ClickShield

Lightweight NLP pipeline to detect clickbait headlines using TF‑IDF features and a Logistic Regression classifier.

• Goal: classify news/article headlines as clickbait vs non‑clickbait
• Tech: Python, scikit‑learn, pandas, NLTK
• Current Version: v1.1
• Artifacts (v1.1):
  • models/clickbait_lr_tfidf_v1.1.pkl
  • models/clickbait_tfidf_vectorizer_v1.1.pkl

Model Performance Summary (v1.1)

The v1.1 model achieved an Accuracy of 0.9809, Precision of 0.9846, Recall of 0.9771, and F1-Score of 0.9808. This represents significant improvements over v1.0. For a detailed breakdown of all metrics and version history, please refer to Model Details.

---

Project layout

• data/raw/ — Original datasets.
• data/processed/ — Cleaned datasets used for training/evaluation.
• src/ — Source code:
  • preprocessing.py — Text cleaning and normalization utilities.
  • dataset.py — Dataset loading and dataset‑level preprocessing.
  • features.py — TF‑IDF feature extraction.
  • train_model.py — End‑to‑end training, evaluation and model saving.
  • predict.py — Inference helpers (single, batch, and CSV‑based predictions).
• models/ — Saved model and vectorizer (*.pkl).
• main.py — CLI entry point to train and run the model.
• requirements.txt — Python dependencies.
• docs/ — Project documentation and visualization figures.

---

Setup

1. Create a virtual environment (Windows PowerShell):

python -m venv venv
.\venv\Scripts\Activate.ps1

2. Install dependencies:

pip install -r requirements.txt

3. Prepare data:

• Place your raw CSV in data/raw/, e.g. data/raw/clickbait_data.csv.
• Required columns:
  • headline — the text of the headline
  • clickbait — label (1 = clickbait, 0 = non‑clickbait)

---

Using main.py (recommended entry point)

Train a model

This will train the Logistic Regression model and save the artifacts with versioned names:

python main.py train `--data-path data/raw/clickbait_data.csv `--model-path models/clickbait_lr_tfidf_v1.1.pkl `--vectorizer-path models/clickbait_tfidf_vectorizer_v1.1.pkl

If you omit the arguments, main.py uses v1.0 defaults. For v1.1, specify the paths explicitly.

Predict for a single headline

Use the trained model and vectorizer to classify one headline:

python main.py predict "You Won't Believe What Happened Next!"`--model-path models/clickbait_lr_tfidf_v1.1.pkl `--vectorizer-path models/clickbait_tfidf_vectorizer_v1.1.pkl

Model Pipeline

---

Module overview (src/)

src/preprocessing.py

Low‑level text preprocessing utilities that operate on raw strings or token lists:

• lowercase_text(text) — converts text to lowercase.
• remove_urls(text) — removes http:// / https:// style URLs.
• remove_punctuation(text) — strips punctuation characters.
• remove_numbers(text) — removes digits from the text.
• tokenize_text(text) — tokenizes text into word tokens using NLTK.
• remove_stopwords(tokens) — removes common English stopwords.
• lemmatize_tokens(tokens) — lemmatizes tokens using WordNet.
• preprocess_text(text) — runs all of the above in sequence and returns a cleaned string; this is the main function used by the rest of the pipeline.

src/dataset.py

Helpers for reading and preparing datasets:

• load_dataset(filepath)
  • Reads a CSV from filepath into a pandas DataFrame.
  • Drops rows with missing values.
• preprocess_dataset(df)
  • Takes a DataFrame with a headline column.
  • Applies preprocess_text to each headline.
  • Returns a copy of the DataFrame with an extra clean_headline column.

src/features.py

TF‑IDF feature extraction:

• extract_tfidf_features(df, text_column="clean_headline")
  • Fits a TfidfVectorizer (1‑, 2‑, and 3‑grams, up to 8000 features) on the specified text column.
  • v1.1 improvements: Extended n-gram range from (1,2) to (1,3) and increased max features from 5000 to 8000 for better context capture.
  • Returns:
    • tfidf_df — a DataFrame of dense TF‑IDF features.
    • vectorizer — the fitted TfidfVectorizer for later use at inference time.

src/train_model.py

End‑to‑end training pipeline:

• train_clickbait_model(data_path, test_size=0.2, random_state=42, max_iter=1000)
  • Loads and cleans the dataset (load_dataset → preprocess_dataset).
  • Splits into train/test sets.
  • Builds TF‑IDF features using extract_tfidf_features.
  • Trains a LogisticRegression classifier with C=10 for improved generalization.
  • v1.1 improvements: Adjusted regularization parameter C from default (1.0) to 10 for better model performance.
  • Evaluates on the test set and returns:
    • model — trained classifier
    • vectorizer — fitted TF‑IDF vectorizer
    • metrics — accuracy, precision, recall, F1, confusion matrix, and classification report
• save_model(model, vectorizer, model_path, vectorizer_path)
  • Saves both objects to disk using joblib.
• print_metrics(metrics)
  • Nicely prints the evaluation metrics to the console.

src/predict.py

Inference utilities for using a trained model and vectorizer:

• load_model(model_path="models/clickbait_lr_tfidf.pkl", vectorizer_path="models/clickbait_tfidf_vectorizer.pkl")
  • Loads a model and vectorizer from disk.
  • For the v1.1 artifacts, call it like:

fromsrc.predictimportload_modelmodel, vectorizer=load_model(
model_path="models/clickbait_lr_tfidf_v1.1.pkl",
vectorizer_path="models/clickbait_tfidf_vectorizer_v1.1.pkl",
)

• predict_headline(headline, model, vectorizer, return_probability=False)
  • Preprocesses a single headline string and predicts 1 (clickbait) / 0 (non‑clickbait).
  • If return_probability=True, returns a dict with prediction, label, and probabilities.
• predict_batch(headlines, model, vectorizer, return_probabilities=False)
  • Predicts for a list/Series of headlines.
  • Optionally returns a DataFrame with predictions and probabilities.
• predict_from_file(filepath, model, vectorizer, text_column="headline", return_probabilities=False)
  • Reads a CSV from filepath.
  • Runs predictions on the text_column.
  • Returns a DataFrame combining original data with predictions.

Example Python usage:

fromsrc.predictimportload_model, predict_headlinemodel, vectorizer=load_model(
model_path="models/clickbait_lr_tfidf_v1.1.pkl",
vectorizer_path="models/clickbait_tfidf_vectorizer_v1.1.pkl",
)
result=predict_headline(
"You Won't Believe What Happened Next!",
model,
vectorizer,
return_probability=True,
)
print(result)

---