Analisis Sentimen Ulasan Pengguna Halodoc Menggunakan TextCNN pada Dataset Tidak Seimbang Berbasis NLP

Fitriyani Fitriyani — 2026-08-31

This study analyzes user sentiment in Halodoc application reviews using a TextCNN architecture on an Indonesian-language dataset with extreme class imbalance. The dataset consists of 2,000 reviews collected from the Google Play Store and manually labeled into three classes: positive, negative, and neutral. The class distribution comprises 1,800 positive reviews (90%), 175 negative reviews (8.75%), and 25 neutral reviews (1.25%). To reduce bias toward the majority class, the TextCNN model was trained using class weighting and evaluated using accuracy, precision, recall, F1-score, and a confusion matrix. The model achieved an accuracy of 93.75% and a weighted F1-score of 0.94, while the macro F1-score was only 0.57. The recall for the neutral class was 0%, indicating that the model was unable to recognize the minority class effectively. The contributions of this study are threefold: an empirical evaluation of TextCNN on Indonesian Halodoc reviews with extreme class imbalance; an analysis of the effect of class distribution using class-level metrics and error analysis; and an identification of the practical implications of the classification results for monitoring digital healthcare service quality. These findings demonstrate that high accuracy alone is insufficient to represent model performance on imbalanced datasets; therefore, strategies such as oversampling, data augmentation, or hybrid approaches should be considered in future research.

Deteksi Cyberbullying pada Komentar Instagram Berbahasa Indonesia Menggunakan Fine-Tuning IndoBERT dengan Analisis Kesalahan Model

Dewi Irma Afriayanti — 2026-08-31

Cyberbullying on social media is difficult to detect automatically because comments may contain informal language, sarcasm, implicit body shaming, and meanings that depend on context. This study evaluates the ability of IndoBERT to detect cyberbullying in Indonesian-language Instagram comments and analyzes the model’s classification errors. The dataset consists of 1,050 comments collected from public Instagram posts and manually labeled into two classes: cyberbullying and non-cyberbullying. The IndoBERT-base-uncased model was fine-tuned and evaluated using a confusion matrix, accuracy, precision, recall, and F1-score. The experimental results show an accuracy of 84.76% and an F1-score of 0.8462. For the cyberbullying class, precision reached 0.9294, while recall was 0.7524, indicating that a portion of bullying comments were still missed as false negatives. Error analysis shows that difficult cases were mainly associated with sarcasm, implicit body shaming, informal language, and ambiguous comments. The contributions of this study are an empirical evaluation of IndoBERT on Indonesian Instagram comments, a clear identification of the research gap between conventional machine-learning approaches and the need for contextual Indonesian-language modeling, and an error analysis that provides practical evidence for improving AI-based content moderation. The findings indicate that IndoBERT has promising performance, but larger datasets, conversational context, and pragmatic language modeling are still required.

Journal of Computer System and Informatics (JoSYC)

Analisis Sentimen Ulasan Pengguna Halodoc Menggunakan TextCNN pada Dataset Tidak Seimbang Berbasis NLP

Deteksi Cyberbullying pada Komentar Instagram Berbahasa Indonesia Menggunakan Fine-Tuning IndoBERT dengan Analisis Kesalahan Model