A credibility classifier that shows its work

The problem

• A model that flags an article as fake without showing why is just another authority asking to be trusted.
• The data is imbalanced (more real than fake), so accuracy misleads: a model can look strong while failing on the fake-news minority it exists to catch.
• Exploratory analysis surfaced a leakage trap: the subject/category field alone was predictive enough to let the model cheat.

Architecture

• A defensive input layer rejects placeholder text, repeated-character spam, and low-quality input before inference.
• Interpretable feature engineering: sentence and readability statistics, punctuation, uppercase ratios, and URL presence, chosen because EDA showed fake articles skew longer and more emphatic.
• XGBoost over those features with a decision threshold tuned to 0.45, not raw text embeddings, so every signal stays explainable.
• SHAP attributions map each prediction back to the features that drove it, surfaced directly in the UI.
• Served as a Dockerized FastAPI API and a Streamlit front end, deployed on Render and Streamlit Cloud.

The decision I'd defend

• I excluded the subject/category field even though it would have inflated the headline number.
• EDA showed it was a leakage shortcut, and a tool that is right for the wrong reason launders a shortcut as judgment.

How I knew it worked

• 98.47% F1 on a held-out test set, measured with F1 and AUC-PR under stratified sampling, with the leak removed before training.
• SHAP made every prediction auditable: I could confirm the model keyed on the linguistic signals EDA flagged, not a spurious artifact.