Klasifikasi pada data yang tidak seimbang

Lihat di TensorFlow.org Jalankan di Google Colab Lihat sumber di GitHub Unduh buku catatan

Tutorial ini menunjukkan bagaimana mengklasifikasikan dataset yang sangat tidak seimbang di mana jumlah contoh di satu kelas jauh lebih banyak daripada contoh di kelas lain. Anda akan bekerja dengan kumpulan data Deteksi Penipuan Kartu Kredit yang dihosting di Kaggle. Tujuannya adalah untuk mendeteksi hanya 492 transaksi penipuan dari total 284.807 transaksi. Anda akan menggunakan Keras untuk menentukan model dan bobot kelas untuk membantu model belajar dari data yang tidak seimbang. .

Tutorial ini berisi kode lengkap untuk:

  • Muat file CSV menggunakan Pandas.
  • Buat rangkaian pelatihan, validasi, dan pengujian.
  • Tentukan dan latih model menggunakan Keras (termasuk pengaturan bobot kelas).
  • Evaluasi model menggunakan berbagai metrik (termasuk presisi dan ingatan).
  • Cobalah teknik umum untuk menangani data yang tidak seimbang seperti:
    • Pembobotan kelas
    • Pengambilan sampel berlebih

Mempersiapkan

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']

Pemrosesan dan eksplorasi data

Unduh kumpulan data Penipuan Kartu Kredit Kaggle

Pandas adalah pustaka Python dengan banyak utilitas bermanfaat untuk memuat dan bekerja dengan data terstruktur. Ini dapat digunakan untuk mengunduh CSV ke dalam Pandas 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()

Periksa ketidakseimbangan label kelas

Mari kita lihat ketidakseimbangan dataset:

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)

Hal ini menunjukkan fraksi kecil dari sampel positif.

Bersihkan, pisahkan, dan normalkan data

Data mentah memiliki beberapa masalah. Pertama, kolom Time dan Amount terlalu bervariasi untuk digunakan secara langsung. Jatuhkan kolom Time (karena tidak jelas artinya) dan ambil log kolom Amount untuk mengurangi jangkauannya.

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)

Pisahkan dataset menjadi train, validasi, dan set test. Set validasi digunakan selama pemasangan model untuk mengevaluasi kerugian dan metrik apa pun, namun model tidak sesuai dengan data ini. Set tes sama sekali tidak digunakan selama fase pelatihan dan hanya digunakan di akhir untuk mengevaluasi seberapa baik model digeneralisasi ke data baru. Hal ini sangat penting dengan set data yang tidak seimbang di mana overfitting menjadi perhatian yang signifikan dari kurangnya data pelatihan.

# 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)

Normalisasikan fitur input menggunakan sklearn StandardScaler. Ini akan mengatur mean ke 0 dan standar deviasi ke 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)

Perhatikan distribusi datanya

Selanjutnya bandingkan distribusi contoh positif dan negatif pada beberapa fitur. Pertanyaan yang bagus untuk ditanyakan pada diri sendiri pada saat ini adalah:

  • Apakah distribusi ini masuk akal?
    • Ya. Anda telah menormalkan input dan ini sebagian besar terkonsentrasi di kisaran +/- 2 .
  • Dapatkah Anda melihat perbedaan antara distribusi?
    • Ya, contoh-contoh positif mengandung tingkat nilai ekstrim yang jauh lebih tinggi.
pos_df = pd.DataFrame(train_features[ bool_train_labels], columns=train_df.columns)
neg_df = pd.DataFrame(train_features[~bool_train_labels], columns=train_df.columns)

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

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

png

png

Tentukan model dan metrik

Tentukan fungsi yang membuat jaringan saraf sederhana dengan lapisan tersembunyi yang terhubung secara padat, lapisan putus sekolah untuk mengurangi overfitting, dan lapisan sigmoid keluaran yang mengembalikan kemungkinan transaksi yang curang:

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

Memahami metrik yang berguna

Perhatikan bahwa ada beberapa metrik yang ditentukan di atas yang dapat dihitung oleh model yang akan membantu saat mengevaluasi kinerja.

  • Negatif palsu dan positif palsu adalah sampel yang salah diklasifikasikan
  • Negatif sejati dan positif sejati adalah sampel yang diklasifikasikan dengan benar
  • Akurasi adalah persentase contoh yang diklasifikasikan dengan benar > \(\frac{\text{true samples} }{\text{total samples} }\)
  • Presisi adalah persentase prediksi positif yang diklasifikasikan dengan benar > \(\frac{\text{true positives} }{\text{true positives + false positives} }\)
  • Ingat adalah persentase positif aktual yang diklasifikasikan dengan benar > \(\frac{\text{true positives} }{\text{true positives + false negatives} }\)
  • AUC mengacu pada Area Di Bawah Kurva dari kurva Karakteristik Operasi Penerima (ROC-AUC). Metrik ini sama dengan probabilitas bahwa pengklasifikasi akan memberi peringkat sampel positif acak lebih tinggi daripada sampel negatif acak.
  • AUPRC mengacu pada Area Di Bawah Kurva Precision-Recall Curve. Metrik ini menghitung pasangan presisi-recall untuk ambang probabilitas yang berbeda.

Baca lebih lajut:

Model dasar

Bangun modelnya

Sekarang buat dan latih model Anda menggunakan fungsi yang telah ditentukan sebelumnya. Perhatikan bahwa model cocok menggunakan ukuran batch standar yang lebih besar dari 2048, ini penting untuk memastikan bahwa setiap batch memiliki peluang yang layak untuk mengandung beberapa sampel positif. Jika ukuran batch terlalu kecil, mereka kemungkinan tidak akan memiliki transaksi penipuan untuk dipelajari.

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
_________________________________________________________________

Uji coba model:

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

Opsional: Tetapkan bias awal yang benar.

Tebakan awal ini tidak bagus. Anda tahu dataset tidak seimbang. Atur bias layer output untuk mencerminkan hal itu (Lihat: A Recipe for Training Neural Networks: "init well" ). Ini dapat membantu dengan konvergensi awal.

Dengan inisialisasi bias default, kerugiannya adalah tentang math.log(2) = 0.69314

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

Bias yang benar untuk ditetapkan dapat diturunkan dari:

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

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

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

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

Tetapkan itu sebagai bias awal, dan model akan memberikan tebakan awal yang jauh lebih masuk akal.

Seharusnya dekat: pos/total = 0.0018

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

Dengan inisialisasi ini, kerugian awal harus kira-kira:

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

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

Kerugian awal ini sekitar 50 kali lebih kecil daripada jika dengan inisialisasi naif.

Dengan cara ini model tidak perlu menghabiskan beberapa zaman pertama hanya belajar bahwa contoh-contoh positif tidak mungkin. Ini juga memudahkan untuk membaca plot kerugian selama pelatihan.

Pos pemeriksaan bobot awal

Untuk membuat berbagai pelatihan berjalan lebih sebanding, simpan bobot model awal ini dalam file pos pemeriksaan, dan muat ke setiap model sebelum pelatihan:

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

Konfirmasikan bahwa perbaikan bias membantu

Sebelum melanjutkan, konfirmasikan dengan cepat bahwa inisialisasi bias yang hati-hati benar-benar membantu.

Latih model selama 20 epoch, dengan dan tanpa inisialisasi yang cermat ini, dan bandingkan kerugiannya:

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

png

Gambar di atas memperjelas: Dalam hal kehilangan validasi, pada masalah ini, inisialisasi yang cermat ini memberikan keuntungan yang jelas.

Latih modelnya

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

Periksa riwayat pelatihan

Di bagian ini, Anda akan menghasilkan plot akurasi dan kehilangan model Anda pada set pelatihan dan validasi. Ini berguna untuk memeriksa overfitting, yang dapat Anda pelajari lebih lanjut di tutorial Overfit dan underfit .

Selain itu, Anda dapat membuat plot ini untuk salah satu metrik yang Anda buat di atas. Negatif palsu dimasukkan sebagai contoh.

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

    plt.legend();
plot_metrics(baseline_history)

png

Evaluasi metrik

Anda dapat menggunakan matriks konfusi untuk meringkas label aktual vs. yang diprediksi, di mana sumbu X adalah label yang diprediksi dan sumbu Y adalah label yang sebenarnya:

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]))

Evaluasi model Anda pada kumpulan data pengujian dan tampilkan hasil untuk metrik yang Anda buat di atas:

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

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

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

png

Jika model telah memprediksi semuanya dengan sempurna, ini akan menjadi matriks diagonal di mana nilai di luar diagonal utama, yang menunjukkan prediksi yang salah, akan menjadi nol. Dalam hal ini matriks menunjukkan bahwa Anda memiliki hasil positif palsu yang relatif sedikit, artinya ada relatif sedikit transaksi sah yang ditandai secara tidak benar. Namun, Anda mungkin ingin memiliki lebih sedikit negatif palsu meskipun biaya untuk meningkatkan jumlah positif palsu. Pertukaran ini mungkin lebih disukai karena negatif palsu akan memungkinkan transaksi penipuan dilakukan, sedangkan positif palsu dapat menyebabkan email dikirim ke pelanggan untuk meminta mereka memverifikasi aktivitas kartu mereka.

Rencanakan ROC

Sekarang plot ROC . Plot ini berguna karena menunjukkan, sekilas, kisaran kinerja yang dapat dicapai model hanya dengan menyetel ambang keluaran.

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

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

png

Plot AUPRC

Sekarang plot AUPRC . Area di bawah kurva presisi-recall interpolasi, diperoleh dengan memplot (recall, presisi) poin untuk nilai-nilai yang berbeda dari ambang klasifikasi. Bergantung pada cara penghitungannya, PR AUC mungkin setara dengan presisi rata-rata 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('Recall')
    plt.ylabel('Precision')
    plt.grid(True)
    ax = plt.gca()
    ax.set_aspect('equal')
plot_prc("Train Baseline", train_labels, train_predictions_baseline, color=colors[0])
plot_prc("Test Baseline", test_labels, test_predictions_baseline, color=colors[0], linestyle='--')
plt.legend(loc='lower right');

png

Sepertinya presisinya relatif tinggi, tetapi daya ingat dan area di bawah kurva ROC (AUC) tidak setinggi yang Anda inginkan. Pengklasifikasi sering menghadapi tantangan saat mencoba memaksimalkan presisi dan ingatan, yang terutama benar saat bekerja dengan kumpulan data yang tidak seimbang. Penting untuk mempertimbangkan biaya dari berbagai jenis kesalahan dalam konteks masalah yang Anda pedulikan. Dalam contoh ini, negatif palsu (transaksi penipuan tidak terjawab) mungkin memiliki biaya finansial, sementara positif palsu (transaksi salah ditandai sebagai penipuan) dapat mengurangi kebahagiaan pengguna.

Bobot kelas

Hitung bobot kelas

Tujuannya adalah untuk mengidentifikasi transaksi penipuan, tetapi Anda tidak memiliki banyak sampel positif untuk digunakan, jadi Anda ingin pengklasifikasi sangat mempertimbangkan beberapa contoh yang tersedia. Anda dapat melakukan ini dengan melewatkan bobot Keras untuk setiap kelas melalui sebuah parameter. Ini akan menyebabkan model "lebih memperhatikan" contoh dari kelas yang kurang terwakili.

# 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

Latih model dengan bobot kelas

Sekarang coba latih ulang dan evaluasi model dengan bobot kelas untuk melihat bagaimana hal itu memengaruhi prediksi.

weighted_model = make_model()
weighted_model.load_weights(initial_weights)

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

Periksa riwayat pelatihan

plot_metrics(weighted_history)

png

Evaluasi metrik

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

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

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

png

Di sini Anda dapat melihat bahwa dengan bobot kelas akurasi dan presisi lebih rendah karena lebih banyak positif palsu, tetapi sebaliknya recall dan AUC lebih tinggi karena model juga menemukan lebih banyak positif benar. Meskipun memiliki akurasi yang lebih rendah, model ini memiliki daya ingat yang lebih tinggi (dan mengidentifikasi lebih banyak transaksi penipuan). Tentu saja, ada biaya untuk kedua jenis kesalahan (Anda juga tidak ingin mengganggu pengguna dengan menandai terlalu banyak transaksi yang sah sebagai penipuan). Pertimbangkan dengan cermat pertukaran antara berbagai jenis kesalahan ini untuk aplikasi Anda.

Rencanakan 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');

png

Plot AUPRC

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

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


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

png

Pengambilan sampel berlebih

Oversample kelas minoritas

Pendekatan terkait adalah dengan mengambil sampel ulang dataset dengan melakukan oversampling kelas minoritas.

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]

Menggunakan NumPy

Anda dapat menyeimbangkan kumpulan data secara manual dengan memilih jumlah indeks acak yang tepat dari contoh positif:

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

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

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

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

resampled_features.shape
(363902, 29)

Menggunakan tf.data

Jika Anda menggunakan tf.data , cara termudah untuk menghasilkan contoh yang seimbang adalah memulai dengan kumpulan data positive dan negative , dan menggabungkannya. Lihat panduan tf.data untuk lebih banyak contoh.

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)

Setiap kumpulan data menyediakan (feature, label) pasangan:

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

Label:  1

Gabungkan keduanya menggunakan tf.data.Dataset.sample_from_datasets :

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

Untuk menggunakan kumpulan data ini, Anda memerlukan jumlah langkah per epoch.

Definisi "zaman" dalam hal ini kurang jelas. Katakanlah itu jumlah batch yang diperlukan untuk melihat setiap contoh negatif satu kali:

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

Latih data yang di-oversampling

Sekarang coba latih model dengan kumpulan data yang disampel ulang alih-alih menggunakan bobot kelas untuk melihat bagaimana metode ini dibandingkan.

resampled_model = make_model()
resampled_model.load_weights(initial_weights)

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

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

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

Jika proses pelatihan mempertimbangkan seluruh dataset pada setiap pembaruan gradien, oversampling ini pada dasarnya akan identik dengan pembobotan kelas.

Namun saat melatih model secara batch, seperti yang Anda lakukan di sini, data yang di-oversampling memberikan sinyal gradien yang lebih halus: Alih-alih setiap contoh positif ditampilkan dalam satu batch dengan bobot besar, mereka ditampilkan dalam banyak batch berbeda setiap kali dengan berat kecil.

Sinyal gradien yang lebih halus ini memudahkan untuk melatih model.

Periksa riwayat pelatihan

Perhatikan bahwa distribusi metrik akan berbeda di sini, karena data pelatihan memiliki distribusi yang sama sekali berbeda dari data validasi dan pengujian.

plot_metrics(resampled_history)

png

Latih ulang

Karena pelatihan lebih mudah pada data yang seimbang, prosedur pelatihan di atas mungkin terlalu cepat.

Jadi, pisahkan epoch untuk memberikan kontrol yang lebih baik kepada tf.keras.callbacks.EarlyStopping kapan harus menghentikan pelatihan.

resampled_model = make_model()
resampled_model.load_weights(initial_weights)

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

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

Periksa kembali riwayat pelatihan

plot_metrics(resampled_history)

png

Evaluasi metrik

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

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

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

png

Rencanakan 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');

png

Plot AUPRC

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

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

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

png

Menerapkan tutorial ini untuk masalah Anda

Klasifikasi data yang tidak seimbang adalah tugas yang sulit karena hanya ada sedikit sampel untuk dipelajari. Anda harus selalu memulai dengan data terlebih dahulu dan melakukan yang terbaik untuk mengumpulkan sampel sebanyak mungkin dan memberikan pemikiran substansial tentang fitur apa yang mungkin relevan sehingga model dapat memaksimalkan kelas minoritas Anda. Pada titik tertentu model Anda mungkin kesulitan untuk meningkatkan dan memberikan hasil yang Anda inginkan, jadi penting untuk mengingat konteks masalah Anda dan pertukaran antara berbagai jenis kesalahan.