Detecting Clickbaits (2/4) - Universal-Sentence-Encoder Transfer Learning
Problem.
Given a set of 32000 headlines and their labels, whether that headline is a clickbait (label 1) or
not (label 0), you’re asked to build a model to detect clickbait headlines.
Solution.
Read data:
df = pd.read_csv("https://raw.github.com/hminooei/DSbyHadi/master/data/clickbait_data.csv.zip")
df.head(3)
Split into train/validation/test sets:
text_train_val, text_test, label_train_val, label_test = train_test_split(
df["headline"],
df["clickbait"],
test_size=0.25,
stratify=df["clickbait"],
random_state=9)
# Split the train_val dataset to train and validation separete portions.
text_train, text_val, label_train, label_val = train_test_split(
text_train_val,
label_train_val,
test_size=0.2,
random_state=9)Load Universal Sentence Encoder pre-trained network and its weights from
tensorflow hub, set the weights as trainable (trainable=True),
and add a final output layer with sigmoid activation since it’s a binary
classifier:
input = tf.keras.layers.Input(shape=[], dtype=tf.string)
use = hub.KerasLayer("https://tfhub.dev/google/universal-sentence-encoder/4", trainable=True)(input)
output = tf.keras.layers.Dense(1, activation="sigmoid")(use)
model = tf.keras.models.Model(input, output)
model.summary()Train for 2 epochs:
model.fit(text_train,
label_train,
epochs=2,
validation_data=(text_val, label_val),
verbose=2)Then we can measure the precision and recall on our test set:
label_preds_classes_test = np.round(model.predict(text_test))
test_metrics = metrics.classification_report(label_test, label_preds_classes_test, labels=[0, 1], digits=4)
print(test_metrics)
Important Points.
- The training time:
45minon Google Colab (TPUs) - Macro precision on test set:
0.9842 - Inference time per record:
~2mson my laptop (MacBook Pro: 2.3 GHz 8-Core Intel Core i9, 32 GB 2667 MHz DDR4)
Note.