News Article Classification with Machine Learning and Deep Learning Models

Overview

The goal of this project is to build a robust classification system for news articles using different models, such as Logistic Regression, Support Vector Machines (SVM), Naive Bayes, K-Nearest Neighbors (KNN), Random Forest, and DistilBERT. The system was tested on a publicly available Kaggle dataset with the aim of comparing the performance of traditional machine learning algorithms versus a transformer-based model (DistilBERT).

Data Preprocessing

• Data was scraped using the newspaper3k library, with additional handling for JavaScript-rendered content using requests-html.
• The text was preprocessed (lowercased, tokenized, lemmatized, and stopwords removed).
• Class imbalance was addressed using undersampling, resulting in a balanced dataset of 26,216 articles.

Models & Techniques

1. Traditional Models:

   • Logistic Regression
   • Support Vector Machine (SVM)
   • Naive Bayes
   • Random Forest
   • K-Nearest Neighbors (KNN) with cosine and Euclidean distances

2. Deep Learning Model:

   • DistilBERT fine-tuned on raw article content

The project uses TF-IDF and BoW for text vectorization and evaluates model performance across multiple metrics, including accuracy, precision, recall, and F1-score.

Results

The results highlight the performance of the models with respect to their classification accuracy, and the following table summarizes the accuracy for the key models evaluated in the project:

Model Accuracy Comparison:

Model	Accuracy
KNN (TF-IDF, Cosine Distance, k=4)	96.803%
KNN (TF-IDF, Euclidean Distance, k=4)	96.777%
Support Vector Machine (TF-IDF)	96.377%
Logistic Regression (TF-IDF)	95.67%
Naive Bayes (TF-IDF)	95.43%
Random Forest	90.25%
DistilBERT (Raw Text)	96.834%

• KNN with TF-IDF and Cosine distance yielded the highest accuracy among traditional models, achieving 96.83% (for k=4).
• The DistilBERT model achieved an impressive accuracy of 96.834% but was computationally more expensive to run.

Conclusion

This project demonstrates the strengths of both traditional machine learning models and deep learning models (such as DistilBERT) for news article classification. The KNN model with TF-IDF offers a compelling trade-off between high accuracy and computational efficiency. On the other hand, DistilBERT provides state-of-the-art accuracy but is more resource-intensive, making it less practical for real-time applications on platforms with limited resources.