Predictive Modeling (Undergraduate Thesis)
An end-to-end multi-label classification pipeline to predict student learning behaviors using ELS activity logs.
Project Gallery
My Role & Contributions
Engineered an end-to-end multi-label classification pipeline to predict student learning behaviors. Resolved extreme dataset class imbalance via Random Oversampling and handled missing data using median imputation. Conducted rigorous hyperparameter tuning and model evaluation using Nested Cross-Validation, demonstrating that Random Forest outperformed XGBoost.
The Problem
Prediksi gaya belajar di ELS saat ini terbatas pada pendekatan single-label, padahal mahasiswa memiliki kombinasi gaya belajar (FSLSM) simultan. Data log juga mengalami class imbalance ekstrem (17.5:1).
The Approach
Menerapkan pendekatan multi-label classification dengan imputasi median dan Random Oversampling untuk menyeimbangkan data.
The Tech
Pengujian algoritma Random Forest, XGBoost, SVM, dan RBF Network, divalidasi dengan Nested Cross-Validation.
The Outcome
Random Forest terbukti paling akurat (F1-Macro 0.7063) dan stabil (standar deviasi ±0.0794) dibandingkan algoritma lainnya.