← AI Engineering / Case study 05
Fake News Detection.
A machine-learning classifier that sorts real headlines from fake ones, and an honest comparison of which models actually earn their place.
- Team
- Rima, Domien, Dimi
- Context
- Ironhack AI Engineering
- Year
- 2026
- Best result
- 95.5% accuracy · F1 0.953
Project presentation · use ← → or swipe
Approach
Pipeline
- Cleaning: NLTK normalisation, stop-word removal and Snowball stemming.
- Features: TF-IDF on 1–2-grams (min_df 5, max_df 0.8) plus FastText word embeddings.
- Models: Logistic Regression, Multinomial Naive Bayes, Random Forest and SVM, compared on the same validation split.
- Ensemble: a probability-averaging hybrid of NB + LR + FastText beat every single model.
Results
| Model | Accuracy | F1 |
|---|---|---|
| Random Forest | 90.51% | — |
| Naive Bayes | 92.45% | 0.921 |
| SVM | 93.53% | 0.934 |
| Logistic Regression | 93.71% | 0.935 |
| NB + LR ensemble | 94.91% | 0.947 |
| Triple hybrid (NB + LR + FastText) | 95.48% | 0.953 |
Random Forest scored 99.97% on training data but only 90.5% on validation, a textbook case of overfitting. Spotting that early is why we went with the lighter, better-generalising ensemble.