Project: News Headlines Classification (Fake vs Real)

This project aims to build a supervised learning model to classify news headlines as fake or real. A Natural Language Processing (NLP) pipeline is implemented to preprocess the data, apply vectorization techniques, and train classification models.

---

1. Project Workflow

Main Steps

1. Data Loading and Exploration:

   • Load the dataset containing two main columns:
     • Headline: The headline text.
     • Veracity: Binary label indicating whether the headline is fake (0) or real (1).
   • Explore the distribution of the labels to understand class balance.

2. Text Preprocessing:

   • Convert headlines to lowercase.
   • Remove special characters and punctuation.
   • (Optional) Remove stopwords and apply lemmatization.

3. Text Vectorization:

   • Use CountVectorizer to convert the headlines into a numerical representation based on word frequency.
   • Experiment with bigrams (ngram_range=(1, 2)) to capture word relationships.

4. Training and Evaluating Base Models:

   • Test multiple supervised classification models:
     • Naive Bayes
     • Logistic Regression
     • Random Forest
     • SVM
   • Evaluate each model on the test set using metrics such as accuracy and classification reports.

5. Selecting the Best Model:

   • The best-performing initial model was Logistic Regression, with an accuracy of ~92.92%.

6. Hyperparameter Tuning:

   • Perform hyperparameter tuning on the winning model (Logistic Regression) using GridSearchCV.
   • Explore combinations of C, solver, and penalty.

7. Evaluating the Optimized Model:

   • Recalculate performance metrics and visualize the confusion matrix.
   • The optimized model maintained similar performance to the default model.

---

2. Tools Used

• Language: Python 3
• Libraries:
  • pandas: For handling tabular data.
  • numpy: For mathematical operations.
  • scikit-learn: For preprocessing, vectorization, and model training.
  • seaborn and matplotlib: For metric visualization.
  • nltk: For text cleaning and tokenization.

---

3. Models and Configurations

Models Evaluated

1. Naive Bayes:

   • A fast and efficient baseline model for sparse data.

2. Logistic Regression:

   • A robust linear model with regularization.
   • Initial winner with an accuracy of 92.92%.

3. Random Forest:

   • Tree-based model.
   • Competitive performance but did not surpass Logistic Regression.

4. SVM:

   • Maximizes decision boundaries.
   • Requires more training time.

Hyperparameters Tuned for Logistic Regression

• C: [0.01, 0.1, 1, 10, 100]
• penalty: ['l1', 'l2']
• solver: ['liblinear', 'lbfgs']

---

4. Evaluation Metrics

• Accuracy: Proportion of correct predictions.
• Classification Report: Detailed metrics (precision, recall, F1-score) for each class.
• Confusion Matrix: Visualization of true positives, true negatives, false positives, and false negatives.

---

5. Final Results

• Winning Model: Logistic Regression (optimized).
• Performance:
  • Accuracy: 92.92%
  • Balanced precision and recall across classes.
• Observations:
  • Hyperparameter tuning did not significantly improve the performance over the default model.
  • The default model settings are near-optimal for this problem.

---

6. Next Steps

1. Improving Preprocessing:

   • Implement TF-IDF instead of CountVectorizer.
   • Experiment with advanced embeddings (Word2Vec, GloVe, or BERT).

2. Collect More Data:

   • A larger dataset could help improve performance and generalization.

3. Try Advanced Models:

   • Gradient boosting models like XGBoost or LightGBM.
   • Transformer-based models (e.g., BERT, RoBERTa) for deeper semantic analysis.

---

7. Contact

If you have any questions or suggestions, feel free to reach out:

• Email: miguelchamizo10@gmail.com
• GitHub: Migueleven
• Linkedin: Miguel Ángel Chamizo Sánchez