Weight-based Oversampling Dan Energy-based Cleaning Untuk Penanganan Ketidakseimbangan Dan Tumpang Tindih Kelas Pada Prediksi Putus Studi Mahasiswa

Hidayat, Rahmad (2026) Weight-based Oversampling Dan Energy-based Cleaning Untuk Penanganan Ketidakseimbangan Dan Tumpang Tindih Kelas Pada Prediksi Putus Studi Mahasiswa. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241005-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Identifikasi dini mahasiswa yang berisiko putus studi menjadi salah satu langkah penting dalam upaya menurunkan angka putus studi di perguruan tinggi. Machine learning telah banyak diterapkan pada prediksi putus studi mahasiswa sehingga perguruan tinggi dapat mengenali mahasiswa yang berisiko agar intervensi akademik dapat segera diberikan. Namun, penerapan machine learning pada permasalahan ini masih menghadapi tantangan berupa ketidakseimbangan kelas (class imbalance) dan tumpang tindih kelas (class overlap), terutama pada skenario klasifikasi multikelas. Sebagian besar metode resampling pada prediksi putus studi masih berfokus pada penanganan class imbalance tanpa mempertimbangkan class overlap, yang dapat menyebabkan batas antar kelas menjadi kurang jelas sehingga menyulitkan model dalam membedakan setiap kelas target secara akurat. Penelitian ini mengusulkan metode resampling hibrida yang mengintegrasikan Weight-Based Oversampling (WBO) dan Energy-Based Cleaning (EBC) untuk menangani kedua permasalahan tersebut secara bersamaan. WBO menghasilkan sampel sintetis secara selektif pada area dengan risiko class overlap yang lebih rendah, sedangkan EBC mengurangi pengaruh class overlap melalui penyesuaian posisi sampel mayoritas tanpa menghapusnya dari data pelatihan. Penelitian ini menggunakan data historis mahasiswa Universitas Maritim Raja Ali Haji angkatan 2015–2021 yang terdiri atas 6.578 mahasiswa dan 24 fitur prediktor. Kelas target dibagi menjadi tiga kategori, yaitu Graduated, Enrolled, dan Dropout. Evaluasi dilakukan menggunakan Stratified 10-Fold Cross Validation dengan XGBoost sebagai algoritma klasifikasi terbaik berdasarkan hasil seleksi model. Kinerja dievaluasi menggunakan metrik multiclass Geometric Mean (mGM), Multiclass Area Under Curve (MAUC), dan recall, dengan fokus utama pada recall kelas dropout. Hasil eksperimen menunjukkan bahwa metode usulan menghasilkan performa terbaik dengan nilai mGM sebesar 0,848 ± 0,017 dan MAUC sebesar 0,923 ± 0,009, lebih tinggi dibandingkan seluruh metode pembanding. Metode usulan juga mencapai recall dropout sebesar 0,918 ± 0,049, meningkat dibandingkan baseline (0,806 ± 0,052), SMOTE (0,830 ± 0,043), GDHS (0,858 ± 0,043), dan MC-MBRC (0,905 ± 0,050). Studi ablasi menunjukkan bahwa WBO dan EBC memberikan kontribusi yang saling melengkapi dalam meningkatkan representasi kelas minoritas dan mengurangi pengaruh overlap. Hasil Wilcoxon Signed-Rank Test menunjukkan bahwa metode usulan menghasilkan perbedaan yang signifikan secara statistik terhadap seluruh metode pembanding pada metrik mGM, serta terhadap sebagian besar metode pembanding pada metrik MAUC dan Recall Dropout. Temuan ini menunjukkan bahwa integrasi WBO dan EBC mampu meningkatkan performa model prediksi putus studi mahasiswa dalam skenario klasifikasi multikelas.
==================================================================================================================================
Early identification of students at risk of dropout is an important step in reducing student dropout rates in higher education. Machine learning has been widely applied to student dropout prediction to identify at-risk students and support timely academic interventions. However, its application still faces challenges related to class imbalance and class overlap, particularly in multiclass classification. Most resampling methods for student dropout prediction focus solely on addressing class imbalance without considering class overlap, which can blur class boundaries and make it more difficult for models to accurately distinguish among target classes. This study proposes a hybrid resampling method integrating Weight-Based Oversampling (WBO) and Energy-Based Cleaning (EBC) to address both problems simultaneously. WBO selectively generates synthetic samples in regions with lower overlap risk, while EBC reduces the effect of overlap by adjusting the majority-class sample positions without removing them from the training data. This study used historical student data from Universitas Maritim Raja Ali Haji for the 2015–2021 cohorts, consisting of 6,578 students and 24 predictive features. The target variable was categorized into three classes: Graduated, Enrolled, and Dropout. The proposed method was evaluated using Stratified 10-Fold Cross Validation with XGBoost, which was selected as the best-performing classification algorithm based on the model selection results. Model performance was evaluated using multiclass Geometric Mean (mGM), Multiclass Area Under the Curve (MAUC), and Recall, with particular emphasis on Dropout Recall. The experimental results show that the proposed method achieved the best performance, with an mGM of 0.848 ± 0.017 and an MAUC of 0.923 ± 0.009, outperforming all comparison methods. The proposed method also achieved a Dropout Recall of 0.918 ± 0.049, representing an improvement over the baseline (0.806 ± 0.052), SMOTE (0.830 ± 0.043), GDHS (0.858 ± 0.043), and MC-MBRC (0.905 ± 0.050). The ablation study demonstrates that WBO and EBC provide complementary contributions by improving minority-class representation and reducing the effect of class overlap. Furthermore, the Wilcoxon Signed-Rank Test confirmed that the proposed method achieved statistically significant improvements over all comparison methods in terms of mGM, and over most comparison methods in terms of MAUC and Dropout Recall. These findings demonstrate that integrating WBO and EBC can improve the performance of student dropout prediction models in multiclass classification.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Class Imbalance, Class Overlap, Klasifikasi Multikelas, Metode Resampling, Prediksi Putus Studi Mahasiswa, Class imbalance, Class overlap, Multiclass classification, Resampling method, Student dropout prediction
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Rahmad Hidayat
Date Deposited:	20 Jul 2026 04:03
Last Modified:	20 Jul 2026 04:03
URI:	http://repository.its.ac.id/id/eprint/135459

Actions (login required)

View Item