A machine learning project that classifies news articles as real or fake using natural language processing and classification models.

Misinformation spreads fast on digital platforms and can have serious real-world consequences. This project builds and compares multiple classifiers to automatically flag news articles that are likely to be fake, based on the text content of the article.

Fake and Real News Dataset (Kaggle) — ~44,900 labeled news articles, each marked as real (1) or fake (0). Only the article text column was used for modeling; title, subject, and date were dropped.

Text preprocessing — removed punctuation, lowercased text, and stripped stopwords using NLTK

Exploratory analysis — generated word clouds for real vs. fake articles, and a bar chart of the most frequent words, to see how the two classes differ in language

Feature extraction — converted the cleaned text into numerical features using TF-IDF (TfidfVectorizer)

Model training — trained and compared four classifiers:

Evaluation — compared train vs. test accuracy for each model (to check for overfitting) and used confusion matrices to see the error breakdown

Model Train Accuracy Test Accuracy Logistic Regression 99.40% 98.75% Decision Tree Classifier 100% 99.62% Gradient Boosting Classifier 99.70% 99.55% Random Forest Classifier 99.99% 98.94%

Best model: Decision Tree Classifier, with 99.62% test accuracy, closely followed by Gradient Boosting at 99.55%.

Both Decision Tree and Random Forest hit near-100% train accuracy, which is a classic overfitting signal — they're memorizing the training data almost perfectly. Gradient Boosting shows the smallest gap between train and test accuracy (0.15 points), making it the most "honest" model of the four in terms of generalization, even though Decision Tree edges it out slightly on raw test accuracy.

The Decision Tree confusion matrix on the test set shows 5,738 correct "Fake" predictions against only 18 misclassifications, and 5,451 correct "Real" predictions against 23 misclassifications — very few errors either way.

Python, Pandas, Scikit-learn, NLTK, WordCloud, Seaborn, Matplotlib

How text needs to be cleaned (stopword removal, punctuation stripping) and vectorized (TF-IDF) before it can be fed into a classical ML model

Word clouds and word-frequency charts are a quick way to sanity-check whether the two classes actually differ in vocabulary before even training a model

Comparing train vs. test accuracy side by side is a simple but effective way to catch overfitting — Decision Tree and Random Forest both hit near-100% train accuracy, a signal that they're closer to memorizing the data than generalizing from it

Higher test accuracy alone doesn't tell the whole story — Gradient Boosting had a smaller train-test gap than Decision Tree despite a marginally lower test score, which matters more if this model were ever used on genuinely new, unseen articles

Comparing multiple classifiers systematically instead of committing to one algorithm upfront

Try a more powerful text representation (word embeddings, or a transformer-based model like BERT) and compare against the classical models here

Deploy the best model as a simple web app (Streamlit/Flask) where a user can paste an article and get a prediction

Test the model on more recent news articles to check how well it generalizes beyond the training dataset's time period