuracy: 0.9994 - val_precision: 0.8615 - val_recall: 0.7671 - val_auc: 0.9449 - val_prc: 0.7965
Epoch 21/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 184.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 139.0000 - accuracy: 0.9991 - precision: 0.8638 - recall: 0.5697 - auc: 0.9294 - prc: 0.6954 - val_loss: 0.0029 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.7957
Epoch 22/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 205.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8798 - recall: 0.6347 - auc: 0.9170 - prc: 0.7232 - val_loss: 0.0029 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.7826
Epoch 23/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 201.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 122.0000 - accuracy: 0.9992 - precision: 0.8664 - recall: 0.6223 - auc: 0.9185 - prc: 0.6926 - val_loss: 0.0028 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.7982
Epoch 24/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 196.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 127.0000 - accuracy: 0.9991 - precision: 0.8711 - recall: 0.6068 - auc: 0.9217 - prc: 0.7141 - val_loss: 0.0028 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.7979
Epoch 25/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 199.0000 - fp: 32.0000 - tn: 181921.0000 - fn: 124.0000 - accuracy: 0.9991 - precision: 0.8615 - recall: 0.6161 - auc: 0.9278 - prc: 0.7046 - val_loss: 0.0027 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8015
Epoch 26/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 191.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 132.0000 - accuracy: 0.9991 - precision: 0.8643 - recall: 0.5913 - auc: 0.9232 - prc: 0.6977 - val_loss: 0.0027 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8049
Epoch 27/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 196.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 127.0000 - accuracy: 0.9991 - precision: 0.8634 - recall: 0.6068 - auc: 0.9170 - prc: 0.6967 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8081
Epoch 28/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 198.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 125.0000 - accuracy: 0.9991 - precision: 0.8684 - recall: 0.6130 - auc: 0.9247 - prc: 0.7149 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9448 - val_prc: 0.8068
Epoch 29/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 194.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 129.0000 - accuracy: 0.9991 - precision: 0.8661 - recall: 0.6006 - auc: 0.9123 - prc: 0.7028 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8058
Epoch 30/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 195.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 128.0000 - accuracy: 0.9991 - precision: 0.8744 - recall: 0.6037 - auc: 0.9217 - prc: 0.7177 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8049
Epoch 31/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 193.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 130.0000 - accuracy: 0.9991 - precision: 0.8694 - recall: 0.5975 - auc: 0.9200 - prc: 0.7029 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8092
Epoch 32/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 195.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 128.0000 - accuracy: 0.9992 - precision: 0.8864 - recall: 0.6037 - auc: 0.9310 - prc: 0.7208 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9448 - val_prc: 0.8128
Epoch 33/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 199.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 124.0000 - accuracy: 0.9992 - precision: 0.8884 - recall: 0.6161 - auc: 0.9217 - prc: 0.7202 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8134
Epoch 34/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 210.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 113.0000 - accuracy: 0.9992 - precision: 0.8824 - recall: 0.6502 - auc: 0.9341 - prc: 0.7453 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9448 - val_prc: 0.8131
Epoch 35/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 200.0000 - fp: 34.0000 - tn: 181919.0000 - fn: 123.0000 - accuracy: 0.9991 - precision: 0.8547 - recall: 0.6192 - auc: 0.9310 - prc: 0.7165 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8199
Epoch 36/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 184.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 139.0000 - accuracy: 0.9991 - precision: 0.8598 - recall: 0.5697 - auc: 0.9278 - prc: 0.7054 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8162
Epoch 37/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 208.0000 - fp: 27.0000 - tn: 181926.0000 - fn: 115.0000 - accuracy: 0.9992 - precision: 0.8851 - recall: 0.6440 - auc: 0.9264 - prc: 0.7460 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8162
Epoch 38/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 201.0000 - fp: 33.0000 - tn: 181920.0000 - fn: 122.0000 - accuracy: 0.9991 - precision: 0.8590 - recall: 0.6223 - auc: 0.9263 - prc: 0.7016 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8199
Epoch 39/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 204.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8793 - recall: 0.6316 - auc: 0.9311 - prc: 0.7300 - val_loss: 0.0026 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8207
Epoch 40/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 202.0000 - fp: 24.0000 - tn: 181929.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8938 - recall: 0.6254 - auc: 0.9296 - prc: 0.7514 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8199
Epoch 41/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 202.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8783 - recall: 0.6254 - auc: 0.9372 - prc: 0.7448 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8200
Epoch 42/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 197.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 126.0000 - accuracy: 0.9991 - precision: 0.8678 - recall: 0.6099 - auc: 0.9325 - prc: 0.7317 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8260
Epoch 43/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 204.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8870 - recall: 0.6316 - auc: 0.9280 - prc: 0.7542 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8231
Epoch 44/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 189.0000 - fp: 36.0000 - tn: 181917.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8400 - recall: 0.5851 - auc: 0.9231 - prc: 0.6901 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8268
Epoch 45/100
90/90 [==============================] - 0s 6ms/step - loss: 0.0040 - tp: 204.0000 - fp: 22.0000 - tn: 181931.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.9027 - recall: 0.6316 - auc: 0.9325 - prc: 0.7257 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8315
Epoch 46/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 198.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 125.0000 - accuracy: 0.9991 - precision: 0.8646 - recall: 0.6130 - auc: 0.9357 - prc: 0.7480 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8322
Epoch 47/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 205.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8723 - recall: 0.6347 - auc: 0.9263 - prc: 0.7404 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8269
Epoch 48/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0036 - tp: 217.0000 - fp: 21.0000 - tn: 181932.0000 - fn: 106.0000 - accuracy: 0.9993 - precision: 0.9118 - recall: 0.6718 - auc: 0.9357 - prc: 0.7610 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8306
Epoch 49/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0036 - tp: 211.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 112.0000 - accuracy: 0.9992 - precision: 0.8941 - recall: 0.6533 - auc: 0.9356 - prc: 0.7632 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8247
Epoch 50/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 205.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8761 - recall: 0.6347 - auc: 0.9340 - prc: 0.7286 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8288
Epoch 51/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 203.0000 - fp: 27.0000 - tn: 181926.0000 - fn: 120.0000 - accuracy: 0.9992 - precision: 0.8826 - recall: 0.6285 - auc: 0.9279 - prc: 0.7410 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8276
Epoch 52/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 201.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 122.0000 - accuracy: 0.9992 - precision: 0.8777 - recall: 0.6223 - auc: 0.9342 - prc: 0.7278 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8306
Epoch 53/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 201.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 122.0000 - accuracy: 0.9992 - precision: 0.8739 - recall: 0.6223 - auc: 0.9325 - prc: 0.7294 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8342
Epoch 54/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 201.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 122.0000 - accuracy: 0.9992 - precision: 0.8739 - recall: 0.6223 - auc: 0.9372 - prc: 0.7202 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8329
Epoch 55/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 196.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 127.0000 - accuracy: 0.9991 - precision: 0.8711 - recall: 0.6068 - auc: 0.9372 - prc: 0.7500 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8337
Epoch 56/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 205.0000 - fp: 35.0000 - tn: 181918.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8542 - recall: 0.6347 - auc: 0.9231 - prc: 0.7172 - val_loss: 0.0024 - val_tp: 55.0000 - val_fp: 7.0000 - val_tn: 45489.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.8871 - val_recall: 0.7534 - val_auc: 0.9449 - val_prc: 0.8335
Epoch 57/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 195.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 128.0000 - accuracy: 0.9991 - precision: 0.8705 - recall: 0.6037 - auc: 0.9341 - prc: 0.7320 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8340
Epoch 58/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 197.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 126.0000 - accuracy: 0.9992 - precision: 0.8756 - recall: 0.6099 - auc: 0.9294 - prc: 0.7280 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8317
Epoch 59/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 206.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 117.0000 - accuracy: 0.9992 - precision: 0.8879 - recall: 0.6378 - auc: 0.9340 - prc: 0.7334 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8415
Epoch 60/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 194.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 129.0000 - accuracy: 0.9991 - precision: 0.8661 - recall: 0.6006 - auc: 0.9325 - prc: 0.7462 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8330
Epoch 61/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 204.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8681 - recall: 0.6316 - auc: 0.9341 - prc: 0.7380 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8344
Epoch 62/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 204.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8908 - recall: 0.6316 - auc: 0.9403 - prc: 0.7393 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8384
Epoch 63/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 204.0000 - fp: 27.0000 - tn: 181926.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8831 - recall: 0.6316 - auc: 0.9372 - prc: 0.7546 - val_loss: 0.0025 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8325
Epoch 64/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 208.0000 - fp: 37.0000 - tn: 181916.0000 - fn: 115.0000 - accuracy: 0.9992 - precision: 0.8490 - recall: 0.6440 - auc: 0.9248 - prc: 0.7373 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8350
Epoch 65/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 199.0000 - fp: 24.0000 - tn: 181929.0000 - fn: 124.0000 - accuracy: 0.9992 - precision: 0.8924 - recall: 0.6161 - auc: 0.9388 - prc: 0.7484 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9449 - val_prc: 0.8350
Epoch 66/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 214.0000 - fp: 23.0000 - tn: 181930.0000 - fn: 109.0000 - accuracy: 0.9993 - precision: 0.9030 - recall: 0.6625 - auc: 0.9403 - prc: 0.7531 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8392
Epoch 67/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 205.0000 - fp: 33.0000 - tn: 181920.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8613 - recall: 0.6347 - auc: 0.9372 - prc: 0.7505 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8314
Epoch 68/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 213.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 110.0000 - accuracy: 0.9992 - precision: 0.8838 - recall: 0.6594 - auc: 0.9465 - prc: 0.7519 - val_loss: 0.0024 - val_tp: 55.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9016 - val_recall: 0.7534 - val_auc: 0.9517 - val_prc: 0.8420
Epoch 69/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 198.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 125.0000 - accuracy: 0.9991 - precision: 0.8684 - recall: 0.6130 - auc: 0.9386 - prc: 0.7377 - val_loss: 0.0023 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8423
Epoch 70/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0036 - tp: 205.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8761 - recall: 0.6347 - auc: 0.9449 - prc: 0.7606 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9517 - val_prc: 0.8433
Epoch 71/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 196.0000 - fp: 24.0000 - tn: 181929.0000 - fn: 127.0000 - accuracy: 0.9992 - precision: 0.8909 - recall: 0.6068 - auc: 0.9340 - prc: 0.7519 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9518 - val_prc: 0.8424
Epoch 72/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0035 - tp: 219.0000 - fp: 28.0000 - tn: 181925.0000 - fn: 104.0000 - accuracy: 0.9993 - precision: 0.8866 - recall: 0.6780 - auc: 0.9373 - prc: 0.7570 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9517 - val_prc: 0.8435
Epoch 73/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 212.0000 - fp: 31.0000 - tn: 181922.0000 - fn: 111.0000 - accuracy: 0.9992 - precision: 0.8724 - recall: 0.6563 - auc: 0.9418 - prc: 0.7575 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9517 - val_prc: 0.8434
Epoch 74/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 201.0000 - fp: 24.0000 - tn: 181929.0000 - fn: 122.0000 - accuracy: 0.9992 - precision: 0.8933 - recall: 0.6223 - auc: 0.9402 - prc: 0.7614 - val_loss: 0.0024 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9518 - val_prc: 0.8406
Epoch 75/100
90/90 [==============================] - 0s 6ms/step - loss: 0.0036 - tp: 210.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 113.0000 - accuracy: 0.9992 - precision: 0.8750 - recall: 0.6502 - auc: 0.9356 - prc: 0.7504 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9517 - val_prc: 0.8416
Epoch 76/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0035 - tp: 206.0000 - fp: 30.0000 - tn: 181923.0000 - fn: 117.0000 - accuracy: 0.9992 - precision: 0.8729 - recall: 0.6378 - auc: 0.9573 - prc: 0.7699 - val_loss: 0.0024 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8415
Epoch 77/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0035 - tp: 216.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 107.0000 - accuracy: 0.9993 - precision: 0.8926 - recall: 0.6687 - auc: 0.9388 - prc: 0.7629 - val_loss: 0.0023 - val_tp: 55.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9016 - val_recall: 0.7534 - val_auc: 0.9517 - val_prc: 0.8398
Epoch 78/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0038 - tp: 202.0000 - fp: 29.0000 - tn: 181924.0000 - fn: 121.0000 - accuracy: 0.9992 - precision: 0.8745 - recall: 0.6254 - auc: 0.9403 - prc: 0.7551 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9518 - val_prc: 0.8426
Epoch 79/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0034 - tp: 213.0000 - fp: 25.0000 - tn: 181928.0000 - fn: 110.0000 - accuracy: 0.9993 - precision: 0.8950 - recall: 0.6594 - auc: 0.9450 - prc: 0.7702 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9517 - val_prc: 0.8427
Epoch 80/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 204.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 119.0000 - accuracy: 0.9992 - precision: 0.8870 - recall: 0.6316 - auc: 0.9417 - prc: 0.7328 - val_loss: 0.0023 - val_tp: 57.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.9048 - val_recall: 0.7808 - val_auc: 0.9518 - val_prc: 0.8401
Epoch 81/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0035 - tp: 205.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8874 - recall: 0.6347 - auc: 0.9402 - prc: 0.7628 - val_loss: 0.0023 - val_tp: 56.0000 - val_fp: 6.0000 - val_tn: 45490.0000 - val_fn: 17.0000 - val_accuracy: 0.9995 - val_precision: 0.9032 - val_recall: 0.7671 - val_auc: 0.9517 - val_prc: 0.8406
Epoch 82/100
82/90 [==========================>...] - ETA: 0s - loss: 0.0037 - tp: 186.0000 - fp: 21.0000 - tn: 167618.0000 - fn: 111.0000 - accuracy: 0.9992 - precision: 0.8986 - recall: 0.6263 - auc: 0.9367 - prc: 0.7555Restoring model weights from the end of the best epoch: 72.
90/90 [==============================] - 1s 6ms/step - loss: 0.0037 - tp: 205.0000 - fp: 26.0000 - tn: 181927.0000 - fn: 118.0000 - accuracy: 0.9992 - precision: 0.8874 - recall: 0.6347 - auc: 0.9386 - prc: 0.7444 - val_loss: 0.0023 - val_tp: 58.0000 - val_fp: 9.0000 - val_tn: 45487.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.8657 - val_recall: 0.7945 - val_auc: 0.9517 - val_prc: 0.8395
Epoch 82: early stopping

Check training history

In this section, you will produce plots of your model's accuracy and loss on the training and validation set. These are useful to check for overfitting, which you can learn more about in the Overfit and underfit tutorial.

Additionally, you can produce these plots for any of the metrics you created above. False negatives are included as an example.

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)

Note: That the validation curve generally performs better than the training curve. This is mainly caused by the fact that the dropout layer is not active when evaluating the model.

Evaluate metrics

You can use a confusion matrix to summarize the actual vs. predicted labels, where the X axis is the predicted label and the Y axis is the actual label:

train_predictions_baseline = model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_baseline = model.predict(test_features, batch_size=BATCH_SIZE)

90/90 [==============================] - 0s 1ms/step
28/28 [==============================] - 0s 1ms/step

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]))

Evaluate your model on the test dataset and display the results for the metrics you created above:

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 :  0.002767325611785054
tp :  78.0
fp :  9.0
tn :  56857.0
fn :  18.0
accuracy :  0.9995260238647461
precision :  0.8965517282485962
recall :  0.8125
auc :  0.9475902915000916
prc :  0.847183108329773

Legitimate Transactions Detected (True Negatives):  56857
Legitimate Transactions Incorrectly Detected (False Positives):  9
Fraudulent Transactions Missed (False Negatives):  18
Fraudulent Transactions Detected (True Positives):  78
Total Fraudulent Transactions:  96

If the model had predicted everything perfectly, this would be a diagonal matrix where values off the main diagonal, indicating incorrect predictions, would be zero. In this case the matrix shows that you have relatively few false positives, meaning that there were relatively few legitimate transactions that were incorrectly flagged. However, you would likely want to have even fewer false negatives despite the cost of increasing the number of false positives. This trade off may be preferable because false negatives would allow fraudulent transactions to go through, whereas false positives may cause an email to be sent to a customer to ask them to verify their card activity.

Plot the ROC

Now plot the ROC. This plot is useful because it shows, at a glance, the range of performance the model can reach just by tuning the output threshold.

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');

Plot the AUPRC

Now plot the AUPRC. Area under the interpolated precision-recall curve, obtained by plotting (recall, precision) points for different values of the classification threshold. Depending on how it's calculated, PR AUC may be equivalent to the average precision of the model.

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('Precision')
    plt.ylabel('Recall')
    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');

It looks like the precision is relatively high, but the recall and the area under the ROC curve (AUC) aren't as high as you might like. Classifiers often face challenges when trying to maximize both precision and recall, which is especially true when working with imbalanced datasets. It is important to consider the costs of different types of errors in the context of the problem you care about. In this example, a false negative (a fraudulent transaction is missed) may have a financial cost, while a false positive (a transaction is incorrectly flagged as fraudulent) may decrease user happiness.

Class weights

Calculate class weights

The goal is to identify fraudulent transactions, but you don't have very many of those positive samples to work with, so you would want to have the classifier heavily weight the few examples that are available. You can do this by passing Keras weights for each class through a parameter. These will cause the model to "pay more attention" to examples from an under-represented class.

# 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

Train a model with class weights

Now try re-training and evaluating the model with class weights to see how that affects the predictions.

Note: Using class_weights changes the range of the loss. This may affect the stability of the training depending on the optimizer. Optimizers whose step size is dependent on the magnitude of the gradient, like tf.keras.optimizers.SGD, may fail. The optimizer used here, tf.keras.optimizers.Adam, is unaffected by the scaling change. Also note that because of the weighting, the total losses are not comparable between the two models.

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 14ms/step - loss: 3.3363 - tp: 112.0000 - fp: 540.0000 - tn: 238279.0000 - fn: 307.0000 - accuracy: 0.9965 - precision: 0.1718 - recall: 0.2673 - auc: 0.7141 - prc: 0.0899 - val_loss: 0.0121 - val_tp: 23.0000 - val_fp: 71.0000 - val_tn: 45425.0000 - val_fn: 50.0000 - val_accuracy: 0.9973 - val_precision: 0.2447 - val_recall: 0.3151 - val_auc: 0.8876 - val_prc: 0.2189
Epoch 2/100
90/90 [==============================] - 1s 6ms/step - loss: 1.2405 - tp: 141.0000 - fp: 1034.0000 - tn: 180919.0000 - fn: 182.0000 - accuracy: 0.9933 - precision: 0.1200 - recall: 0.4365 - auc: 0.8562 - prc: 0.1437 - val_loss: 0.0157 - val_tp: 48.0000 - val_fp: 124.0000 - val_tn: 45372.0000 - val_fn: 25.0000 - val_accuracy: 0.9967 - val_precision: 0.2791 - val_recall: 0.6575 - val_auc: 0.9392 - val_prc: 0.4586
Epoch 3/100
90/90 [==============================] - 1s 6ms/step - loss: 0.8540 - tp: 198.0000 - fp: 1492.0000 - tn: 180461.0000 - fn: 125.0000 - accuracy: 0.9911 - precision: 0.1172 - recall: 0.6130 - auc: 0.8861 - prc: 0.1848 - val_loss: 0.0216 - val_tp: 59.0000 - val_fp: 184.0000 - val_tn: 45312.0000 - val_fn: 14.0000 - val_accuracy: 0.9957 - val_precision: 0.2428 - val_recall: 0.8082 - val_auc: 0.9582 - val_prc: 0.5453
Epoch 4/100
90/90 [==============================] - 1s 6ms/step - loss: 0.6551 - tp: 225.0000 - fp: 2166.0000 - tn: 179787.0000 - fn: 98.0000 - accuracy: 0.9876 - precision: 0.0941 - recall: 0.6966 - auc: 0.9086 - prc: 0.2311 - val_loss: 0.0281 - val_tp: 62.0000 - val_fp: 235.0000 - val_tn: 45261.0000 - val_fn: 11.0000 - val_accuracy: 0.9946 - val_precision: 0.2088 - val_recall: 0.8493 - val_auc: 0.9627 - val_prc: 0.5504
Epoch 5/100
90/90 [==============================] - 1s 6ms/step - loss: 0.5006 - tp: 247.0000 - fp: 2986.0000 - tn: 178967.0000 - fn: 76.0000 - accuracy: 0.9832 - precision: 0.0764 - recall: 0.7647 - auc: 0.9274 - prc: 0.2182 - val_loss: 0.0367 - val_tp: 64.0000 - val_fp: 356.0000 - val_tn: 45140.0000 - val_fn: 9.0000 - val_accuracy: 0.9920 - val_precision: 0.1524 - val_recall: 0.8767 - val_auc: 0.9676 - val_prc: 0.5536
Epoch 6/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4861 - tp: 250.0000 - fp: 3848.0000 - tn: 178105.0000 - fn: 73.0000 - accuracy: 0.9785 - precision: 0.0610 - recall: 0.7740 - auc: 0.9276 - prc: 0.1976 - val_loss: 0.0454 - val_tp: 64.0000 - val_fp: 507.0000 - val_tn: 44989.0000 - val_fn: 9.0000 - val_accuracy: 0.9887 - val_precision: 0.1121 - val_recall: 0.8767 - val_auc: 0.9736 - val_prc: 0.5522
Epoch 7/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4173 - tp: 260.0000 - fp: 4601.0000 - tn: 177352.0000 - fn: 63.0000 - accuracy: 0.9744 - precision: 0.0535 - recall: 0.8050 - auc: 0.9349 - prc: 0.1939 - val_loss: 0.0533 - val_tp: 65.0000 - val_fp: 622.0000 - val_tn: 44874.0000 - val_fn: 8.0000 - val_accuracy: 0.9862 - val_precision: 0.0946 - val_recall: 0.8904 - val_auc: 0.9757 - val_prc: 0.5307
Epoch 8/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4603 - tp: 258.0000 - fp: 5399.0000 - tn: 176554.0000 - fn: 65.0000 - accuracy: 0.9700 - precision: 0.0456 - recall: 0.7988 - auc: 0.9225 - prc: 0.1723 - val_loss: 0.0617 - val_tp: 65.0000 - val_fp: 728.0000 - val_tn: 44768.0000 - val_fn: 8.0000 - val_accuracy: 0.9838 - val_precision: 0.0820 - val_recall: 0.8904 - val_auc: 0.9811 - val_prc: 0.5223
Epoch 9/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3903 - tp: 260.0000 - fp: 6230.0000 - tn: 175723.0000 - fn: 63.0000 - accuracy: 0.9655 - precision: 0.0401 - recall: 0.8050 - auc: 0.9424 - prc: 0.1609 - val_loss: 0.0720 - val_tp: 65.0000 - val_fp: 852.0000 - val_tn: 44644.0000 - val_fn: 8.0000 - val_accuracy: 0.9811 - val_precision: 0.0709 - val_recall: 0.8904 - val_auc: 0.9822 - val_prc: 0.5203
Epoch 10/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3662 - tp: 270.0000 - fp: 6567.0000 - tn: 175386.0000 - fn: 53.0000 - accuracy: 0.9637 - precision: 0.0395 - recall: 0.8359 - auc: 0.9400 - prc: 0.1673 - val_loss: 0.0744 - val_tp: 65.0000 - val_fp: 870.0000 - val_tn: 44626.0000 - val_fn: 8.0000 - val_accuracy: 0.9807 - val_precision: 0.0695 - val_recall: 0.8904 - val_auc: 0.9828 - val_prc: 0.5165
Epoch 11/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3496 - tp: 274.0000 - fp: 7039.0000 - tn: 174914.0000 - fn: 49.0000 - accuracy: 0.9611 - precision: 0.0375 - recall: 0.8483 - auc: 0.9415 - prc: 0.1715 - val_loss: 0.0760 - val_tp: 65.0000 - val_fp: 881.0000 - val_tn: 44615.0000 - val_fn: 8.0000 - val_accuracy: 0.9805 - val_precision: 0.0687 - val_recall: 0.8904 - val_auc: 0.9830 - val_prc: 0.5239
Epoch 12/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4208 - tp: 264.0000 - fp: 6923.0000 - tn: 175030.0000 - fn: 59.0000 - accuracy: 0.9617 - precision: 0.0367 - recall: 0.8173 - auc: 0.9244 - prc: 0.1632 - val_loss: 0.0788 - val_tp: 65.0000 - val_fp: 905.0000 - val_tn: 44591.0000 - val_fn: 8.0000 - val_accuracy: 0.9800 - val_precision: 0.0670 - val_recall: 0.8904 - val_auc: 0.9837 - val_prc: 0.5425
Epoch 13/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3524 - tp: 274.0000 - fp: 7253.0000 - tn: 174700.0000 - fn: 49.0000 - accuracy: 0.9599 - precision: 0.0364 - recall: 0.8483 - auc: 0.9415 - prc: 0.1615 - val_loss: 0.0842 - val_tp: 65.0000 - val_fp: 956.0000 - val_tn: 44540.0000 - val_fn: 8.0000 - val_accuracy: 0.9788 - val_precision: 0.0637 - val_recall: 0.8904 - val_auc: 0.9837 - val_prc: 0.5044
Epoch 14/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3047 - tp: 278.0000 - fp: 7509.0000 - tn: 174444.0000 - fn: 45.0000 - accuracy: 0.9586 - precision: 0.0357 - recall: 0.8607 - auc: 0.9541 - prc: 0.1673 - val_loss: 0.0879 - val_tp: 65.0000 - val_fp: 990.0000 - val_tn: 44506.0000 - val_fn: 8.0000 - val_accuracy: 0.9781 - val_precision: 0.0616 - val_recall: 0.8904 - val_auc: 0.9840 - val_prc: 0.4995
Epoch 15/100
89/90 [============================>.] - ETA: 0s - loss: 0.2886 - tp: 283.0000 - fp: 7751.0000 - tn: 174198.0000 - fn: 40.0000 - accuracy: 0.9573 - precision: 0.0352 - recall: 0.8762 - auc: 0.9542 - prc: 0.1691Restoring model weights from the end of the best epoch: 5.
90/90 [==============================] - 1s 6ms/step - loss: 0.2886 - tp: 283.0000 - fp: 7752.0000 - tn: 174201.0000 - fn: 40.0000 - accuracy: 0.9573 - precision: 0.0352 - recall: 0.8762 - auc: 0.9542 - prc: 0.1691 - val_loss: 0.0896 - val_tp: 66.0000 - val_fp: 1005.0000 - val_tn: 44491.0000 - val_fn: 7.0000 - val_accuracy: 0.9778 - val_precision: 0.0616 - val_recall: 0.9041 - val_auc: 0.9843 - val_prc: 0.4827
Epoch 15: early stopping

Check training history

plot_metrics(weighted_history)

Evaluate metrics

train_predictions_weighted = weighted_model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_weighted = weighted_model.predict(test_features, batch_size=BATCH_SIZE)

90/90 [==============================] - 0s 1ms/step
28/28 [==============================] - 0s 1ms/step

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 :  0.0376315712928772
tp :  84.0
fp :  454.0
tn :  56412.0
fn :  12.0
accuracy :  0.9918190836906433
precision :  0.15613383054733276
recall :  0.875
auc :  0.9684351086616516
prc :  0.5385129451751709

Legitimate Transactions Detected (True Negatives):  56412
Legitimate Transactions Incorrectly Detected (False Positives):  454
Fraudulent Transactions Missed (False Negatives):  12
Fraudulent Transactions Detected (True Positives):  84
Total Fraudulent Transactions:  96

Here you can see that with class weights the accuracy and precision are lower because there are more false positives, but conversely the recall and AUC are higher because the model also found more true positives. Despite having lower accuracy, this model has higher recall (and identifies more fraudulent transactions). Of course, there is a cost to both types of error (you wouldn't want to bug users by flagging too many legitimate transactions as fraudulent, either). Carefully consider the trade-offs between these different types of errors for your application.

Plot the ROC

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');

Plot the 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');

Oversampling

Oversample the minority class

A related approach would be to resample the dataset by oversampling the minority class.

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]

Using NumPy

You can balance the dataset manually by choosing the right number of random indices from the positive examples:

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

(181953, 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

(363906, 29)

Using tf.data

If you're using tf.data the easiest way to produce balanced examples is to start with a positive and a negative dataset, and merge them. See the tf.data guide for more examples.

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)

Each dataset provides (feature, label) pairs:

for features, label in pos_ds.take(1):
  print("Features:\n", features.numpy())
  print()
  print("Label: ", label.numpy())

Features:
 [-2.25920707 -0.0072081  -3.40638234  4.89298529  3.87783585 -3.25107162
 -2.14285219  0.28594133 -3.52636688 -3.65233852  5.         -5.
 -1.1397116  -5.          0.28999929 -1.26264317  0.74955735  1.1855264
 -0.84786706 -0.13985126  0.33724841 -0.06842706 -0.3580829  -0.65930289
  1.63937591  1.36920315  1.22563042  2.86007288 -1.4516158 ]

Label:  1

Merge the two together using 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())

0.498046875

To use this dataset, you'll need the number of steps per epoch.

The definition of "epoch" in this case is less clear. Say it's the number of batches required to see each negative example once:

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

278.0

Train on the oversampled data

Now try training the model with the resampled data set instead of using class weights to see how these methods compare.

Note: Because the data was balanced by replicating the positive examples, the total dataset size is larger, and each epoch runs for more training steps.

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 [==============================] - 8s 22ms/step - loss: 0.6368 - tp: 214568.0000 - fp: 82297.0000 - tn: 259040.0000 - fn: 70401.0000 - accuracy: 0.7562 - precision: 0.7228 - recall: 0.7530 - auc: 0.8294 - prc: 0.8637 - val_loss: 0.2618 - val_tp: 68.0000 - val_fp: 1550.0000 - val_tn: 43946.0000 - val_fn: 5.0000 - val_accuracy: 0.9659 - val_precision: 0.0420 - val_recall: 0.9315 - val_auc: 0.9874 - val_prc: 0.7153
Epoch 2/100
278/278 [==============================] - 5s 19ms/step - loss: 0.2339 - tp: 253726.0000 - fp: 20423.0000 - tn: 263865.0000 - fn: 31330.0000 - accuracy: 0.9091 - precision: 0.9255 - recall: 0.8901 - auc: 0.9628 - prc: 0.9711 - val_loss: 0.1346 - val_tp: 67.0000 - val_fp: 1014.0000 - val_tn: 44482.0000 - val_fn: 6.0000 - val_accuracy: 0.9776 - val_precision: 0.0620 - val_recall: 0.9178 - val_auc: 0.9945 - val_prc: 0.7481
Epoch 3/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1753 - tp: 259428.0000 - fp: 12400.0000 - tn: 271726.0000 - fn: 25790.0000 - accuracy: 0.9329 - precision: 0.9544 - recall: 0.9096 - auc: 0.9792 - prc: 0.9830 - val_loss: 0.0976 - val_tp: 67.0000 - val_fp: 867.0000 - val_tn: 44629.0000 - val_fn: 6.0000 - val_accuracy: 0.9808 - val_precision: 0.0717 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7559
Epoch 4/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1520 - tp: 261697.0000 - fp: 10110.0000 - tn: 274249.0000 - fn: 23288.0000 - accuracy: 0.9413 - precision: 0.9628 - recall: 0.9183 - auc: 0.9849 - prc: 0.9871 - val_loss: 0.0841 - val_tp: 67.0000 - val_fp: 876.0000 - val_tn: 44620.0000 - val_fn: 6.0000 - val_accuracy: 0.9806 - val_precision: 0.0710 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7494
Epoch 5/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1376 - tp: 262584.0000 - fp: 9120.0000 - tn: 275971.0000 - fn: 21669.0000 - accuracy: 0.9459 - precision: 0.9664 - recall: 0.9238 - auc: 0.9882 - prc: 0.9894 - val_loss: 0.0762 - val_tp: 67.0000 - val_fp: 853.0000 - val_tn: 44643.0000 - val_fn: 6.0000 - val_accuracy: 0.9811 - val_precision: 0.0728 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7422
Epoch 6/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1267 - tp: 264133.0000 - fp: 8618.0000 - tn: 276116.0000 - fn: 20477.0000 - accuracy: 0.9489 - precision: 0.9684 - recall: 0.9281 - auc: 0.9904 - prc: 0.9911 - val_loss: 0.0696 - val_tp: 67.0000 - val_fp: 822.0000 - val_tn: 44674.0000 - val_fn: 6.0000 - val_accuracy: 0.9818 - val_precision: 0.0754 - val_recall: 0.9178 - val_auc: 0.9946 - val_prc: 0.7418
Epoch 7/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1189 - tp: 265211.0000 - fp: 8283.0000 - tn: 276289.0000 - fn: 19561.0000 - accuracy: 0.9511 - precision: 0.9697 - recall: 0.9313 - auc: 0.9918 - prc: 0.9922 - val_loss: 0.0640 - val_tp: 67.0000 - val_fp: 771.0000 - val_tn: 44725.0000 - val_fn: 6.0000 - val_accuracy: 0.9829 - val_precision: 0.0800 - val_recall: 0.9178 - val_auc: 0.9942 - val_prc: 0.7416
Epoch 8/100
278/278 [==============================] - 5s 20ms/step - loss: 0.1113 - tp: 265521.0000 - fp: 7871.0000 - tn: 277346.0000 - fn: 18606.0000 - accuracy: 0.9535 - precision: 0.9712 - recall: 0.9345 - auc: 0.9931 - prc: 0.9932 - val_loss: 0.0580 - val_tp: 67.0000 - val_fp: 702.0000 - val_tn: 44794.0000 - val_fn: 6.0000 - val_accuracy: 0.9845 - val_precision: 0.0871 - val_recall: 0.9178 - val_auc: 0.9938 - val_prc: 0.7420
Epoch 9/100
278/278 [==============================] - 5s 20ms/step - loss: 0.1055 - tp: 266759.0000 - fp: 7648.0000 - tn: 277146.0000 - fn: 17791.0000 - accuracy: 0.9553 - precision: 0.9721 - recall: 0.9375 - auc: 0.9938 - prc: 0.9939 - val_loss: 0.0528 - val_tp: 67.0000 - val_fp: 675.0000 - val_tn: 44821.0000 - val_fn: 6.0000 - val_accuracy: 0.9851 - val_precision: 0.0903 - val_recall: 0.9178 - val_auc: 0.9934 - val_prc: 0.7426
Epoch 10/100
278/278 [==============================] - 5s 19ms/step - loss: 0.1016 - tp: 267885.0000 - fp: 7444.0000 - tn: 276857.0000 - fn: 17158.0000 - accuracy: 0.9568 - precision: 0.9730 - recall: 0.9398 - auc: 0.9943 - prc: 0.9942 - val_loss: 0.0496 - val_tp: 67.0000 - val_fp: 662.0000 - val_tn: 44834.0000 - val_fn: 6.0000 - val_accuracy: 0.9853 - val_precision: 0.0919 - val_recall: 0.9178 - val_auc: 0.9930 - val_prc: 0.7243
Epoch 11/100
278/278 [==============================] - 5s 19ms/step - loss: 0.0974 - tp: 268749.0000 - fp: 7419.0000 - tn: 277074.0000 - fn: 16102.0000 - accuracy: 0.9587 - precision: 0.9731 - recall: 0.9435 - auc: 0.9947 - prc: 0.9946 - val_loss: 0.0465 - val_tp: 67.0000 - val_fp: 632.0000 - val_tn: 44864.0000 - val_fn: 6.0000 - val_accuracy: 0.9860 - val_precision: 0.0959 - val_recall: 0.9178 - val_auc: 0.9928 - val_prc: 0.7248
Epoch 12/100
278/278 [==============================] - 5s 19ms/step - loss: 0.0944 - tp: 269498.0000 - fp: 7384.0000 - tn: 277085.0000 - fn: 15377.0000 - accuracy: 0.9600 - precision: 0.9733 - recall: 0.9460 - auc: 0.9950 - prc: 0.9949 - val_loss: 0.0455 - val_tp: 67.0000 - val_fp: 641.0000 - val_tn: 44855.0000 - val_fn: 6.0000 - val_accuracy: 0.9858 - val_precision: 0.0946 - val_recall: 0.9178 - val_auc: 0.9927 - val_prc: 0.7070
Epoch 13/100
278/278 [==============================] - ETA: 0s - loss: 0.0912 - tp: 269571.0000 - fp: 7399.0000 - tn: 277708.0000 - fn: 14666.0000 - accuracy: 0.9612 - precision: 0.9733 - recall: 0.9484 - auc: 0.9953 - prc: 0.9951Restoring model weights from the end of the best epoch: 3.
278/278 [==============================] - 5s 19ms/step - loss: 0.0912 - tp: 269571.0000 - fp: 7399.0000 - tn: 277708.0000 - fn: 14666.0000 - accuracy: 0.9612 - precision: 0.9733 - recall: 0.9484 - auc: 0.9953 - prc: 0.9951 - val_loss: 0.0422 - val_tp: 67.0000 - val_fp: 591.0000 - val_tn: 44905.0000 - val_fn: 6.0000 - val_accuracy: 0.9869 - val_precision: 0.1018 - val_recall: 0.9178 - val_auc: 0.9926 - val_prc: 0.7169
Epoch 13: early stopping

If the training process were considering the whole dataset on each gradient update, this oversampling would be basically identical to the class weighting.

But when training the model batch-wise, as you did here, the oversampled data provides a smoother gradient signal: Instead of each positive example being shown in one batch with a large weight, they're shown in many different batches each time with a small weight.

This smoother gradient signal makes it easier to train the model.

Check training history

Note that the distributions of metrics will be different here, because the training data has a totally different distribution from the validation and test data.

plot_metrics(resampled_history)

Re-train

Because training is easier on the balanced data, the above training procedure may overfit quickly.

So break up the epochs to give the tf.keras.callbacks.EarlyStopping finer control over when to stop training.

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 76ms/step - loss: 2.0054 - tp: 6299.0000 - fp: 10480.0000 - tn: 55419.0000 - fn: 14331.0000 - accuracy: 0.7133 - precision: 0.3754 - recall: 0.3053 - auc: 0.6631 - prc: 0.4161 - val_loss: 0.7601 - val_tp: 21.0000 - val_fp: 21680.0000 - val_tn: 23816.0000 - val_fn: 52.0000 - val_accuracy: 0.5231 - val_precision: 9.6770e-04 - val_recall: 0.2877 - val_auc: 0.3172 - val_prc: 0.0014
Epoch 2/1000
20/20 [==============================] - 0s 24ms/step - loss: 1.2359 - tp: 10247.0000 - fp: 10132.0000 - tn: 10397.0000 - fn: 10184.0000 - accuracy: 0.5040 - precision: 0.5028 - recall: 0.5015 - auc: 0.4999 - prc: 0.6413 - val_loss: 0.7564 - val_tp: 48.0000 - val_fp: 21320.0000 - val_tn: 24176.0000 - val_fn: 25.0000 - val_accuracy: 0.5316 - val_precision: 0.0022 - val_recall: 0.6575 - val_auc: 0.6533 - val_prc: 0.0485
Epoch 3/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.8576 - tp: 13612.0000 - fp: 9694.0000 - tn: 10667.0000 - fn: 6987.0000 - accuracy: 0.5927 - precision: 0.5841 - recall: 0.6608 - auc: 0.6607 - prc: 0.7634 - val_loss: 0.7143 - val_tp: 63.0000 - val_fp: 19194.0000 - val_tn: 26302.0000 - val_fn: 10.0000 - val_accuracy: 0.5786 - val_precision: 0.0033 - val_recall: 0.8630 - val_auc: 0.8577 - val_prc: 0.1507
Epoch 4/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.6847 - tp: 15195.0000 - fp: 8650.0000 - tn: 11869.0000 - fn: 5246.0000 - accuracy: 0.6607 - precision: 0.6372 - recall: 0.7434 - auc: 0.7550 - prc: 0.8294 - val_loss: 0.6522 - val_tp: 65.0000 - val_fp: 16012.0000 - val_tn: 29484.0000 - val_fn: 8.0000 - val_accuracy: 0.6484 - val_precision: 0.0040 - val_recall: 0.8904 - val_auc: 0.9065 - val_prc: 0.2609
Epoch 5/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.5836 - tp: 16133.0000 - fp: 7573.0000 - tn: 12751.0000 - fn: 4503.0000 - accuracy: 0.7052 - precision: 0.6805 - recall: 0.7818 - auc: 0.8069 - prc: 0.8688 - val_loss: 0.5896 - val_tp: 66.0000 - val_fp: 12822.0000 - val_tn: 32674.0000 - val_fn: 7.0000 - val_accuracy: 0.7185 - val_precision: 0.0051 - val_recall: 0.9041 - val_auc: 0.9259 - val_prc: 0.3880
Epoch 6/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.5186 - tp: 16451.0000 - fp: 6555.0000 - tn: 13918.0000 - fn: 4036.0000 - accuracy: 0.7414 - precision: 0.7151 - recall: 0.8030 - auc: 0.8409 - prc: 0.8907 - val_loss: 0.5300 - val_tp: 67.0000 - val_fp: 9680.0000 - val_tn: 35816.0000 - val_fn: 6.0000 - val_accuracy: 0.7874 - val_precision: 0.0069 - val_recall: 0.9178 - val_auc: 0.9434 - val_prc: 0.5008
Epoch 7/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.4782 - tp: 16712.0000 - fp: 5743.0000 - tn: 14709.0000 - fn: 3796.0000 - accuracy: 0.7671 - precision: 0.7442 - recall: 0.8149 - auc: 0.8600 - prc: 0.9032 - val_loss: 0.4763 - val_tp: 68.0000 - val_fp: 7080.0000 - val_tn: 38416.0000 - val_fn: 5.0000 - val_accuracy: 0.8445 - val_precision: 0.0095 - val_recall: 0.9315 - val_auc: 0.9555 - val_prc: 0.5643
Epoch 8/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.4357 - tp: 16815.0000 - fp: 4970.0000 - tn: 15687.0000 - fn: 3488.0000 - accuracy: 0.7935 - precision: 0.7719 - recall: 0.8282 - auc: 0.8799 - prc: 0.9157 - val_loss: 0.4287 - val_tp: 68.0000 - val_fp: 4974.0000 - val_tn: 40522.0000 - val_fn: 5.0000 - val_accuracy: 0.8907 - val_precision: 0.0135 - val_recall: 0.9315 - val_auc: 0.9642 - val_prc: 0.6281
Epoch 9/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.4052 - tp: 16931.0000 - fp: 4132.0000 - tn: 16459.0000 - fn: 3438.0000 - accuracy: 0.8152 - precision: 0.8038 - recall: 0.8312 - auc: 0.8922 - prc: 0.9238 - val_loss: 0.3893 - val_tp: 68.0000 - val_fp: 3646.0000 - val_tn: 41850.0000 - val_fn: 5.0000 - val_accuracy: 0.9199 - val_precision: 0.0183 - val_recall: 0.9315 - val_auc: 0.9709 - val_prc: 0.6681
Epoch 10/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.3701 - tp: 17362.0000 - fp: 3603.0000 - tn: 16820.0000 - fn: 3175.0000 - accuracy: 0.8345 - precision: 0.8281 - recall: 0.8454 - auc: 0.9093 - prc: 0.9364 - val_loss: 0.3556 - val_tp: 69.0000 - val_fp: 2818.0000 - val_tn: 42678.0000 - val_fn: 4.0000 - val_accuracy: 0.9381 - val_precision: 0.0239 - val_recall: 0.9452 - val_auc: 0.9771 - val_prc: 0.6858
Epoch 11/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.3507 - tp: 17566.0000 - fp: 3172.0000 - tn: 17183.0000 - fn: 3039.0000 - accuracy: 0.8484 - precision: 0.8470 - recall: 0.8525 - auc: 0.9176 - prc: 0.9416 - val_loss: 0.3269 - val_tp: 69.0000 - val_fp: 2256.0000 - val_tn: 43240.0000 - val_fn: 4.0000 - val_accuracy: 0.9504 - val_precision: 0.0297 - val_recall: 0.9452 - val_auc: 0.9811 - val_prc: 0.6952
Epoch 12/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.3325 - tp: 17372.0000 - fp: 2859.0000 - tn: 17890.0000 - fn: 2839.0000 - accuracy: 0.8609 - precision: 0.8587 - recall: 0.8595 - auc: 0.9268 - prc: 0.9460 - val_loss: 0.2991 - val_tp: 68.0000 - val_fp: 1867.0000 - val_tn: 43629.0000 - val_fn: 5.0000 - val_accuracy: 0.9589 - val_precision: 0.0351 - val_recall: 0.9315 - val_auc: 0.9837 - val_prc: 0.7017
Epoch 13/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.3169 - tp: 17690.0000 - fp: 2524.0000 - tn: 17896.0000 - fn: 2850.0000 - accuracy: 0.8688 - precision: 0.8751 - recall: 0.8612 - auc: 0.9308 - prc: 0.9503 - val_loss: 0.2772 - val_tp: 68.0000 - val_fp: 1629.0000 - val_tn: 43867.0000 - val_fn: 5.0000 - val_accuracy: 0.9641 - val_precision: 0.0401 - val_recall: 0.9315 - val_auc: 0.9858 - val_prc: 0.7106
Epoch 14/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2969 - tp: 17677.0000 - fp: 2190.0000 - tn: 18385.0000 - fn: 2708.0000 - accuracy: 0.8804 - precision: 0.8898 - recall: 0.8672 - auc: 0.9400 - prc: 0.9555 - val_loss: 0.2586 - val_tp: 68.0000 - val_fp: 1477.0000 - val_tn: 44019.0000 - val_fn: 5.0000 - val_accuracy: 0.9675 - val_precision: 0.0440 - val_recall: 0.9315 - val_auc: 0.9874 - val_prc: 0.7132
Epoch 15/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2867 - tp: 17761.0000 - fp: 2037.0000 - tn: 18478.0000 - fn: 2684.0000 - accuracy: 0.8847 - precision: 0.8971 - recall: 0.8687 - auc: 0.9435 - prc: 0.9583 - val_loss: 0.2422 - val_tp: 68.0000 - val_fp: 1392.0000 - val_tn: 44104.0000 - val_fn: 5.0000 - val_accuracy: 0.9693 - val_precision: 0.0466 - val_recall: 0.9315 - val_auc: 0.9886 - val_prc: 0.7195
Epoch 16/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.2744 - tp: 18001.0000 - fp: 1857.0000 - tn: 18569.0000 - fn: 2533.0000 - accuracy: 0.8928 - precision: 0.9065 - recall: 0.8766 - auc: 0.9482 - prc: 0.9613 - val_loss: 0.2278 - val_tp: 67.0000 - val_fp: 1315.0000 - val_tn: 44181.0000 - val_fn: 6.0000 - val_accuracy: 0.9710 - val_precision: 0.0485 - val_recall: 0.9178 - val_auc: 0.9898 - val_prc: 0.7274
Epoch 17/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2630 - tp: 17977.0000 - fp: 1754.0000 - tn: 18791.0000 - fn: 2438.0000 - accuracy: 0.8977 - precision: 0.9111 - recall: 0.8806 - auc: 0.9528 - prc: 0.9643 - val_loss: 0.2148 - val_tp: 67.0000 - val_fp: 1254.0000 - val_tn: 44242.0000 - val_fn: 6.0000 - val_accuracy: 0.9723 - val_precision: 0.0507 - val_recall: 0.9178 - val_auc: 0.9909 - val_prc: 0.7300
Epoch 18/1000
20/20 [==============================] - 1s 26ms/step - loss: 0.2541 - tp: 18176.0000 - fp: 1670.0000 - tn: 18783.0000 - fn: 2331.0000 - accuracy: 0.9023 - precision: 0.9159 - recall: 0.8863 - auc: 0.9562 - prc: 0.9667 - val_loss: 0.2034 - val_tp: 66.0000 - val_fp: 1235.0000 - val_tn: 44261.0000 - val_fn: 7.0000 - val_accuracy: 0.9727 - val_precision: 0.0507 - val_recall: 0.9041 - val_auc: 0.9915 - val_prc: 0.7287
Epoch 19/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2463 - tp: 18088.0000 - fp: 1630.0000 - tn: 18963.0000 - fn: 2279.0000 - accuracy: 0.9046 - precision: 0.9173 - recall: 0.8881 - auc: 0.9588 - prc: 0.9683 - val_loss: 0.1922 - val_tp: 66.0000 - val_fp: 1178.0000 - val_tn: 44318.0000 - val_fn: 7.0000 - val_accuracy: 0.9740 - val_precision: 0.0531 - val_recall: 0.9041 - val_auc: 0.9921 - val_prc: 0.7335
Epoch 20/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.2341 - tp: 18073.0000 - fp: 1470.0000 - tn: 19196.0000 - fn: 2221.0000 - accuracy: 0.9099 - precision: 0.9248 - recall: 0.8906 - auc: 0.9623 - prc: 0.9709 - val_loss: 0.1821 - val_tp: 66.0000 - val_fp: 1144.0000 - val_tn: 44352.0000 - val_fn: 7.0000 - val_accuracy: 0.9747 - val_precision: 0.0545 - val_recall: 0.9041 - val_auc: 0.9927 - val_prc: 0.7345
Epoch 21/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2331 - tp: 18126.0000 - fp: 1436.0000 - tn: 19189.0000 - fn: 2209.0000 - accuracy: 0.9110 - precision: 0.9266 - recall: 0.8914 - auc: 0.9632 - prc: 0.9712 - val_loss: 0.1746 - val_tp: 66.0000 - val_fp: 1125.0000 - val_tn: 44371.0000 - val_fn: 7.0000 - val_accuracy: 0.9752 - val_precision: 0.0554 - val_recall: 0.9041 - val_auc: 0.9930 - val_prc: 0.7352
Epoch 22/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2236 - tp: 18526.0000 - fp: 1343.0000 - tn: 18861.0000 - fn: 2230.0000 - accuracy: 0.9128 - precision: 0.9324 - recall: 0.8926 - auc: 0.9655 - prc: 0.9737 - val_loss: 0.1672 - val_tp: 66.0000 - val_fp: 1097.0000 - val_tn: 44399.0000 - val_fn: 7.0000 - val_accuracy: 0.9758 - val_precision: 0.0567 - val_recall: 0.9041 - val_auc: 0.9935 - val_prc: 0.7383
Epoch 23/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2197 - tp: 18364.0000 - fp: 1309.0000 - tn: 19065.0000 - fn: 2222.0000 - accuracy: 0.9138 - precision: 0.9335 - recall: 0.8921 - auc: 0.9673 - prc: 0.9744 - val_loss: 0.1606 - val_tp: 66.0000 - val_fp: 1095.0000 - val_tn: 44401.0000 - val_fn: 7.0000 - val_accuracy: 0.9758 - val_precision: 0.0568 - val_recall: 0.9041 - val_auc: 0.9938 - val_prc: 0.7416
Epoch 24/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2136 - tp: 18337.0000 - fp: 1215.0000 - tn: 19260.0000 - fn: 2148.0000 - accuracy: 0.9179 - precision: 0.9379 - recall: 0.8951 - auc: 0.9689 - prc: 0.9754 - val_loss: 0.1540 - val_tp: 66.0000 - val_fp: 1069.0000 - val_tn: 44427.0000 - val_fn: 7.0000 - val_accuracy: 0.9764 - val_precision: 0.0581 - val_recall: 0.9041 - val_auc: 0.9941 - val_prc: 0.7417
Epoch 25/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2137 - tp: 18513.0000 - fp: 1200.0000 - tn: 19114.0000 - fn: 2133.0000 - accuracy: 0.9186 - precision: 0.9391 - recall: 0.8967 - auc: 0.9692 - prc: 0.9759 - val_loss: 0.1487 - val_tp: 66.0000 - val_fp: 1055.0000 - val_tn: 44441.0000 - val_fn: 7.0000 - val_accuracy: 0.9767 - val_precision: 0.0589 - val_recall: 0.9041 - val_auc: 0.9941 - val_prc: 0.7407
Epoch 26/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2051 - tp: 18203.0000 - fp: 1144.0000 - tn: 19546.0000 - fn: 2067.0000 - accuracy: 0.9216 - precision: 0.9409 - recall: 0.8980 - auc: 0.9716 - prc: 0.9767 - val_loss: 0.1432 - val_tp: 67.0000 - val_fp: 1023.0000 - val_tn: 44473.0000 - val_fn: 6.0000 - val_accuracy: 0.9774 - val_precision: 0.0615 - val_recall: 0.9178 - val_auc: 0.9942 - val_prc: 0.7437
Epoch 27/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.2017 - tp: 18422.0000 - fp: 1109.0000 - tn: 19356.0000 - fn: 2073.0000 - accuracy: 0.9223 - precision: 0.9432 - recall: 0.8989 - auc: 0.9727 - prc: 0.9778 - val_loss: 0.1376 - val_tp: 67.0000 - val_fp: 995.0000 - val_tn: 44501.0000 - val_fn: 6.0000 - val_accuracy: 0.9780 - val_precision: 0.0631 - val_recall: 0.9178 - val_auc: 0.9944 - val_prc: 0.7452
Epoch 28/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1993 - tp: 18544.0000 - fp: 1137.0000 - tn: 19232.0000 - fn: 2047.0000 - accuracy: 0.9223 - precision: 0.9422 - recall: 0.9006 - auc: 0.9736 - prc: 0.9787 - val_loss: 0.1327 - val_tp: 67.0000 - val_fp: 985.0000 - val_tn: 44511.0000 - val_fn: 6.0000 - val_accuracy: 0.9783 - val_precision: 0.0637 - val_recall: 0.9178 - val_auc: 0.9946 - val_prc: 0.7468
Epoch 29/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1967 - tp: 18420.0000 - fp: 1099.0000 - tn: 19427.0000 - fn: 2014.0000 - accuracy: 0.9240 - precision: 0.9437 - recall: 0.9014 - auc: 0.9741 - prc: 0.9790 - val_loss: 0.1287 - val_tp: 67.0000 - val_fp: 971.0000 - val_tn: 44525.0000 - val_fn: 6.0000 - val_accuracy: 0.9786 - val_precision: 0.0645 - val_recall: 0.9178 - val_auc: 0.9947 - val_prc: 0.7482
Epoch 30/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1900 - tp: 18615.0000 - fp: 974.0000 - tn: 19378.0000 - fn: 1993.0000 - accuracy: 0.9276 - precision: 0.9503 - recall: 0.9033 - auc: 0.9756 - prc: 0.9805 - val_loss: 0.1254 - val_tp: 67.0000 - val_fp: 975.0000 - val_tn: 44521.0000 - val_fn: 6.0000 - val_accuracy: 0.9785 - val_precision: 0.0643 - val_recall: 0.9178 - val_auc: 0.9948 - val_prc: 0.7485
Epoch 31/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1842 - tp: 18482.0000 - fp: 958.0000 - tn: 19592.0000 - fn: 1928.0000 - accuracy: 0.9295 - precision: 0.9507 - recall: 0.9055 - auc: 0.9770 - prc: 0.9811 - val_loss: 0.1222 - val_tp: 67.0000 - val_fp: 967.0000 - val_tn: 44529.0000 - val_fn: 6.0000 - val_accuracy: 0.9786 - val_precision: 0.0648 - val_recall: 0.9178 - val_auc: 0.9949 - val_prc: 0.7494
Epoch 32/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1817 - tp: 18693.0000 - fp: 926.0000 - tn: 19423.0000 - fn: 1918.0000 - accuracy: 0.9306 - precision: 0.9528 - recall: 0.9069 - auc: 0.9776 - prc: 0.9817 - val_loss: 0.1195 - val_tp: 67.0000 - val_fp: 963.0000 - val_tn: 44533.0000 - val_fn: 6.0000 - val_accuracy: 0.9787 - val_precision: 0.0650 - val_recall: 0.9178 - val_auc: 0.9949 - val_prc: 0.7502
Epoch 33/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1804 - tp: 18668.0000 - fp: 971.0000 - tn: 19405.0000 - fn: 1916.0000 - accuracy: 0.9295 - precision: 0.9506 - recall: 0.9069 - auc: 0.9780 - prc: 0.9821 - val_loss: 0.1164 - val_tp: 67.0000 - val_fp: 957.0000 - val_tn: 44539.0000 - val_fn: 6.0000 - val_accuracy: 0.9789 - val_precision: 0.0654 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7516
Epoch 34/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1772 - tp: 18630.0000 - fp: 878.0000 - tn: 19607.0000 - fn: 1845.0000 - accuracy: 0.9335 - precision: 0.9550 - recall: 0.9099 - auc: 0.9789 - prc: 0.9825 - val_loss: 0.1132 - val_tp: 67.0000 - val_fp: 938.0000 - val_tn: 44558.0000 - val_fn: 6.0000 - val_accuracy: 0.9793 - val_precision: 0.0667 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7518
Epoch 35/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1747 - tp: 18664.0000 - fp: 893.0000 - tn: 19523.0000 - fn: 1880.0000 - accuracy: 0.9323 - precision: 0.9543 - recall: 0.9085 - auc: 0.9793 - prc: 0.9829 - val_loss: 0.1110 - val_tp: 67.0000 - val_fp: 936.0000 - val_tn: 44560.0000 - val_fn: 6.0000 - val_accuracy: 0.9793 - val_precision: 0.0668 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7523
Epoch 36/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1753 - tp: 18668.0000 - fp: 877.0000 - tn: 19531.0000 - fn: 1884.0000 - accuracy: 0.9326 - precision: 0.9551 - recall: 0.9083 - auc: 0.9790 - prc: 0.9830 - val_loss: 0.1089 - val_tp: 67.0000 - val_fp: 936.0000 - val_tn: 44560.0000 - val_fn: 6.0000 - val_accuracy: 0.9793 - val_precision: 0.0668 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7560
Epoch 37/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1691 - tp: 18684.0000 - fp: 866.0000 - tn: 19598.0000 - fn: 1812.0000 - accuracy: 0.9346 - precision: 0.9557 - recall: 0.9116 - auc: 0.9810 - prc: 0.9842 - val_loss: 0.1062 - val_tp: 67.0000 - val_fp: 909.0000 - val_tn: 44587.0000 - val_fn: 6.0000 - val_accuracy: 0.9799 - val_precision: 0.0686 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7560
Epoch 38/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1691 - tp: 18741.0000 - fp: 830.0000 - tn: 19553.0000 - fn: 1836.0000 - accuracy: 0.9349 - precision: 0.9576 - recall: 0.9108 - auc: 0.9810 - prc: 0.9842 - val_loss: 0.1041 - val_tp: 67.0000 - val_fp: 903.0000 - val_tn: 44593.0000 - val_fn: 6.0000 - val_accuracy: 0.9801 - val_precision: 0.0691 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7564
Epoch 39/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1697 - tp: 18739.0000 - fp: 841.0000 - tn: 19556.0000 - fn: 1824.0000 - accuracy: 0.9349 - precision: 0.9570 - recall: 0.9113 - auc: 0.9811 - prc: 0.9845 - val_loss: 0.1018 - val_tp: 67.0000 - val_fp: 888.0000 - val_tn: 44608.0000 - val_fn: 6.0000 - val_accuracy: 0.9804 - val_precision: 0.0702 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7563
Epoch 40/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1637 - tp: 18653.0000 - fp: 808.0000 - tn: 19662.0000 - fn: 1837.0000 - accuracy: 0.9354 - precision: 0.9585 - recall: 0.9103 - auc: 0.9820 - prc: 0.9850 - val_loss: 0.1005 - val_tp: 67.0000 - val_fp: 893.0000 - val_tn: 44603.0000 - val_fn: 6.0000 - val_accuracy: 0.9803 - val_precision: 0.0698 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7564
Epoch 41/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1624 - tp: 18820.0000 - fp: 790.0000 - tn: 19638.0000 - fn: 1712.0000 - accuracy: 0.9389 - precision: 0.9597 - recall: 0.9166 - auc: 0.9827 - prc: 0.9855 - val_loss: 0.0986 - val_tp: 67.0000 - val_fp: 888.0000 - val_tn: 44608.0000 - val_fn: 6.0000 - val_accuracy: 0.9804 - val_precision: 0.0702 - val_recall: 0.9178 - val_auc: 0.9950 - val_prc: 0.7568
Epoch 42/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1605 - tp: 18760.0000 - fp: 761.0000 - tn: 19682.0000 - fn: 1757.0000 - accuracy: 0.9385 - precision: 0.9610 - recall: 0.9144 - auc: 0.9828 - prc: 0.9857 - val_loss: 0.0976 - val_tp: 67.0000 - val_fp: 899.0000 - val_tn: 44597.0000 - val_fn: 6.0000 - val_accuracy: 0.9801 - val_precision: 0.0694 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7566
Epoch 43/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1600 - tp: 18767.0000 - fp: 766.0000 - tn: 19694.0000 - fn: 1733.0000 - accuracy: 0.9390 - precision: 0.9608 - recall: 0.9155 - auc: 0.9832 - prc: 0.9859 - val_loss: 0.0958 - val_tp: 67.0000 - val_fp: 881.0000 - val_tn: 44615.0000 - val_fn: 6.0000 - val_accuracy: 0.9805 - val_precision: 0.0707 - val_recall: 0.9178 - val_auc: 0.9951 - val_prc: 0.7568
Epoch 44/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1598 - tp: 18753.0000 - fp: 778.0000 - tn: 19681.0000 - fn: 1748.0000 - accuracy: 0.9383 - precision: 0.9602 - recall: 0.9147 - auc: 0.9831 - prc: 0.9856 - val_loss: 0.0945 - val_tp: 67.0000 - val_fp: 876.0000 - val_tn: 44620.0000 - val_fn: 6.0000 - val_accuracy: 0.9806 - val_precision: 0.0710 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7568
Epoch 45/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1579 - tp: 18598.0000 - fp: 790.0000 - tn: 19840.0000 - fn: 1732.0000 - accuracy: 0.9384 - precision: 0.9593 - recall: 0.9148 - auc: 0.9834 - prc: 0.9859 - val_loss: 0.0932 - val_tp: 67.0000 - val_fp: 869.0000 - val_tn: 44627.0000 - val_fn: 6.0000 - val_accuracy: 0.9808 - val_precision: 0.0716 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7566
Epoch 46/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1561 - tp: 18780.0000 - fp: 760.0000 - tn: 19692.0000 - fn: 1728.0000 - accuracy: 0.9393 - precision: 0.9611 - recall: 0.9157 - auc: 0.9843 - prc: 0.9864 - val_loss: 0.0917 - val_tp: 67.0000 - val_fp: 858.0000 - val_tn: 44638.0000 - val_fn: 6.0000 - val_accuracy: 0.9810 - val_precision: 0.0724 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7566
Epoch 47/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1583 - tp: 18786.0000 - fp: 769.0000 - tn: 19661.0000 - fn: 1744.0000 - accuracy: 0.9386 - precision: 0.9607 - recall: 0.9151 - auc: 0.9835 - prc: 0.9861 - val_loss: 0.0905 - val_tp: 67.0000 - val_fp: 851.0000 - val_tn: 44645.0000 - val_fn: 6.0000 - val_accuracy: 0.9812 - val_precision: 0.0730 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7568
Epoch 48/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1526 - tp: 18597.0000 - fp: 679.0000 - tn: 19964.0000 - fn: 1720.0000 - accuracy: 0.9414 - precision: 0.9648 - recall: 0.9153 - auc: 0.9848 - prc: 0.9868 - val_loss: 0.0891 - val_tp: 67.0000 - val_fp: 842.0000 - val_tn: 44654.0000 - val_fn: 6.0000 - val_accuracy: 0.9814 - val_precision: 0.0737 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7489
Epoch 49/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1523 - tp: 18795.0000 - fp: 701.0000 - tn: 19731.0000 - fn: 1733.0000 - accuracy: 0.9406 - precision: 0.9640 - recall: 0.9156 - auc: 0.9845 - prc: 0.9870 - val_loss: 0.0886 - val_tp: 67.0000 - val_fp: 851.0000 - val_tn: 44645.0000 - val_fn: 6.0000 - val_accuracy: 0.9812 - val_precision: 0.0730 - val_recall: 0.9178 - val_auc: 0.9954 - val_prc: 0.7494
Epoch 50/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1479 - tp: 18782.0000 - fp: 682.0000 - tn: 19808.0000 - fn: 1688.0000 - accuracy: 0.9421 - precision: 0.9650 - recall: 0.9175 - auc: 0.9858 - prc: 0.9877 - val_loss: 0.0883 - val_tp: 67.0000 - val_fp: 867.0000 - val_tn: 44629.0000 - val_fn: 6.0000 - val_accuracy: 0.9808 - val_precision: 0.0717 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7491
Epoch 51/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1490 - tp: 18867.0000 - fp: 712.0000 - tn: 19683.0000 - fn: 1698.0000 - accuracy: 0.9412 - precision: 0.9636 - recall: 0.9174 - auc: 0.9856 - prc: 0.9877 - val_loss: 0.0883 - val_tp: 67.0000 - val_fp: 884.0000 - val_tn: 44612.0000 - val_fn: 6.0000 - val_accuracy: 0.9805 - val_precision: 0.0705 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7487
Epoch 52/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1485 - tp: 18946.0000 - fp: 716.0000 - tn: 19653.0000 - fn: 1645.0000 - accuracy: 0.9424 - precision: 0.9636 - recall: 0.9201 - auc: 0.9858 - prc: 0.9878 - val_loss: 0.0876 - val_tp: 67.0000 - val_fp: 891.0000 - val_tn: 44605.0000 - val_fn: 6.0000 - val_accuracy: 0.9803 - val_precision: 0.0699 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7488
Epoch 53/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1466 - tp: 18795.0000 - fp: 684.0000 - tn: 19882.0000 - fn: 1599.0000 - accuracy: 0.9443 - precision: 0.9649 - recall: 0.9216 - auc: 0.9860 - prc: 0.9877 - val_loss: 0.0870 - val_tp: 67.0000 - val_fp: 897.0000 - val_tn: 44599.0000 - val_fn: 6.0000 - val_accuracy: 0.9802 - val_precision: 0.0695 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7488
Epoch 54/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1482 - tp: 19074.0000 - fp: 738.0000 - tn: 19480.0000 - fn: 1668.0000 - accuracy: 0.9413 - precision: 0.9627 - recall: 0.9196 - auc: 0.9860 - prc: 0.9880 - val_loss: 0.0856 - val_tp: 67.0000 - val_fp: 885.0000 - val_tn: 44611.0000 - val_fn: 6.0000 - val_accuracy: 0.9804 - val_precision: 0.0704 - val_recall: 0.9178 - val_auc: 0.9953 - val_prc: 0.7490
Epoch 55/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1456 - tp: 18935.0000 - fp: 713.0000 - tn: 19657.0000 - fn: 1655.0000 - accuracy: 0.9422 - precision: 0.9637 - recall: 0.9196 - auc: 0.9864 - prc: 0.9883 - val_loss: 0.0850 - val_tp: 67.0000 - val_fp: 894.0000 - val_tn: 44602.0000 - val_fn: 6.0000 - val_accuracy: 0.9802 - val_precision: 0.0697 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7491
Epoch 56/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1457 - tp: 18855.0000 - fp: 730.0000 - tn: 19772.0000 - fn: 1603.0000 - accuracy: 0.9430 - precision: 0.9627 - recall: 0.9216 - auc: 0.9865 - prc: 0.9879 - val_loss: 0.0834 - val_tp: 67.0000 - val_fp: 881.0000 - val_tn: 44615.0000 - val_fn: 6.0000 - val_accuracy: 0.9805 - val_precision: 0.0707 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7494
Epoch 57/1000
20/20 [==============================] - ETA: 0s - loss: 0.1436 - tp: 18714.0000 - fp: 662.0000 - tn: 19927.0000 - fn: 1657.0000 - accuracy: 0.9434 - precision: 0.9658 - recall: 0.9187 - auc: 0.9868 - prc: 0.9884Restoring model weights from the end of the best epoch: 47.
20/20 [==============================] - 0s 24ms/step - loss: 0.1436 - tp: 18714.0000 - fp: 662.0000 - tn: 19927.0000 - fn: 1657.0000 - accuracy: 0.9434 - precision: 0.9658 - recall: 0.9187 - auc: 0.9868 - prc: 0.9884 - val_loss: 0.0822 - val_tp: 67.0000 - val_fp: 865.0000 - val_tn: 44631.0000 - val_fn: 6.0000 - val_accuracy: 0.9809 - val_precision: 0.0719 - val_recall: 0.9178 - val_auc: 0.9952 - val_prc: 0.7494
Epoch 57: early stopping

Re-check training history

plot_metrics(resampled_history)

Evaluate metrics

train_predictions_resampled = resampled_model.predict(train_features, batch_size=BATCH_SIZE)
test_predictions_resampled = resampled_model.predict(test_features, batch_size=BATCH_SIZE)

90/90 [==============================] - 0s 1ms/step
28/28 [==============================] - 0s 1ms/step

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 :  0.0926642045378685
tp :  87.0
fp :  1080.0
tn :  55786.0
fn :  9.0
accuracy :  0.9808819890022278
precision :  0.07455012947320938
recall :  0.90625
auc :  0.9739617109298706
prc :  0.6837594509124756

Legitimate Transactions Detected (True Negatives):  55786
Legitimate Transactions Incorrectly Detected (False Positives):  1080
Fraudulent Transactions Missed (False Negatives):  9
Fraudulent Transactions Detected (True Positives):  87
Total Fraudulent Transactions:  96

Plot the ROC

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');

Plot the 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');

Applying this tutorial to your problem

Imbalanced data classification is an inherently difficult task since there are so few samples to learn from. You should always start with the data first and do your best to collect as many samples as possible and give substantial thought to what features may be relevant so the model can get the most out of your minority class. At some point your model may struggle to improve and yield the results you want, so it is important to keep in mind the context of your problem and the trade offs between different types of errors.