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 and can be used to download CSVs into a dataframe.

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 Ammount'] = 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 our 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.

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)

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

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

png

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'),
]

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(lr=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.

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.

EPOCHS = 100
BATCH_SIZE = 2048

early_stopping = tf.keras.callbacks.EarlyStopping(
    monitor='val_auc', 
    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])

array([[0.5788107 ],
       [0.44979692],
       [0.5427961 ],
       [0.5985188 ],
       [0.7758075 ],
       [0.3417888 ],
       [0.39359283],
       [0.5399953 ],
       [0.3551327 ],
       [0.47230086]], 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: 0.7817

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])

array([[0.00093563],
       [0.00187903],
       [0.00109238],
       [0.00117128],
       [0.00134988],
       [0.00090826],
       [0.00099455],
       [0.00154405],
       [0.00100204],
       [0.0004291 ]], 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: 0.0146

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 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')
  
  plt.legend()

plot_loss(zero_bias_history, "Zero Bias", 0)
plot_loss(careful_bias_history, "Careful Bias", 1)

png

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 [==============================] - 1s 13ms/step - loss: 0.0112 - tp: 100.0000 - fp: 25.0000 - tn: 227419.0000 - fn: 301.0000 - accuracy: 0.9986 - precision: 0.8000 - recall: 0.2494 - auc: 0.7615 - val_loss: 0.0067 - val_tp: 15.0000 - val_fp: 2.0000 - val_tn: 45480.0000 - val_fn: 72.0000 - val_accuracy: 0.9984 - val_precision: 0.8824 - val_recall: 0.1724 - val_auc: 0.9077
Epoch 2/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0075 - tp: 108.0000 - fp: 24.0000 - tn: 181938.0000 - fn: 206.0000 - accuracy: 0.9987 - precision: 0.8182 - recall: 0.3439 - auc: 0.8491 - val_loss: 0.0046 - val_tp: 45.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 42.0000 - val_accuracy: 0.9989 - val_precision: 0.8824 - val_recall: 0.5172 - val_auc: 0.9308
Epoch 3/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0065 - tp: 138.0000 - fp: 27.0000 - tn: 181935.0000 - fn: 176.0000 - accuracy: 0.9989 - precision: 0.8364 - recall: 0.4395 - auc: 0.8567 - val_loss: 0.0040 - val_tp: 54.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 33.0000 - val_accuracy: 0.9991 - val_precision: 0.8852 - val_recall: 0.6207 - val_auc: 0.9365
Epoch 4/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0060 - tp: 154.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 160.0000 - accuracy: 0.9989 - precision: 0.8235 - recall: 0.4904 - auc: 0.8848 - val_loss: 0.0037 - val_tp: 61.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 26.0000 - val_accuracy: 0.9993 - val_precision: 0.8841 - val_recall: 0.7011 - val_auc: 0.9422
Epoch 5/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0057 - tp: 157.0000 - fp: 36.0000 - tn: 181926.0000 - fn: 157.0000 - accuracy: 0.9989 - precision: 0.8135 - recall: 0.5000 - auc: 0.8982 - val_loss: 0.0035 - val_tp: 62.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 25.0000 - val_accuracy: 0.9993 - val_precision: 0.8857 - val_recall: 0.7126 - val_auc: 0.9422
Epoch 6/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0057 - tp: 152.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 162.0000 - accuracy: 0.9989 - precision: 0.8261 - recall: 0.4841 - auc: 0.8934 - val_loss: 0.0033 - val_tp: 65.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8904 - val_recall: 0.7471 - val_auc: 0.9479
Epoch 7/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0052 - tp: 174.0000 - fp: 30.0000 - tn: 181932.0000 - fn: 140.0000 - accuracy: 0.9991 - precision: 0.8529 - recall: 0.5541 - auc: 0.8983 - val_loss: 0.0032 - val_tp: 66.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.8919 - val_recall: 0.7586 - val_auc: 0.9479
Epoch 8/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0054 - tp: 161.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 153.0000 - accuracy: 0.9990 - precision: 0.8342 - recall: 0.5127 - auc: 0.8983 - val_loss: 0.0031 - val_tp: 66.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.8919 - val_recall: 0.7586 - val_auc: 0.9479
Epoch 9/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0050 - tp: 167.0000 - fp: 37.0000 - tn: 181925.0000 - fn: 147.0000 - accuracy: 0.9990 - precision: 0.8186 - recall: 0.5318 - auc: 0.9064 - val_loss: 0.0030 - val_tp: 65.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 22.0000 - val_accuracy: 0.9993 - val_precision: 0.8904 - val_recall: 0.7471 - val_auc: 0.9479
Epoch 10/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0053 - tp: 156.0000 - fp: 34.0000 - tn: 181928.0000 - fn: 158.0000 - accuracy: 0.9989 - precision: 0.8211 - recall: 0.4968 - auc: 0.9046 - val_loss: 0.0029 - val_tp: 67.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 20.0000 - val_accuracy: 0.9994 - val_precision: 0.8933 - val_recall: 0.7701 - val_auc: 0.9479
Epoch 11/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0048 - tp: 165.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 149.0000 - accuracy: 0.9990 - precision: 0.8376 - recall: 0.5255 - auc: 0.9063 - val_loss: 0.0029 - val_tp: 68.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8947 - val_recall: 0.7816 - val_auc: 0.9479
Epoch 12/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0051 - tp: 165.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 149.0000 - accuracy: 0.9990 - precision: 0.8250 - recall: 0.5255 - auc: 0.9110 - val_loss: 0.0028 - val_tp: 67.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 20.0000 - val_accuracy: 0.9994 - val_precision: 0.8933 - val_recall: 0.7701 - val_auc: 0.9480
Epoch 13/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0050 - tp: 157.0000 - fp: 29.0000 - tn: 181933.0000 - fn: 157.0000 - accuracy: 0.9990 - precision: 0.8441 - recall: 0.5000 - auc: 0.9031 - val_loss: 0.0028 - val_tp: 69.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 18.0000 - val_accuracy: 0.9994 - val_precision: 0.8961 - val_recall: 0.7931 - val_auc: 0.9479
Epoch 14/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0053 - tp: 160.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 154.0000 - accuracy: 0.9990 - precision: 0.8205 - recall: 0.5096 - auc: 0.8934 - val_loss: 0.0027 - val_tp: 69.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 18.0000 - val_accuracy: 0.9994 - val_precision: 0.8961 - val_recall: 0.7931 - val_auc: 0.9479
Epoch 15/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0049 - tp: 168.0000 - fp: 36.0000 - tn: 181926.0000 - fn: 146.0000 - accuracy: 0.9990 - precision: 0.8235 - recall: 0.5350 - auc: 0.9031 - val_loss: 0.0027 - val_tp: 68.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8947 - val_recall: 0.7816 - val_auc: 0.9479
Epoch 16/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0046 - tp: 169.0000 - fp: 30.0000 - tn: 181932.0000 - fn: 145.0000 - accuracy: 0.9990 - precision: 0.8492 - recall: 0.5382 - auc: 0.9143 - val_loss: 0.0027 - val_tp: 68.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8947 - val_recall: 0.7816 - val_auc: 0.9537
Epoch 17/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 181.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 133.0000 - accuracy: 0.9991 - precision: 0.8498 - recall: 0.5764 - auc: 0.9144 - val_loss: 0.0027 - val_tp: 70.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 17.0000 - val_accuracy: 0.9995 - val_precision: 0.8974 - val_recall: 0.8046 - val_auc: 0.9537
Epoch 18/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 181.0000 - fp: 29.0000 - tn: 181933.0000 - fn: 133.0000 - accuracy: 0.9991 - precision: 0.8619 - recall: 0.5764 - auc: 0.9112 - val_loss: 0.0026 - val_tp: 69.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 18.0000 - val_accuracy: 0.9994 - val_precision: 0.8961 - val_recall: 0.7931 - val_auc: 0.9537
Epoch 19/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0046 - tp: 172.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 142.0000 - accuracy: 0.9990 - precision: 0.8431 - recall: 0.5478 - auc: 0.9096 - val_loss: 0.0026 - val_tp: 68.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 19.0000 - val_accuracy: 0.9994 - val_precision: 0.8947 - val_recall: 0.7816 - val_auc: 0.9537
Epoch 20/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 177.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 137.0000 - accuracy: 0.9991 - precision: 0.8349 - recall: 0.5637 - auc: 0.9128 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 21/100
90/90 [==============================] - 1s 7ms/step - loss: 0.0045 - tp: 176.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 138.0000 - accuracy: 0.9991 - precision: 0.8462 - recall: 0.5605 - auc: 0.9096 - val_loss: 0.0026 - val_tp: 66.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.9167 - val_recall: 0.7586 - val_auc: 0.9537
Epoch 22/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0047 - tp: 163.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 151.0000 - accuracy: 0.9990 - precision: 0.8316 - recall: 0.5191 - auc: 0.9096 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 23/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0046 - tp: 183.0000 - fp: 38.0000 - tn: 181924.0000 - fn: 131.0000 - accuracy: 0.9991 - precision: 0.8281 - recall: 0.5828 - auc: 0.9113 - val_loss: 0.0026 - val_tp: 66.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 21.0000 - val_accuracy: 0.9994 - val_precision: 0.9041 - val_recall: 0.7586 - val_auc: 0.9537
Epoch 24/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 168.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 146.0000 - accuracy: 0.9990 - precision: 0.8400 - recall: 0.5350 - auc: 0.9128 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 25/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 179.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 135.0000 - accuracy: 0.9991 - precision: 0.8483 - recall: 0.5701 - auc: 0.9161 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 26/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 173.0000 - fp: 38.0000 - tn: 181924.0000 - fn: 141.0000 - accuracy: 0.9990 - precision: 0.8199 - recall: 0.5510 - auc: 0.9208 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 27/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 172.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 142.0000 - accuracy: 0.9990 - precision: 0.8431 - recall: 0.5478 - auc: 0.9081 - val_loss: 0.0026 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 28/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 181.0000 - fp: 39.0000 - tn: 181923.0000 - fn: 133.0000 - accuracy: 0.9991 - precision: 0.8227 - recall: 0.5764 - auc: 0.9193 - val_loss: 0.0025 - val_tp: 68.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 19.0000 - val_accuracy: 0.9995 - val_precision: 0.9189 - val_recall: 0.7816 - val_auc: 0.9537
Epoch 29/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 177.0000 - fp: 38.0000 - tn: 181924.0000 - fn: 137.0000 - accuracy: 0.9990 - precision: 0.8233 - recall: 0.5637 - auc: 0.9305 - val_loss: 0.0025 - val_tp: 67.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 20.0000 - val_accuracy: 0.9994 - val_precision: 0.9054 - val_recall: 0.7701 - val_auc: 0.9538
Epoch 30/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 168.0000 - fp: 31.0000 - tn: 181931.0000 - fn: 146.0000 - accuracy: 0.9990 - precision: 0.8442 - recall: 0.5350 - auc: 0.9161 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9200 - val_recall: 0.7931 - val_auc: 0.9537
Epoch 31/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 172.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 142.0000 - accuracy: 0.9990 - precision: 0.8309 - recall: 0.5478 - auc: 0.9176 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9538
Epoch 32/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 188.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 126.0000 - accuracy: 0.9991 - precision: 0.8507 - recall: 0.5987 - auc: 0.9162 - val_loss: 0.0025 - val_tp: 70.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 17.0000 - val_accuracy: 0.9995 - val_precision: 0.9091 - val_recall: 0.8046 - val_auc: 0.9538
Epoch 33/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 184.0000 - fp: 27.0000 - tn: 181935.0000 - fn: 130.0000 - accuracy: 0.9991 - precision: 0.8720 - recall: 0.5860 - auc: 0.9225 - val_loss: 0.0025 - val_tp: 72.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.9000 - val_recall: 0.8276 - val_auc: 0.9537
Epoch 34/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 185.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 129.0000 - accuracy: 0.9991 - precision: 0.8486 - recall: 0.5892 - auc: 0.9273 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 35/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0044 - tp: 178.0000 - fp: 36.0000 - tn: 181926.0000 - fn: 136.0000 - accuracy: 0.9991 - precision: 0.8318 - recall: 0.5669 - auc: 0.9160 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 36/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0045 - tp: 171.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 143.0000 - accuracy: 0.9990 - precision: 0.8382 - recall: 0.5446 - auc: 0.9192 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9538
Epoch 37/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 189.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 125.0000 - accuracy: 0.9991 - precision: 0.8438 - recall: 0.6019 - auc: 0.9242 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9200 - val_recall: 0.7931 - val_auc: 0.9538
Epoch 38/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 185.0000 - fp: 25.0000 - tn: 181937.0000 - fn: 129.0000 - accuracy: 0.9992 - precision: 0.8810 - recall: 0.5892 - auc: 0.9176 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 39/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 181.0000 - fp: 35.0000 - tn: 181927.0000 - fn: 133.0000 - accuracy: 0.9991 - precision: 0.8380 - recall: 0.5764 - auc: 0.9225 - val_loss: 0.0025 - val_tp: 68.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 19.0000 - val_accuracy: 0.9995 - val_precision: 0.9189 - val_recall: 0.7816 - val_auc: 0.9538
Epoch 40/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 175.0000 - fp: 30.0000 - tn: 181932.0000 - fn: 139.0000 - accuracy: 0.9991 - precision: 0.8537 - recall: 0.5573 - auc: 0.9209 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9200 - val_recall: 0.7931 - val_auc: 0.9538
Epoch 41/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0041 - tp: 180.0000 - fp: 32.0000 - tn: 181930.0000 - fn: 134.0000 - accuracy: 0.9991 - precision: 0.8491 - recall: 0.5732 - auc: 0.9320 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9537
Epoch 42/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0040 - tp: 188.0000 - fp: 34.0000 - tn: 181928.0000 - fn: 126.0000 - accuracy: 0.9991 - precision: 0.8468 - recall: 0.5987 - auc: 0.9209 - val_loss: 0.0025 - val_tp: 71.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 16.0000 - val_accuracy: 0.9995 - val_precision: 0.8987 - val_recall: 0.8161 - val_auc: 0.9538
Epoch 43/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 176.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 138.0000 - accuracy: 0.9991 - precision: 0.8421 - recall: 0.5605 - auc: 0.9225 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9200 - val_recall: 0.7931 - val_auc: 0.9538
Epoch 44/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 172.0000 - fp: 37.0000 - tn: 181925.0000 - fn: 142.0000 - accuracy: 0.9990 - precision: 0.8230 - recall: 0.5478 - auc: 0.9129 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9079 - val_recall: 0.7931 - val_auc: 0.9537
Epoch 45/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 175.0000 - fp: 36.0000 - tn: 181926.0000 - fn: 139.0000 - accuracy: 0.9990 - precision: 0.8294 - recall: 0.5573 - auc: 0.9368 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9079 - val_recall: 0.7931 - val_auc: 0.9537
Epoch 46/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0043 - tp: 176.0000 - fp: 33.0000 - tn: 181929.0000 - fn: 138.0000 - accuracy: 0.9991 - precision: 0.8421 - recall: 0.5605 - auc: 0.9240 - val_loss: 0.0025 - val_tp: 69.0000 - val_fp: 7.0000 - val_tn: 45475.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9079 - val_recall: 0.7931 - val_auc: 0.9538
Epoch 47/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 178.0000 - fp: 27.0000 - tn: 181935.0000 - fn: 136.0000 - accuracy: 0.9991 - precision: 0.8683 - recall: 0.5669 - auc: 0.9273 - val_loss: 0.0025 - val_tp: 72.0000 - val_fp: 8.0000 - val_tn: 45474.0000 - val_fn: 15.0000 - val_accuracy: 0.9995 - val_precision: 0.9000 - val_recall: 0.8276 - val_auc: 0.9537
Epoch 48/100
90/90 [==============================] - 1s 6ms/step - loss: 0.0039 - tp: 198.0000 - fp: 34.0000 - tn: 181928.0000 - fn: 116.0000 - accuracy: 0.9992 - precision: 0.8534 - recall: 0.6306 - auc: 0.9256 - val_loss: 0.0025 - val_tp: 68.0000 - val_fp: 5.0000 - val_tn: 45477.0000 - val_fn: 19.0000 - val_accuracy: 0.9995 - val_precision: 0.9315 - val_recall: 0.7816 - val_auc: 0.9538
Epoch 49/100
85/90 [===========================>..] - ETA: 0s - loss: 0.0043 - tp: 162.0000 - fp: 29.0000 - tn: 173750.0000 - fn: 139.0000 - accuracy: 0.9990 - precision: 0.8482 - recall: 0.5382 - auc: 0.9157Restoring model weights from the end of the best epoch.
90/90 [==============================] - 1s 6ms/step - loss: 0.0042 - tp: 171.0000 - fp: 30.0000 - tn: 181932.0000 - fn: 143.0000 - accuracy: 0.9991 - precision: 0.8507 - recall: 0.5446 - auc: 0.9191 - val_loss: 0.0024 - val_tp: 69.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 18.0000 - val_accuracy: 0.9995 - val_precision: 0.9200 - val_recall: 0.7931 - val_auc: 0.9537
Epoch 00049: 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 this 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', 'auc', 'precision', 'recall']
  for n, metric in enumerate(metrics):
    name = metric.replace("_"," ").capitalize()
    plt.subplot(2,2,n+1)
    plt.plot(history.epoch,  history.history[metric], color=colors[0], label='Train')
    plt.plot(history.epoch, history.history['val_'+metric],
             color=colors[0], linestyle="--", label='Val')
    plt.xlabel('Epoch')
    plt.ylabel(name)
    if metric == 'loss':
      plt.ylim([0, plt.ylim()[1]])
    elif metric == 'auc':
      plt.ylim([0.8,1])
    else:
      plt.ylim([0,1])

    plt.legend()

plot_metrics(baseline_history)

png

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)

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 :  0.002310588490217924
tp :  69.0
fp :  5.0
tn :  56866.0
fn :  22.0
accuracy :  0.9995260238647461
precision :  0.9324324131011963
recall :  0.7582417726516724
auc :  0.9557874202728271

Legitimate Transactions Detected (True Negatives):  56866
Legitimate Transactions Incorrectly Detected (False Positives):  5
Fraudulent Transactions Missed (False Negatives):  22
Fraudulent Transactions Detected (True Positives):  69
Total Fraudulent Transactions:  91

png

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')

<matplotlib.legend.Legend at 0x7fa50c5adef0>

png

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.

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 [==============================] - 1s 15ms/step - loss: 2.5149 - tp: 105.0000 - fp: 66.0000 - tn: 238767.0000 - fn: 300.0000 - accuracy: 0.9985 - precision: 0.6140 - recall: 0.2593 - auc: 0.7803 - val_loss: 0.0067 - val_tp: 25.0000 - val_fp: 6.0000 - val_tn: 45476.0000 - val_fn: 62.0000 - val_accuracy: 0.9985 - val_precision: 0.8065 - val_recall: 0.2874 - val_auc: 0.9211
Epoch 2/100
90/90 [==============================] - 1s 6ms/step - loss: 1.2482 - tp: 145.0000 - fp: 124.0000 - tn: 181838.0000 - fn: 169.0000 - accuracy: 0.9984 - precision: 0.5390 - recall: 0.4618 - auc: 0.8560 - val_loss: 0.0062 - val_tp: 68.0000 - val_fp: 12.0000 - val_tn: 45470.0000 - val_fn: 19.0000 - val_accuracy: 0.9993 - val_precision: 0.8500 - val_recall: 0.7816 - val_auc: 0.9408
Epoch 3/100
90/90 [==============================] - 1s 6ms/step - loss: 0.8972 - tp: 177.0000 - fp: 237.0000 - tn: 181725.0000 - fn: 137.0000 - accuracy: 0.9979 - precision: 0.4275 - recall: 0.5637 - auc: 0.8876 - val_loss: 0.0079 - val_tp: 73.0000 - val_fp: 16.0000 - val_tn: 45466.0000 - val_fn: 14.0000 - val_accuracy: 0.9993 - val_precision: 0.8202 - val_recall: 0.8391 - val_auc: 0.9518
Epoch 4/100
90/90 [==============================] - 1s 6ms/step - loss: 0.6983 - tp: 210.0000 - fp: 387.0000 - tn: 181575.0000 - fn: 104.0000 - accuracy: 0.9973 - precision: 0.3518 - recall: 0.6688 - auc: 0.9028 - val_loss: 0.0098 - val_tp: 74.0000 - val_fp: 19.0000 - val_tn: 45463.0000 - val_fn: 13.0000 - val_accuracy: 0.9993 - val_precision: 0.7957 - val_recall: 0.8506 - val_auc: 0.9600
Epoch 5/100
90/90 [==============================] - 1s 6ms/step - loss: 0.6417 - tp: 220.0000 - fp: 583.0000 - tn: 181379.0000 - fn: 94.0000 - accuracy: 0.9963 - precision: 0.2740 - recall: 0.7006 - auc: 0.9084 - val_loss: 0.0119 - val_tp: 74.0000 - val_fp: 25.0000 - val_tn: 45457.0000 - val_fn: 13.0000 - val_accuracy: 0.9992 - val_precision: 0.7475 - val_recall: 0.8506 - val_auc: 0.9777
Epoch 6/100
90/90 [==============================] - 1s 6ms/step - loss: 0.5846 - tp: 232.0000 - fp: 977.0000 - tn: 180985.0000 - fn: 82.0000 - accuracy: 0.9942 - precision: 0.1919 - recall: 0.7389 - auc: 0.9048 - val_loss: 0.0148 - val_tp: 74.0000 - val_fp: 34.0000 - val_tn: 45448.0000 - val_fn: 13.0000 - val_accuracy: 0.9990 - val_precision: 0.6852 - val_recall: 0.8506 - val_auc: 0.9802
Epoch 7/100
90/90 [==============================] - 1s 6ms/step - loss: 0.5404 - tp: 234.0000 - fp: 1464.0000 - tn: 180498.0000 - fn: 80.0000 - accuracy: 0.9915 - precision: 0.1378 - recall: 0.7452 - auc: 0.9190 - val_loss: 0.0183 - val_tp: 74.0000 - val_fp: 50.0000 - val_tn: 45432.0000 - val_fn: 13.0000 - val_accuracy: 0.9986 - val_precision: 0.5968 - val_recall: 0.8506 - val_auc: 0.9823
Epoch 8/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4714 - tp: 241.0000 - fp: 1862.0000 - tn: 180100.0000 - fn: 73.0000 - accuracy: 0.9894 - precision: 0.1146 - recall: 0.7675 - auc: 0.9252 - val_loss: 0.0225 - val_tp: 76.0000 - val_fp: 84.0000 - val_tn: 45398.0000 - val_fn: 11.0000 - val_accuracy: 0.9979 - val_precision: 0.4750 - val_recall: 0.8736 - val_auc: 0.9851
Epoch 9/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4329 - tp: 247.0000 - fp: 2508.0000 - tn: 179454.0000 - fn: 67.0000 - accuracy: 0.9859 - precision: 0.0897 - recall: 0.7866 - auc: 0.9345 - val_loss: 0.0282 - val_tp: 76.0000 - val_fp: 170.0000 - val_tn: 45312.0000 - val_fn: 11.0000 - val_accuracy: 0.9960 - val_precision: 0.3089 - val_recall: 0.8736 - val_auc: 0.9873
Epoch 10/100
90/90 [==============================] - 1s 6ms/step - loss: 0.4467 - tp: 249.0000 - fp: 3175.0000 - tn: 178787.0000 - fn: 65.0000 - accuracy: 0.9822 - precision: 0.0727 - recall: 0.7930 - auc: 0.9210 - val_loss: 0.0341 - val_tp: 78.0000 - val_fp: 282.0000 - val_tn: 45200.0000 - val_fn: 9.0000 - val_accuracy: 0.9936 - val_precision: 0.2167 - val_recall: 0.8966 - val_auc: 0.9881
Epoch 11/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3947 - tp: 260.0000 - fp: 3569.0000 - tn: 178393.0000 - fn: 54.0000 - accuracy: 0.9801 - precision: 0.0679 - recall: 0.8280 - auc: 0.9290 - val_loss: 0.0394 - val_tp: 78.0000 - val_fp: 346.0000 - val_tn: 45136.0000 - val_fn: 9.0000 - val_accuracy: 0.9922 - val_precision: 0.1840 - val_recall: 0.8966 - val_auc: 0.9877
Epoch 12/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3694 - tp: 257.0000 - fp: 4294.0000 - tn: 177668.0000 - fn: 57.0000 - accuracy: 0.9761 - precision: 0.0565 - recall: 0.8185 - auc: 0.9418 - val_loss: 0.0473 - val_tp: 78.0000 - val_fp: 504.0000 - val_tn: 44978.0000 - val_fn: 9.0000 - val_accuracy: 0.9887 - val_precision: 0.1340 - val_recall: 0.8966 - val_auc: 0.9879
Epoch 13/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3479 - tp: 262.0000 - fp: 4886.0000 - tn: 177076.0000 - fn: 52.0000 - accuracy: 0.9729 - precision: 0.0509 - recall: 0.8344 - auc: 0.9403 - val_loss: 0.0539 - val_tp: 78.0000 - val_fp: 586.0000 - val_tn: 44896.0000 - val_fn: 9.0000 - val_accuracy: 0.9869 - val_precision: 0.1175 - val_recall: 0.8966 - val_auc: 0.9881
Epoch 14/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3653 - tp: 263.0000 - fp: 5360.0000 - tn: 176602.0000 - fn: 51.0000 - accuracy: 0.9703 - precision: 0.0468 - recall: 0.8376 - auc: 0.9370 - val_loss: 0.0610 - val_tp: 78.0000 - val_fp: 664.0000 - val_tn: 44818.0000 - val_fn: 9.0000 - val_accuracy: 0.9852 - val_precision: 0.1051 - val_recall: 0.8966 - val_auc: 0.9876
Epoch 15/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3673 - tp: 262.0000 - fp: 5820.0000 - tn: 176142.0000 - fn: 52.0000 - accuracy: 0.9678 - precision: 0.0431 - recall: 0.8344 - auc: 0.9316 - val_loss: 0.0658 - val_tp: 78.0000 - val_fp: 715.0000 - val_tn: 44767.0000 - val_fn: 9.0000 - val_accuracy: 0.9841 - val_precision: 0.0984 - val_recall: 0.8966 - val_auc: 0.9877
Epoch 16/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3228 - tp: 262.0000 - fp: 6230.0000 - tn: 175732.0000 - fn: 52.0000 - accuracy: 0.9655 - precision: 0.0404 - recall: 0.8344 - auc: 0.9445 - val_loss: 0.0716 - val_tp: 79.0000 - val_fp: 805.0000 - val_tn: 44677.0000 - val_fn: 8.0000 - val_accuracy: 0.9822 - val_precision: 0.0894 - val_recall: 0.9080 - val_auc: 0.9877
Epoch 17/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3299 - tp: 268.0000 - fp: 6572.0000 - tn: 175390.0000 - fn: 46.0000 - accuracy: 0.9637 - precision: 0.0392 - recall: 0.8535 - auc: 0.9423 - val_loss: 0.0757 - val_tp: 81.0000 - val_fp: 846.0000 - val_tn: 44636.0000 - val_fn: 6.0000 - val_accuracy: 0.9813 - val_precision: 0.0874 - val_recall: 0.9310 - val_auc: 0.9878
Epoch 18/100
90/90 [==============================] - 1s 6ms/step - loss: 0.2522 - tp: 276.0000 - fp: 6934.0000 - tn: 175028.0000 - fn: 38.0000 - accuracy: 0.9618 - precision: 0.0383 - recall: 0.8790 - auc: 0.9610 - val_loss: 0.0779 - val_tp: 81.0000 - val_fp: 874.0000 - val_tn: 44608.0000 - val_fn: 6.0000 - val_accuracy: 0.9807 - val_precision: 0.0848 - val_recall: 0.9310 - val_auc: 0.9877
Epoch 19/100
90/90 [==============================] - 1s 6ms/step - loss: 0.3607 - tp: 264.0000 - fp: 6790.0000 - tn: 175172.0000 - fn: 50.0000 - accuracy: 0.9625 - precision: 0.0374 - recall: 0.8408 - auc: 0.9303 - val_loss: 0.0781 - val_tp: 81.0000 - val_fp: 865.0000 - val_tn: 44617.0000 - val_fn: 6.0000 - val_accuracy: 0.9809 - val_precision: 0.0856 - val_recall: 0.9310 - val_auc: 0.9879
Epoch 20/100
89/90 [============================>.] - ETA: 0s - loss: 0.2977 - tp: 269.0000 - fp: 6769.0000 - tn: 175189.0000 - fn: 45.0000 - accuracy: 0.9626 - precision: 0.0382 - recall: 0.8567 - auc: 0.9488Restoring model weights from the end of the best epoch.
90/90 [==============================] - 1s 6ms/step - loss: 0.2977 - tp: 269.0000 - fp: 6769.0000 - tn: 175193.0000 - fn: 45.0000 - accuracy: 0.9626 - precision: 0.0382 - recall: 0.8567 - auc: 0.9488 - val_loss: 0.0780 - val_tp: 81.0000 - val_fp: 853.0000 - val_tn: 44629.0000 - val_fn: 6.0000 - val_accuracy: 0.9811 - val_precision: 0.0867 - val_recall: 0.9310 - val_auc: 0.9879
Epoch 00020: early stopping

Check training history

plot_metrics(weighted_history)

png

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)

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 :  0.03226418048143387
tp :  82.0
fp :  352.0
tn :  56519.0
fn :  9.0
accuracy :  0.993662416934967
precision :  0.18894009292125702
recall :  0.901098906993866
auc :  0.9671803712844849

Legitimate Transactions Detected (True Negatives):  56519
Legitimate Transactions Incorrectly Detected (False Positives):  352
Fraudulent Transactions Missed (False Negatives):  9
Fraudulent Transactions Detected (True Positives):  82
Total Fraudulent Transactions:  91

png

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')

<matplotlib.legend.Legend at 0x7fa54c0729e8>

png

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

(181962, 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

(363924, 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:
 [ 0.23104754  0.83661044 -0.31875356  1.9796369   1.28403692  0.07389102
  1.03350673 -0.11568355 -1.54396817  0.88004244 -1.66944551 -0.24324391
  0.45900013  0.14583622 -2.06637388  0.42470592 -0.94489216 -0.83112221
 -1.83416278 -0.34138858  0.14130878  0.51019975  0.08224586  0.6642136
 -1.39031637 -0.42194185  0.22525572  0.28277796 -4.86369823]

Label:  1

Merge the two together using experimental.sample_from_datasets:

resampled_ds = tf.data.experimental.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())

0.49609375

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.

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 [==============================] - 6s 23ms/step - loss: 0.4356 - tp: 223484.0000 - fp: 51288.0000 - tn: 290777.0000 - fn: 60757.0000 - accuracy: 0.8211 - precision: 0.8133 - recall: 0.7862 - auc: 0.8933 - val_loss: 0.2172 - val_tp: 79.0000 - val_fp: 1076.0000 - val_tn: 44406.0000 - val_fn: 8.0000 - val_accuracy: 0.9762 - val_precision: 0.0684 - val_recall: 0.9080 - val_auc: 0.9792
Epoch 2/100
278/278 [==============================] - 6s 20ms/step - loss: 0.2177 - tp: 246785.0000 - fp: 12557.0000 - tn: 271871.0000 - fn: 38131.0000 - accuracy: 0.9110 - precision: 0.9516 - recall: 0.8662 - auc: 0.9686 - val_loss: 0.1226 - val_tp: 80.0000 - val_fp: 951.0000 - val_tn: 44531.0000 - val_fn: 7.0000 - val_accuracy: 0.9790 - val_precision: 0.0776 - val_recall: 0.9195 - val_auc: 0.9835
Epoch 3/100
278/278 [==============================] - 6s 21ms/step - loss: 0.1751 - tp: 250631.0000 - fp: 9797.0000 - tn: 275174.0000 - fn: 33742.0000 - accuracy: 0.9235 - precision: 0.9624 - recall: 0.8813 - auc: 0.9810 - val_loss: 0.0940 - val_tp: 82.0000 - val_fp: 966.0000 - val_tn: 44516.0000 - val_fn: 5.0000 - val_accuracy: 0.9787 - val_precision: 0.0782 - val_recall: 0.9425 - val_auc: 0.9836
Epoch 4/100
278/278 [==============================] - 6s 22ms/step - loss: 0.1532 - tp: 254169.0000 - fp: 9171.0000 - tn: 275694.0000 - fn: 30310.0000 - accuracy: 0.9307 - precision: 0.9652 - recall: 0.8935 - auc: 0.9861 - val_loss: 0.0802 - val_tp: 82.0000 - val_fp: 918.0000 - val_tn: 44564.0000 - val_fn: 5.0000 - val_accuracy: 0.9797 - val_precision: 0.0820 - val_recall: 0.9425 - val_auc: 0.9847
Epoch 5/100
278/278 [==============================] - 6s 22ms/step - loss: 0.1372 - tp: 257034.0000 - fp: 9061.0000 - tn: 275758.0000 - fn: 27491.0000 - accuracy: 0.9358 - precision: 0.9659 - recall: 0.9034 - auc: 0.9892 - val_loss: 0.0720 - val_tp: 82.0000 - val_fp: 910.0000 - val_tn: 44572.0000 - val_fn: 5.0000 - val_accuracy: 0.9799 - val_precision: 0.0827 - val_recall: 0.9425 - val_auc: 0.9854
Epoch 6/100
278/278 [==============================] - 6s 22ms/step - loss: 0.1260 - tp: 258997.0000 - fp: 9079.0000 - tn: 275819.0000 - fn: 25449.0000 - accuracy: 0.9394 - precision: 0.9661 - recall: 0.9105 - auc: 0.9911 - val_loss: 0.0666 - val_tp: 81.0000 - val_fp: 915.0000 - val_tn: 44567.0000 - val_fn: 6.0000 - val_accuracy: 0.9798 - val_precision: 0.0813 - val_recall: 0.9310 - val_auc: 0.9856
Epoch 7/100
278/278 [==============================] - 6s 21ms/step - loss: 0.1167 - tp: 261100.0000 - fp: 9112.0000 - tn: 276180.0000 - fn: 22952.0000 - accuracy: 0.9437 - precision: 0.9663 - recall: 0.9192 - auc: 0.9925 - val_loss: 0.0623 - val_tp: 81.0000 - val_fp: 911.0000 - val_tn: 44571.0000 - val_fn: 6.0000 - val_accuracy: 0.9799 - val_precision: 0.0817 - val_recall: 0.9310 - val_auc: 0.9858
Epoch 8/100
278/278 [==============================] - 6s 22ms/step - loss: 0.1082 - tp: 263945.0000 - fp: 9428.0000 - tn: 275276.0000 - fn: 20695.0000 - accuracy: 0.9471 - precision: 0.9655 - recall: 0.9273 - auc: 0.9937 - val_loss: 0.0587 - val_tp: 81.0000 - val_fp: 910.0000 - val_tn: 44572.0000 - val_fn: 6.0000 - val_accuracy: 0.9799 - val_precision: 0.0817 - val_recall: 0.9310 - val_auc: 0.9857
Epoch 9/100
278/278 [==============================] - 6s 21ms/step - loss: 0.1014 - tp: 268108.0000 - fp: 10376.0000 - tn: 274312.0000 - fn: 16548.0000 - accuracy: 0.9527 - precision: 0.9627 - recall: 0.9419 - auc: 0.9944 - val_loss: 0.0543 - val_tp: 80.0000 - val_fp: 873.0000 - val_tn: 44609.0000 - val_fn: 7.0000 - val_accuracy: 0.9807 - val_precision: 0.0839 - val_recall: 0.9195 - val_auc: 0.9857
Epoch 10/100
278/278 [==============================] - 6s 22ms/step - loss: 0.0951 - tp: 277520.0000 - fp: 12692.0000 - tn: 271795.0000 - fn: 7337.0000 - accuracy: 0.9648 - precision: 0.9563 - recall: 0.9742 - auc: 0.9950 - val_loss: 0.0495 - val_tp: 79.0000 - val_fp: 829.0000 - val_tn: 44653.0000 - val_fn: 8.0000 - val_accuracy: 0.9816 - val_precision: 0.0870 - val_recall: 0.9080 - val_auc: 0.9855
Epoch 11/100
278/278 [==============================] - 6s 21ms/step - loss: 0.0895 - tp: 278865.0000 - fp: 12938.0000 - tn: 271719.0000 - fn: 5822.0000 - accuracy: 0.9670 - precision: 0.9557 - recall: 0.9795 - auc: 0.9955 - val_loss: 0.0450 - val_tp: 79.0000 - val_fp: 789.0000 - val_tn: 44693.0000 - val_fn: 8.0000 - val_accuracy: 0.9825 - val_precision: 0.0910 - val_recall: 0.9080 - val_auc: 0.9859
Epoch 12/100
278/278 [==============================] - 6s 21ms/step - loss: 0.0842 - tp: 279845.0000 - fp: 13187.0000 - tn: 272121.0000 - fn: 4191.0000 - accuracy: 0.9695 - precision: 0.9550 - recall: 0.9852 - auc: 0.9960 - val_loss: 0.0410 - val_tp: 79.0000 - val_fp: 733.0000 - val_tn: 44749.0000 - val_fn: 8.0000 - val_accuracy: 0.9837 - val_precision: 0.0973 - val_recall: 0.9080 - val_auc: 0.9813
Epoch 13/100
278/278 [==============================] - 6s 22ms/step - loss: 0.0792 - tp: 281765.0000 - fp: 12977.0000 - tn: 271393.0000 - fn: 3209.0000 - accuracy: 0.9716 - precision: 0.9560 - recall: 0.9887 - auc: 0.9963 - val_loss: 0.0389 - val_tp: 79.0000 - val_fp: 721.0000 - val_tn: 44761.0000 - val_fn: 8.0000 - val_accuracy: 0.9840 - val_precision: 0.0988 - val_recall: 0.9080 - val_auc: 0.9814
Epoch 14/100
278/278 [==============================] - 6s 21ms/step - loss: 0.0754 - tp: 281962.0000 - fp: 13026.0000 - tn: 272154.0000 - fn: 2202.0000 - accuracy: 0.9733 - precision: 0.9558 - recall: 0.9923 - auc: 0.9966 - val_loss: 0.0348 - val_tp: 79.0000 - val_fp: 646.0000 - val_tn: 44836.0000 - val_fn: 8.0000 - val_accuracy: 0.9856 - val_precision: 0.1090 - val_recall: 0.9080 - val_auc: 0.9763
Epoch 15/100
278/278 [==============================] - 6s 23ms/step - loss: 0.0722 - tp: 283858.0000 - fp: 12932.0000 - tn: 271419.0000 - fn: 1135.0000 - accuracy: 0.9753 - precision: 0.9564 - recall: 0.9960 - auc: 0.9967 - val_loss: 0.0331 - val_tp: 79.0000 - val_fp: 640.0000 - val_tn: 44842.0000 - val_fn: 8.0000 - val_accuracy: 0.9858 - val_precision: 0.1099 - val_recall: 0.9080 - val_auc: 0.9714
Epoch 16/100
278/278 [==============================] - 6s 22ms/step - loss: 0.0689 - tp: 283059.0000 - fp: 12757.0000 - tn: 273004.0000 - fn: 524.0000 - accuracy: 0.9767 - precision: 0.9569 - recall: 0.9982 - auc: 0.9970 - val_loss: 0.0308 - val_tp: 79.0000 - val_fp: 583.0000 - val_tn: 44899.0000 - val_fn: 8.0000 - val_accuracy: 0.9870 - val_precision: 0.1193 - val_recall: 0.9080 - val_auc: 0.9667
Epoch 17/100
278/278 [==============================] - 6s 23ms/step - loss: 0.0661 - tp: 283879.0000 - fp: 12340.0000 - tn: 272779.0000 - fn: 346.0000 - accuracy: 0.9777 - precision: 0.9583 - recall: 0.9988 - auc: 0.9971 - val_loss: 0.0289 - val_tp: 79.0000 - val_fp: 542.0000 - val_tn: 44940.0000 - val_fn: 8.0000 - val_accuracy: 0.9879 - val_precision: 0.1272 - val_recall: 0.9080 - val_auc: 0.9618
Epoch 18/100
278/278 [==============================] - 6s 22ms/step - loss: 0.0635 - tp: 284858.0000 - fp: 12157.0000 - tn: 272120.0000 - fn: 209.0000 - accuracy: 0.9783 - precision: 0.9591 - recall: 0.9993 - auc: 0.9973 - val_loss: 0.0277 - val_tp: 79.0000 - val_fp: 511.0000 - val_tn: 44971.0000 - val_fn: 8.0000 - val_accuracy: 0.9886 - val_precision: 0.1339 - val_recall: 0.9080 - val_auc: 0.9621
Epoch 19/100
278/278 [==============================] - 6s 23ms/step - loss: 0.0620 - tp: 284459.0000 - fp: 11978.0000 - tn: 272718.0000 - fn: 189.0000 - accuracy: 0.9786 - precision: 0.9596 - recall: 0.9993 - auc: 0.9973 - val_loss: 0.0261 - val_tp: 79.0000 - val_fp: 478.0000 - val_tn: 45004.0000 - val_fn: 8.0000 - val_accuracy: 0.9893 - val_precision: 0.1418 - val_recall: 0.9080 - val_auc: 0.9624
Epoch 20/100
278/278 [==============================] - 6s 23ms/step - loss: 0.0600 - tp: 284950.0000 - fp: 11793.0000 - tn: 272572.0000 - fn: 29.0000 - accuracy: 0.9792 - precision: 0.9603 - recall: 0.9999 - auc: 0.9974 - val_loss: 0.0252 - val_tp: 79.0000 - val_fp: 463.0000 - val_tn: 45019.0000 - val_fn: 8.0000 - val_accuracy: 0.9897 - val_precision: 0.1458 - val_recall: 0.9080 - val_auc: 0.9626
Epoch 21/100
276/278 [============================>.] - ETA: 0s - loss: 0.0581 - tp: 282210.0000 - fp: 11270.0000 - tn: 271768.0000 - fn: 0.0000e+00 - accuracy: 0.9801 - precision: 0.9616 - recall: 1.0000 - auc: 0.9975Restoring model weights from the end of the best epoch.
278/278 [==============================] - 6s 22ms/step - loss: 0.0581 - tp: 284274.0000 - fp: 11360.0000 - tn: 273710.0000 - fn: 0.0000e+00 - accuracy: 0.9800 - precision: 0.9616 - recall: 1.0000 - auc: 0.9975 - val_loss: 0.0241 - val_tp: 79.0000 - val_fp: 444.0000 - val_tn: 45038.0000 - val_fn: 8.0000 - val_accuracy: 0.9901 - val_precision: 0.1511 - val_recall: 0.9080 - val_auc: 0.9628
Epoch 00021: 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 )

png

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 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 [==============================] - 1s 60ms/step - loss: 1.0656 - tp: 9507.0000 - fp: 7370.0000 - tn: 58667.0000 - fn: 10985.0000 - accuracy: 0.7879 - precision: 0.5633 - recall: 0.4639 - auc: 0.8255 - val_loss: 0.5792 - val_tp: 66.0000 - val_fp: 13452.0000 - val_tn: 32030.0000 - val_fn: 21.0000 - val_accuracy: 0.7043 - val_precision: 0.0049 - val_recall: 0.7586 - val_auc: 0.7866
Epoch 2/1000
20/20 [==============================] - 1s 26ms/step - loss: 0.6996 - tp: 13383.0000 - fp: 7208.0000 - tn: 13397.0000 - fn: 6972.0000 - accuracy: 0.6538 - precision: 0.6499 - recall: 0.6575 - auc: 0.7027 - val_loss: 0.5702 - val_tp: 76.0000 - val_fp: 12408.0000 - val_tn: 33074.0000 - val_fn: 11.0000 - val_accuracy: 0.7275 - val_precision: 0.0061 - val_recall: 0.8736 - val_auc: 0.9076
Epoch 3/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.5532 - tp: 15127.0000 - fp: 6665.0000 - tn: 14055.0000 - fn: 5113.0000 - accuracy: 0.7125 - precision: 0.6942 - recall: 0.7474 - auc: 0.7952 - val_loss: 0.5335 - val_tp: 79.0000 - val_fp: 9006.0000 - val_tn: 36476.0000 - val_fn: 8.0000 - val_accuracy: 0.8022 - val_precision: 0.0087 - val_recall: 0.9080 - val_auc: 0.9408
Epoch 4/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.4738 - tp: 16061.0000 - fp: 5669.0000 - tn: 14890.0000 - fn: 4340.0000 - accuracy: 0.7556 - precision: 0.7391 - recall: 0.7873 - auc: 0.8495 - val_loss: 0.4883 - val_tp: 78.0000 - val_fp: 5756.0000 - val_tn: 39726.0000 - val_fn: 9.0000 - val_accuracy: 0.8735 - val_precision: 0.0134 - val_recall: 0.8966 - val_auc: 0.9489
Epoch 5/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.4266 - tp: 16612.0000 - fp: 4719.0000 - tn: 15715.0000 - fn: 3914.0000 - accuracy: 0.7892 - precision: 0.7788 - recall: 0.8093 - auc: 0.8786 - val_loss: 0.4435 - val_tp: 78.0000 - val_fp: 3758.0000 - val_tn: 41724.0000 - val_fn: 9.0000 - val_accuracy: 0.9173 - val_precision: 0.0203 - val_recall: 0.8966 - val_auc: 0.9539
Epoch 6/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.3908 - tp: 16911.0000 - fp: 3861.0000 - tn: 16514.0000 - fn: 3674.0000 - accuracy: 0.8160 - precision: 0.8141 - recall: 0.8215 - auc: 0.8976 - val_loss: 0.4032 - val_tp: 79.0000 - val_fp: 2770.0000 - val_tn: 42712.0000 - val_fn: 8.0000 - val_accuracy: 0.9390 - val_precision: 0.0277 - val_recall: 0.9080 - val_auc: 0.9590
Epoch 7/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.3664 - tp: 17049.0000 - fp: 3209.0000 - tn: 17179.0000 - fn: 3523.0000 - accuracy: 0.8356 - precision: 0.8416 - recall: 0.8287 - auc: 0.9108 - val_loss: 0.3682 - val_tp: 79.0000 - val_fp: 2119.0000 - val_tn: 43363.0000 - val_fn: 8.0000 - val_accuracy: 0.9533 - val_precision: 0.0359 - val_recall: 0.9080 - val_auc: 0.9634
Epoch 8/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.3467 - tp: 17100.0000 - fp: 2699.0000 - tn: 17686.0000 - fn: 3475.0000 - accuracy: 0.8493 - precision: 0.8637 - recall: 0.8311 - auc: 0.9193 - val_loss: 0.3373 - val_tp: 79.0000 - val_fp: 1753.0000 - val_tn: 43729.0000 - val_fn: 8.0000 - val_accuracy: 0.9614 - val_precision: 0.0431 - val_recall: 0.9080 - val_auc: 0.9675
Epoch 9/1000
20/20 [==============================] - 1s 29ms/step - loss: 0.3285 - tp: 17043.0000 - fp: 2345.0000 - tn: 18228.0000 - fn: 3344.0000 - accuracy: 0.8611 - precision: 0.8790 - recall: 0.8360 - auc: 0.9271 - val_loss: 0.3104 - val_tp: 79.0000 - val_fp: 1495.0000 - val_tn: 43987.0000 - val_fn: 8.0000 - val_accuracy: 0.9670 - val_precision: 0.0502 - val_recall: 0.9080 - val_auc: 0.9702
Epoch 10/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.3094 - tp: 17322.0000 - fp: 2012.0000 - tn: 18405.0000 - fn: 3221.0000 - accuracy: 0.8722 - precision: 0.8959 - recall: 0.8432 - auc: 0.9361 - val_loss: 0.2865 - val_tp: 79.0000 - val_fp: 1332.0000 - val_tn: 44150.0000 - val_fn: 8.0000 - val_accuracy: 0.9706 - val_precision: 0.0560 - val_recall: 0.9080 - val_auc: 0.9721
Epoch 11/1000
20/20 [==============================] - 1s 29ms/step - loss: 0.2962 - tp: 17184.0000 - fp: 1757.0000 - tn: 18853.0000 - fn: 3166.0000 - accuracy: 0.8798 - precision: 0.9072 - recall: 0.8444 - auc: 0.9406 - val_loss: 0.2654 - val_tp: 79.0000 - val_fp: 1228.0000 - val_tn: 44254.0000 - val_fn: 8.0000 - val_accuracy: 0.9729 - val_precision: 0.0604 - val_recall: 0.9080 - val_auc: 0.9739
Epoch 12/1000
20/20 [==============================] - 1s 30ms/step - loss: 0.2835 - tp: 17373.0000 - fp: 1543.0000 - tn: 18909.0000 - fn: 3135.0000 - accuracy: 0.8858 - precision: 0.9184 - recall: 0.8471 - auc: 0.9458 - val_loss: 0.2469 - val_tp: 79.0000 - val_fp: 1155.0000 - val_tn: 44327.0000 - val_fn: 8.0000 - val_accuracy: 0.9745 - val_precision: 0.0640 - val_recall: 0.9080 - val_auc: 0.9759
Epoch 13/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.2710 - tp: 17386.0000 - fp: 1395.0000 - tn: 19124.0000 - fn: 3055.0000 - accuracy: 0.8914 - precision: 0.9257 - recall: 0.8505 - auc: 0.9502 - val_loss: 0.2302 - val_tp: 79.0000 - val_fp: 1092.0000 - val_tn: 44390.0000 - val_fn: 8.0000 - val_accuracy: 0.9759 - val_precision: 0.0675 - val_recall: 0.9080 - val_auc: 0.9782
Epoch 14/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2618 - tp: 17336.0000 - fp: 1343.0000 - tn: 19296.0000 - fn: 2985.0000 - accuracy: 0.8943 - precision: 0.9281 - recall: 0.8531 - auc: 0.9541 - val_loss: 0.2156 - val_tp: 79.0000 - val_fp: 1053.0000 - val_tn: 44429.0000 - val_fn: 8.0000 - val_accuracy: 0.9767 - val_precision: 0.0698 - val_recall: 0.9080 - val_auc: 0.9797
Epoch 15/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2529 - tp: 17466.0000 - fp: 1154.0000 - tn: 19366.0000 - fn: 2974.0000 - accuracy: 0.8992 - precision: 0.9380 - recall: 0.8545 - auc: 0.9574 - val_loss: 0.2026 - val_tp: 79.0000 - val_fp: 1029.0000 - val_tn: 44453.0000 - val_fn: 8.0000 - val_accuracy: 0.9772 - val_precision: 0.0713 - val_recall: 0.9080 - val_auc: 0.9806
Epoch 16/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2456 - tp: 17579.0000 - fp: 1075.0000 - tn: 19322.0000 - fn: 2984.0000 - accuracy: 0.9009 - precision: 0.9424 - recall: 0.8549 - auc: 0.9590 - val_loss: 0.1923 - val_tp: 79.0000 - val_fp: 1017.0000 - val_tn: 44465.0000 - val_fn: 8.0000 - val_accuracy: 0.9775 - val_precision: 0.0721 - val_recall: 0.9080 - val_auc: 0.9813
Epoch 17/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.2382 - tp: 17573.0000 - fp: 982.0000 - tn: 19540.0000 - fn: 2865.0000 - accuracy: 0.9061 - precision: 0.9471 - recall: 0.8598 - auc: 0.9620 - val_loss: 0.1828 - val_tp: 79.0000 - val_fp: 1005.0000 - val_tn: 44477.0000 - val_fn: 8.0000 - val_accuracy: 0.9778 - val_precision: 0.0729 - val_recall: 0.9080 - val_auc: 0.9819
Epoch 18/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.2307 - tp: 17711.0000 - fp: 966.0000 - tn: 19448.0000 - fn: 2835.0000 - accuracy: 0.9072 - precision: 0.9483 - recall: 0.8620 - auc: 0.9644 - val_loss: 0.1736 - val_tp: 80.0000 - val_fp: 990.0000 - val_tn: 44492.0000 - val_fn: 7.0000 - val_accuracy: 0.9781 - val_precision: 0.0748 - val_recall: 0.9195 - val_auc: 0.9825
Epoch 19/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.2280 - tp: 17732.0000 - fp: 952.0000 - tn: 19442.0000 - fn: 2834.0000 - accuracy: 0.9076 - precision: 0.9490 - recall: 0.8622 - auc: 0.9653 - val_loss: 0.1660 - val_tp: 80.0000 - val_fp: 974.0000 - val_tn: 44508.0000 - val_fn: 7.0000 - val_accuracy: 0.9785 - val_precision: 0.0759 - val_recall: 0.9195 - val_auc: 0.9826
Epoch 20/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.2224 - tp: 17725.0000 - fp: 939.0000 - tn: 19538.0000 - fn: 2758.0000 - accuracy: 0.9097 - precision: 0.9497 - recall: 0.8654 - auc: 0.9667 - val_loss: 0.1591 - val_tp: 80.0000 - val_fp: 962.0000 - val_tn: 44520.0000 - val_fn: 7.0000 - val_accuracy: 0.9787 - val_precision: 0.0768 - val_recall: 0.9195 - val_auc: 0.9831
Epoch 21/1000
20/20 [==============================] - 1s 29ms/step - loss: 0.2168 - tp: 17757.0000 - fp: 826.0000 - tn: 19618.0000 - fn: 2759.0000 - accuracy: 0.9125 - precision: 0.9556 - recall: 0.8655 - auc: 0.9689 - val_loss: 0.1531 - val_tp: 80.0000 - val_fp: 967.0000 - val_tn: 44515.0000 - val_fn: 7.0000 - val_accuracy: 0.9786 - val_precision: 0.0764 - val_recall: 0.9195 - val_auc: 0.9831
Epoch 22/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.2112 - tp: 17833.0000 - fp: 883.0000 - tn: 19522.0000 - fn: 2722.0000 - accuracy: 0.9120 - precision: 0.9528 - recall: 0.8676 - auc: 0.9703 - val_loss: 0.1479 - val_tp: 80.0000 - val_fp: 975.0000 - val_tn: 44507.0000 - val_fn: 7.0000 - val_accuracy: 0.9785 - val_precision: 0.0758 - val_recall: 0.9195 - val_auc: 0.9832
Epoch 23/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.2058 - tp: 17865.0000 - fp: 835.0000 - tn: 19580.0000 - fn: 2680.0000 - accuracy: 0.9142 - precision: 0.9553 - recall: 0.8696 - auc: 0.9723 - val_loss: 0.1427 - val_tp: 80.0000 - val_fp: 977.0000 - val_tn: 44505.0000 - val_fn: 7.0000 - val_accuracy: 0.9784 - val_precision: 0.0757 - val_recall: 0.9195 - val_auc: 0.9834
Epoch 24/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.2053 - tp: 17856.0000 - fp: 802.0000 - tn: 19599.0000 - fn: 2703.0000 - accuracy: 0.9144 - precision: 0.9570 - recall: 0.8685 - auc: 0.9727 - val_loss: 0.1375 - val_tp: 80.0000 - val_fp: 969.0000 - val_tn: 44513.0000 - val_fn: 7.0000 - val_accuracy: 0.9786 - val_precision: 0.0763 - val_recall: 0.9195 - val_auc: 0.9833
Epoch 25/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.2004 - tp: 17854.0000 - fp: 809.0000 - tn: 19690.0000 - fn: 2607.0000 - accuracy: 0.9166 - precision: 0.9567 - recall: 0.8726 - auc: 0.9740 - val_loss: 0.1331 - val_tp: 80.0000 - val_fp: 976.0000 - val_tn: 44506.0000 - val_fn: 7.0000 - val_accuracy: 0.9784 - val_precision: 0.0758 - val_recall: 0.9195 - val_auc: 0.9837
Epoch 26/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1991 - tp: 17857.0000 - fp: 793.0000 - tn: 19690.0000 - fn: 2620.0000 - accuracy: 0.9167 - precision: 0.9575 - recall: 0.8721 - auc: 0.9747 - val_loss: 0.1291 - val_tp: 80.0000 - val_fp: 968.0000 - val_tn: 44514.0000 - val_fn: 7.0000 - val_accuracy: 0.9786 - val_precision: 0.0763 - val_recall: 0.9195 - val_auc: 0.9836
Epoch 27/1000
20/20 [==============================] - 1s 40ms/step - loss: 0.1929 - tp: 17836.0000 - fp: 750.0000 - tn: 19833.0000 - fn: 2541.0000 - accuracy: 0.9197 - precision: 0.9596 - recall: 0.8753 - auc: 0.9760 - val_loss: 0.1252 - val_tp: 80.0000 - val_fp: 960.0000 - val_tn: 44522.0000 - val_fn: 7.0000 - val_accuracy: 0.9788 - val_precision: 0.0769 - val_recall: 0.9195 - val_auc: 0.9839
Epoch 28/1000
20/20 [==============================] - 1s 29ms/step - loss: 0.1935 - tp: 17776.0000 - fp: 753.0000 - tn: 19827.0000 - fn: 2604.0000 - accuracy: 0.9180 - precision: 0.9594 - recall: 0.8722 - auc: 0.9763 - val_loss: 0.1215 - val_tp: 80.0000 - val_fp: 946.0000 - val_tn: 44536.0000 - val_fn: 7.0000 - val_accuracy: 0.9791 - val_precision: 0.0780 - val_recall: 0.9195 - val_auc: 0.9836
Epoch 29/1000
20/20 [==============================] - 1s 32ms/step - loss: 0.1892 - tp: 17877.0000 - fp: 746.0000 - tn: 19791.0000 - fn: 2546.0000 - accuracy: 0.9196 - precision: 0.9599 - recall: 0.8753 - auc: 0.9773 - val_loss: 0.1183 - val_tp: 80.0000 - val_fp: 944.0000 - val_tn: 44538.0000 - val_fn: 7.0000 - val_accuracy: 0.9791 - val_precision: 0.0781 - val_recall: 0.9195 - val_auc: 0.9840
Epoch 30/1000
20/20 [==============================] - 1s 30ms/step - loss: 0.1855 - tp: 18053.0000 - fp: 746.0000 - tn: 19673.0000 - fn: 2488.0000 - accuracy: 0.9210 - precision: 0.9603 - recall: 0.8789 - auc: 0.9779 - val_loss: 0.1157 - val_tp: 80.0000 - val_fp: 949.0000 - val_tn: 44533.0000 - val_fn: 7.0000 - val_accuracy: 0.9790 - val_precision: 0.0777 - val_recall: 0.9195 - val_auc: 0.9835
Epoch 31/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.1843 - tp: 18042.0000 - fp: 723.0000 - tn: 19656.0000 - fn: 2539.0000 - accuracy: 0.9204 - precision: 0.9615 - recall: 0.8766 - auc: 0.9783 - val_loss: 0.1137 - val_tp: 80.0000 - val_fp: 958.0000 - val_tn: 44524.0000 - val_fn: 7.0000 - val_accuracy: 0.9788 - val_precision: 0.0771 - val_recall: 0.9195 - val_auc: 0.9836
Epoch 32/1000
20/20 [==============================] - 1s 26ms/step - loss: 0.1831 - tp: 17974.0000 - fp: 743.0000 - tn: 19741.0000 - fn: 2502.0000 - accuracy: 0.9208 - precision: 0.9603 - recall: 0.8778 - auc: 0.9789 - val_loss: 0.1112 - val_tp: 80.0000 - val_fp: 958.0000 - val_tn: 44524.0000 - val_fn: 7.0000 - val_accuracy: 0.9788 - val_precision: 0.0771 - val_recall: 0.9195 - val_auc: 0.9840
Epoch 33/1000
20/20 [==============================] - 1s 26ms/step - loss: 0.1805 - tp: 18172.0000 - fp: 775.0000 - tn: 19591.0000 - fn: 2422.0000 - accuracy: 0.9219 - precision: 0.9591 - recall: 0.8824 - auc: 0.9796 - val_loss: 0.1088 - val_tp: 81.0000 - val_fp: 956.0000 - val_tn: 44526.0000 - val_fn: 6.0000 - val_accuracy: 0.9789 - val_precision: 0.0781 - val_recall: 0.9310 - val_auc: 0.9841
Epoch 34/1000
20/20 [==============================] - 0s 24ms/step - loss: 0.1749 - tp: 18125.0000 - fp: 715.0000 - tn: 19698.0000 - fn: 2422.0000 - accuracy: 0.9234 - precision: 0.9620 - recall: 0.8821 - auc: 0.9812 - val_loss: 0.1068 - val_tp: 81.0000 - val_fp: 964.0000 - val_tn: 44518.0000 - val_fn: 6.0000 - val_accuracy: 0.9787 - val_precision: 0.0775 - val_recall: 0.9310 - val_auc: 0.9836
Epoch 35/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1769 - tp: 18135.0000 - fp: 715.0000 - tn: 19694.0000 - fn: 2416.0000 - accuracy: 0.9236 - precision: 0.9621 - recall: 0.8824 - auc: 0.9809 - val_loss: 0.1048 - val_tp: 81.0000 - val_fp: 978.0000 - val_tn: 44504.0000 - val_fn: 6.0000 - val_accuracy: 0.9784 - val_precision: 0.0765 - val_recall: 0.9310 - val_auc: 0.9838
Epoch 36/1000
20/20 [==============================] - 1s 30ms/step - loss: 0.1739 - tp: 18006.0000 - fp: 704.0000 - tn: 19827.0000 - fn: 2423.0000 - accuracy: 0.9237 - precision: 0.9624 - recall: 0.8814 - auc: 0.9814 - val_loss: 0.1029 - val_tp: 81.0000 - val_fp: 986.0000 - val_tn: 44496.0000 - val_fn: 6.0000 - val_accuracy: 0.9782 - val_precision: 0.0759 - val_recall: 0.9310 - val_auc: 0.9839
Epoch 37/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.1687 - tp: 18002.0000 - fp: 660.0000 - tn: 19879.0000 - fn: 2419.0000 - accuracy: 0.9248 - precision: 0.9646 - recall: 0.8815 - auc: 0.9826 - val_loss: 0.1011 - val_tp: 81.0000 - val_fp: 984.0000 - val_tn: 44498.0000 - val_fn: 6.0000 - val_accuracy: 0.9783 - val_precision: 0.0761 - val_recall: 0.9310 - val_auc: 0.9841
Epoch 38/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.1699 - tp: 17932.0000 - fp: 677.0000 - tn: 19986.0000 - fn: 2365.0000 - accuracy: 0.9257 - precision: 0.9636 - recall: 0.8835 - auc: 0.9825 - val_loss: 0.0995 - val_tp: 82.0000 - val_fp: 979.0000 - val_tn: 44503.0000 - val_fn: 5.0000 - val_accuracy: 0.9784 - val_precision: 0.0773 - val_recall: 0.9425 - val_auc: 0.9842
Epoch 39/1000
20/20 [==============================] - 1s 30ms/step - loss: 0.1676 - tp: 18086.0000 - fp: 736.0000 - tn: 19780.0000 - fn: 2358.0000 - accuracy: 0.9245 - precision: 0.9609 - recall: 0.8847 - auc: 0.9826 - val_loss: 0.0980 - val_tp: 82.0000 - val_fp: 975.0000 - val_tn: 44507.0000 - val_fn: 5.0000 - val_accuracy: 0.9785 - val_precision: 0.0776 - val_recall: 0.9425 - val_auc: 0.9844
Epoch 40/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.1670 - tp: 18066.0000 - fp: 685.0000 - tn: 19868.0000 - fn: 2341.0000 - accuracy: 0.9261 - precision: 0.9635 - recall: 0.8853 - auc: 0.9832 - val_loss: 0.0964 - val_tp: 82.0000 - val_fp: 965.0000 - val_tn: 44517.0000 - val_fn: 5.0000 - val_accuracy: 0.9787 - val_precision: 0.0783 - val_recall: 0.9425 - val_auc: 0.9845
Epoch 41/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1640 - tp: 17950.0000 - fp: 645.0000 - tn: 19995.0000 - fn: 2370.0000 - accuracy: 0.9264 - precision: 0.9653 - recall: 0.8834 - auc: 0.9839 - val_loss: 0.0950 - val_tp: 82.0000 - val_fp: 956.0000 - val_tn: 44526.0000 - val_fn: 5.0000 - val_accuracy: 0.9789 - val_precision: 0.0790 - val_recall: 0.9425 - val_auc: 0.9835
Epoch 42/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1641 - tp: 18083.0000 - fp: 665.0000 - tn: 19842.0000 - fn: 2370.0000 - accuracy: 0.9259 - precision: 0.9645 - recall: 0.8841 - auc: 0.9839 - val_loss: 0.0938 - val_tp: 82.0000 - val_fp: 949.0000 - val_tn: 44533.0000 - val_fn: 5.0000 - val_accuracy: 0.9791 - val_precision: 0.0795 - val_recall: 0.9425 - val_auc: 0.9837
Epoch 43/1000
20/20 [==============================] - 0s 23ms/step - loss: 0.1600 - tp: 18012.0000 - fp: 684.0000 - tn: 19970.0000 - fn: 2294.0000 - accuracy: 0.9273 - precision: 0.9634 - recall: 0.8870 - auc: 0.9845 - val_loss: 0.0925 - val_tp: 82.0000 - val_fp: 949.0000 - val_tn: 44533.0000 - val_fn: 5.0000 - val_accuracy: 0.9791 - val_precision: 0.0795 - val_recall: 0.9425 - val_auc: 0.9837
Epoch 44/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.1597 - tp: 18346.0000 - fp: 657.0000 - tn: 19657.0000 - fn: 2300.0000 - accuracy: 0.9278 - precision: 0.9654 - recall: 0.8886 - auc: 0.9847 - val_loss: 0.0919 - val_tp: 82.0000 - val_fp: 955.0000 - val_tn: 44527.0000 - val_fn: 5.0000 - val_accuracy: 0.9789 - val_precision: 0.0791 - val_recall: 0.9425 - val_auc: 0.9838
Epoch 45/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.1607 - tp: 18109.0000 - fp: 726.0000 - tn: 19836.0000 - fn: 2289.0000 - accuracy: 0.9264 - precision: 0.9615 - recall: 0.8878 - auc: 0.9846 - val_loss: 0.0908 - val_tp: 82.0000 - val_fp: 948.0000 - val_tn: 44534.0000 - val_fn: 5.0000 - val_accuracy: 0.9791 - val_precision: 0.0796 - val_recall: 0.9425 - val_auc: 0.9839
Epoch 46/1000
20/20 [==============================] - 1s 27ms/step - loss: 0.1581 - tp: 18192.0000 - fp: 650.0000 - tn: 19833.0000 - fn: 2285.0000 - accuracy: 0.9283 - precision: 0.9655 - recall: 0.8884 - auc: 0.9849 - val_loss: 0.0902 - val_tp: 82.0000 - val_fp: 955.0000 - val_tn: 44527.0000 - val_fn: 5.0000 - val_accuracy: 0.9789 - val_precision: 0.0791 - val_recall: 0.9425 - val_auc: 0.9839
Epoch 47/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.1579 - tp: 18301.0000 - fp: 676.0000 - tn: 19760.0000 - fn: 2223.0000 - accuracy: 0.9292 - precision: 0.9644 - recall: 0.8917 - auc: 0.9853 - val_loss: 0.0892 - val_tp: 82.0000 - val_fp: 956.0000 - val_tn: 44526.0000 - val_fn: 5.0000 - val_accuracy: 0.9789 - val_precision: 0.0790 - val_recall: 0.9425 - val_auc: 0.9840
Epoch 48/1000
20/20 [==============================] - 1s 28ms/step - loss: 0.1503 - tp: 18172.0000 - fp: 593.0000 - tn: 19959.0000 - fn: 2236.0000 - accuracy: 0.9309 - precision: 0.9684 - recall: 0.8904 - auc: 0.9867 - val_loss: 0.0887 - val_tp: 82.0000 - val_fp: 970.0000 - val_tn: 44512.0000 - val_fn: 5.0000 - val_accuracy: 0.9786 - val_precision: 0.0779 - val_recall: 0.9425 - val_auc: 0.9840
Epoch 49/1000
20/20 [==============================] - 0s 25ms/step - loss: 0.1572 - tp: 18217.0000 - fp: 750.0000 - tn: 19709.0000 - fn: 2284.0000 - accuracy: 0.9259 - precision: 0.9605 - recall: 0.8886 - auc: 0.9852 - val_loss: 0.0876 - val_tp: 82.0000 - val_fp: 964.0000 - val_tn: 44518.0000 - val_fn: 5.0000 - val_accuracy: 0.9787 - val_precision: 0.0784 - val_recall: 0.9425 - val_auc: 0.9841
Epoch 50/1000
20/20 [==============================] - ETA: 0s - loss: 0.1529 - tp: 18230.0000 - fp: 696.0000 - tn: 19874.0000 - fn: 2160.0000 - accuracy: 0.9303 - precision: 0.9632 - recall: 0.8941 - auc: 0.9860Restoring model weights from the end of the best epoch.
20/20 [==============================] - 0s 23ms/step - loss: 0.1529 - tp: 18230.0000 - fp: 696.0000 - tn: 19874.0000 - fn: 2160.0000 - accuracy: 0.9303 - precision: 0.9632 - recall: 0.8941 - auc: 0.9860 - val_loss: 0.0860 - val_tp: 82.0000 - val_fp: 941.0000 - val_tn: 44541.0000 - val_fn: 5.0000 - val_accuracy: 0.9792 - val_precision: 0.0802 - val_recall: 0.9425 - val_auc: 0.9843
Epoch 00050: early stopping

Re-check training history

plot_metrics(resampled_history)

png

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)

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 :  0.09607589244842529
tp :  84.0
fp :  1195.0
tn :  55676.0
fn :  7.0
accuracy :  0.9788982272148132
precision :  0.06567630916833878
recall :  0.9230769276618958
auc :  0.9697299599647522

Legitimate Transactions Detected (True Negatives):  55676
Legitimate Transactions Incorrectly Detected (False Positives):  1195
Fraudulent Transactions Missed (False Negatives):  7
Fraudulent Transactions Detected (True Positives):  84
Total Fraudulent Transactions:  91

png

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')

<matplotlib.legend.Legend at 0x7fa4bc66c9b0>

png

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.

Classification on imbalanced data

Setup

Data processing and exploration

Download the Kaggle Credit Card Fraud data set

Examine the class label imbalance

Clean, split and normalize the data

Look at the data distribution

Define the model and metrics

Understanding useful metrics

Baseline model

Build the model

Optional: Set the correct initial bias.

Checkpoint the initial weights

Confirm that the bias fix helps

Train the model

Check training history

Evaluate metrics

Plot the ROC

Class weights

Calculate class weights

Train a model with class weights

Check training history

Evaluate metrics

Plot the ROC

Oversampling

Oversample the minority class

Using NumPy

Using tf.data

Train on the oversampled data

Check training history

Re-train

Re-check training history

Evaluate metrics

Plot the ROC

Applying this tutorial to your problem

Using `tf.data`