Pemodelan Random Forest dan XGBoost dalam Memprediksi Jawaban Non-response pada Exit Poll Pilgub di Jawa Tengah

Abstract

Pelaksanaan exit poll sering menghadapi permasalahan non-response, seperti yang terjadi pada exit poll Pilgub di Jawa Tengah. Adanya non-response dapat memengaruhi prediksi hasil Pilgub yang kurang akurat sehingga mengurangi tingkat kepercayaan publik terhadap lembaga survei. Penelitian ini bertujuan untuk membandingkan kinerja metode Random Forest dan XGBoost tanpa dan dengan SMOTE-NC, serta menerapkan model klasifikasi terbaik untuk memprediksi preferensi Calon Gubernur dan Wakil Gubernur di Jawa Tengah. Data yang digunakan dalam penelitian ini sebanyak 1.733 responden yang meliputi 1.560 responden untuk pemodelan dan 173 responden memberikan jawaban non-response. Permasalahan non-response diatasi dengan pemodelan klasifikasi menggunakan Random Forest dan XGBoost tanpa dan dengan SMOTE-NC. Proses optimasi hyperparameter dilakukan dengan grid search dan 10-fold cross validation. Pemodelan diulang sebanyak 10 kali untuk memperoleh performa model yang optimal. Model XGBoost dengan SMOTE-NC mencatat balanced accuracy 84,64%, F1-score 81,40%, precision 82,52%, dan recall 80,40%, sedikit lebih baik dibandingkan Random Forest dengan SMOTE-NC yang mencatat balanced accuracy 84,23%, F1-score 80,85%, precision 79,87%, dan recall 81,93%. Berdasarkan metrik evaluasi tersebut, XGBoost dengan SMOTE-NC dipilih sebagai model terbaik. Model ini kemudian diterapkan untuk memprediksi jawaban responden yang non-response dan menghasilkan balanced accuracy 86,02%, F1-score 83,21%, precision 82,61%, dan recall 83,82%.

The implementation of exit polls often faces non-response issues, as was the case with the exit poll for the gubernatorial election in Central Java. Non-response can affect the accuracy of gubernatorial election predictions, thereby reducing public trust in survey institutions. This study aims to compare the performance of the Random Forest and XGBoost methods with and without SMOTE-NC, and to apply the best classification model to predict preferences for the Governor and Vice Governor candidates in Central Java. The data used in this study consists of 1.733 respondents, including 1.560 respondents for modeling and 173 respondents who provided non-response answers. The non-response issue was addressed by classification modeling using Random Forest and XGBoost with and without SMOTE-NC. Hyperparameter optimization was performed using grid search and 10-fold cross-validation. Modeling was repeated 10 times to obtain optimal model performance. The XGBoost model with SMOTE-NC recorded a balanced accuracy of 84.64%, an F1-score of 81.40%, precision of 82.52%, and recall of 80.40%, slightly better than Random Forest with SMOTE-NC, which recorded a balanced accuracy of 84.23%, an F1-score of 80.85%, precision of 79.87%, and recall of 81.93%. Based on these evaluation metrics, XGBoost with SMOTE-NC was selected as the best model. This model was then applied to predict non-response answers from respondents and produced a balanced accuracy of 86.02%, an F1-score of 83.21%, precision of 82.61%, and recall of 83.82%.