<![CDATA[Seiring meningkatnya jumlah ulasan pengguna pada platform investasi digital, diperlukan metode analisis sentimen untuk membantu evaluasi kualitas layanan. Penelitian ini bertujuan membandingkan kinerja model Support Vector Machine (SVM) berbasis TF-IDF dan IndoBERT dalam klasifikasi sentimen ulasan aplikasi Bibit serta memperoleh konfigurasi hyperparameter yang sesuai untuk masing-masing model. Dataset diperoleh melalui web scraping sebanyak 4.123 ulasan Google Play Store. Setelah penyaringan dengan mengeluarkan ulasan berating 3, diperoleh 3.687 ulasan yang terdiri atas 1.956 positif dan 1.731 negatif. Penelitian menggunakan kerangka Cross-Industry Standard Process for Data Mining (CRISP-DM) yang meliputi data understanding, data preparation, modeling, evaluation, dan deployment. Optimasi dilakukan melalui hyperparameter tuning, sedangkan evaluasi menggunakan Stratified 5-Fold Cross Validation dengan metrik accuracy, precision, recall, F1-score, ROC-AUC, dan confusion matrix. Hasil evaluasi menunjukkan bahwa SVM memperoleh rata-rata akurasi sebesar 87,28%, sedangkan IndoBERT sebesar 87,25%. Selisih rata-rata akurasi sebesar 0,03 poin persentase menunjukkan bahwa kinerja kedua model berdasarkan metrik akurasi sangat berdekatan pada dataset penelitian. Analisis kesalahan (Error Analysis) digunakan untuk mengidentifikasi pola kesalahan klasifikasi, sedangkan Explainable Artificial Intelligence (XAI) diterapkan menggunakan LIME pada SVM dan Token Saliency pada IndoBERT untuk menginterpretasikan kontribusi fitur terhadap hasil prediksi. Model kemudian diimplementasikan dalam prototipe aplikasi berbasis Streamlit sebagai bentuk penerapan tahap deployment dalam CRISP-DM. Sistem masih dijalankan pada lingkungan pengembangan dan belum tersedia untuk akses publik. Abstract As the number of user reviews on digital investment platforms continues to increase, sentiment analysis methods are needed to support service quality evaluation. This study aims to compare the performance of Support Vector Machine (SVM) based on Term Frequency–Inverse Document Frequency (TF-IDF) and IndoBERT for sentiment classification of Bibit application reviews, as well as to obtain appropriate hyperparameter configurations for each model. The dataset was collected through web scraping, resulting in 4,123 Google Play Store reviews. After filtering out reviews with a rating of 3, a total of 3,687 reviews remained, consisting of 1,956 positive and 1,731 negative reviews. The study adopts the Cross-Industry Standard Process for Data Mining (CRISP-DM) framework, comprising data understanding, data preparation, modeling, evaluation, and deployment. Optimization was performed through hyperparameter tuning, while evaluation employed Stratified 5-Fold Cross Validation using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix metrics. The evaluation results show that SVM achieved an average accuracy of 87.28%, while IndoBERT achieved 87.25%. The 0.03 percentage-point difference in average accuracy indicates that the performance of both models based on accuracy was very close on the research dataset. Error Analysis was conducted to identify patterns of misclassification, while Explainable Artificial Intelligence (XAI) was applied using LIME for SVM and Token Saliency for IndoBERT to interpret feature contributions to prediction outcomes. The models were subsequently implemented in a Streamlit-based application prototype as an application of the deployment stage in CRISP-DM. The system is currently operated in a development environment and is not yet publicly accessible.]]>
Copyrights © 2026
