This project explores and predicts the popularity of online news articles using machine learning techniques. By analyzing a dataset of articles from Mashable.com, we aim to identify key factors that influence article engagement and develop models to predict the number of shares an article will receive.
The dataset contains metadata from 39,797 articles published on Mashable.com, with 61 attributes including:
- Content features (word count, images, links)
- Keyword metrics
- Natural language processing metrics (sentiment, subjectivity)
- Temporal features (day of week, weekend)
- Social metrics (shares, author followers) - additional data we collected ourselves after recognizing that the original dataset focused too heavily on article structure while neglecting content factors
- Analyzed class distribution of article shares
- Examined content, channel, and author-related features
- Identified temporal patterns in publishing and engagement
- Performed correlation analysis to find predictive features
- Transformed skewed features to improve distribution
- Applied standardization to numeric features
- Created interaction features and polynomial features
- Reduced dimensionality by removing highly correlated features
- Applied SMOTE to handle class imbalance
- Developed and compared multiple classification models:
- Logistic Regression
- Random Forest
- Gradient Boosting
- XGBoost
- LightGBM
- Neural Networks
- Created ensemble models through stacking
- Performed hyperparameter tuning for optimal performance
- Evaluated with accuracy, F1-score and confusion matrices
- Weekends show higher sharing rates despite fewer articles being published
- Certain content channels (Tech, Social Media) correlate with higher shares
- Keyword metrics strongly influence article shareability
- Best model achieved ~68% accuracy in predicting share categories
- Important features include: keyword metrics, whether published on weekend, channel type, and sentiment measures
| Model | Test Accuracy | F1 Score |
|---|---|---|
| Optimized XGBoost | 67.84% | 67.85% |
| Top 3 Stacking | 67.51% | 67.55% |
| LightGBM | 67.39% | 67.35% |
| Gradient Boosting | 67.00% | 67.01% |
| Random Forest | 66.97% | 66.99% |
| Logistic Regression | 66.41% | 66.34% |
| XGBoost | 66.12% | 66.11% |
| Neural Network | 61.45% | 61.46% |
- Limited predictive accuracy suggests external factors influence sharing behavior
- Feature engineering and hyperparameter optimization provided modest gains
- Future approaches could incorporate real-time social media data
- Integrating audience demographics and browsing behavior could improve prediction accuracy
Anqi Gu, Evelyn Zhou, Han Zhang, Hongyu Liao, Yiwei Li
MSDS 422 Practical Machine Learning (March 2025)
For a comprehensive analysis and in-depth discussion of this project, please refer to our full report.