Selected as an Outstanding Project in the Pattern Recognition course at Ewha Womans University.
This project focuses on predicting the "virality" of online news articles (whether an article exceeds 1,400 shares) by analyzing various features of the content. Our team achieved high predictive performance through advanced data preprocessing and ensemble modeling.
- Selected as a Top Project: Recognized for excellence in methodology and analysis within the Pattern Recognition course, followed by an official class presentation.
- Doyoon Kim, Minha Woo, Yeeun Kim, Jiwon Lee
- Goal: Build a binary classification model to predict news popularity.
- Target Variable:
y=1(Popular, >1,400 shares),y=0(Not popular) - Dataset: 22,200 samples with 46 features (approx. 10% missing values).
- Regression-based Imputation: Instead of simple mean/median filling, we performed Linear/Polynomial Regression to impute missing values for highly correlated variable pairs (r > 0.7).
- Outlier Handling with DBSCAN: Utilized Density-Based Spatial Clustering (DBSCAN) to identify and handle outliers in high-dimensional spaces, improving model robustness.
- Feature Engineering: Applied Log Transformation to skewed numerical data and utilized StandardScaler for normalization.
- Algorithm Exploration: Evaluated multiple models including Logistic Regression, Random Forest, XGBoost, CatBoost, and LightGBM.
- Final Model: Developed a Stacking Ensemble (Meta Model: Logistic Regression) to maximize predictive accuracy.
- Performance: Achieved a final Mean Score of 0.6955.
- FINAL CODES: Final Python source code for preprocessing and modeling (
code.ipynb). - REPORTS: Detailed project report and presentation slides (
report.pdf,PPT.pdf).
(I recommend inserting the following images here for maximum impact)
- DBSCAN Clustering Plot: To showcase your advanced outlier detection skills.
- Correlation Heatmap: To justify your regression-based imputation strategy.
- Model Performance Comparison: To highlight the superiority of the Stacking model.