Electronic Theses and Dissertation

Roblox merupakan salah satu platform permainan daring yang memungkinkan pengguna membuat, membagikan, dan memainkan berbagai konten interaktif. Banyaknya pengguna Roblox di Indonesia menghasilkan beragam ulasan yang memuat pengalaman positif, keluhan teknis, masalah keamanan, serta penilaian terhadap pengalaman bermain. Ulasan pengguna pada Google Play Store dapat dimanfaatkan untuk mengidentifikasi opini dan persepsi pengguna melalui analisis sentimen. Penelitian ini bertujuan membangun dataset ulasan aplikasi Roblox berbahasa Indonesia serta membandingkan kinerja model Support Vector Machine (SVM), IndoBERT, dan IndoBERTweet dalam mengklasifikasikan sentimen negatif, netral, dan positif. Sebanyak 50.000 ulasan dikumpulkan dari Google Play Store menggunakan teknik web scraping. Setelah dilakukan penghapusan data kosong dan duplikat, diperoleh dataset akhir sebanyak 40.298 ulasan yang terdiri atas 19.488 ulasan negatif, 2.471 ulasan netral, dan 18.339 ulasan positif. Data kemudian melalui tahapan prapemrosesan teks, pelabelan sentimen otomatis menggunakan IndoBERT, pembagian data dengan rasio 80:10:10, pelatihan model, dan evaluasi. SVM menggunakan representasi fitur TF-IDF, sedangkan IndoBERT dan IndoBERTweet menggunakan representasi kontekstual berbasis Transformer. Kinerja model dievaluasi menggunakan metrik accuracy, precision, recall, dan F1-Score. Hasil pengujian menunjukkan bahwa IndoBERT memperoleh performa terbaik dengan accuracy sebesar 0,8891, macro precision sebesar 0,7909, macro recall sebesar 0,8679, dan macro F1-Score sebesar 0,8178. IndoBERTweet memperoleh accuracy sebesar 0,8663 dan macro F1-Score sebesar 0,7842, sedangkan SVM memperoleh accuracy sebesar 0,8407 dan macro F1-Score sebesar 0,7349. Hasil tersebut menunjukkan bahwa IndoBERT dan IndoBERTweet memberikan performa klasifikasi yang lebih tinggi dibandingkan SVM pada dataset ulasan Roblox berbahasa Indonesia yang digunakan dalam penelitian ini. IndoBERT menjadi model terbaik berdasarkan nilai accuracy dan macro F1-Score, meskipun seluruh model masih menghadapi kesulitan dalam mengklasifikasikan kelas netral akibat ketidakseimbangan jumlah data dan karakteristik ulasan yang cenderung ambigu.

Roblox is an online gaming platform that allows users to create, share, and play various forms of interactive content. The large number of Roblox users in Indonesia has generated diverse reviews containing positive experiences, technical complaints, security concerns, and evaluations of the overall gaming experience. User reviews on the Google Play Store can be utilized to identify users’ opinions and perceptions through sentiment analysis. This study aims to construct an Indonesian-language dataset of Roblox application reviews and compare the performance of Support Vector Machine (SVM), IndoBERT, and IndoBERTweet in classifying negative, neutral, and positive sentiments. A total of 50.000 reviews were collected from the Google Play Store using a web-scraping technique. After removing empty and duplicate entries, a final dataset of 40.298 reviews was obtained, consisting of 19,488 negative reviews, 2,471 neutral reviews, and 18,339 positive reviews. The data subsequently underwent text preprocessing, automatic sentiment labeling using IndoBERT, data splitting with an 80:10:10 ratio, model training, and evaluation. SVM employed TF-IDF feature representation, whereas IndoBERT and IndoBERTweet utilized contextual representations based on Transformer architectures. Model performance was evaluated using accuracy, precision, recall, and F1-Score. The results showed that IndoBERT achieved the best performance, with an accuracy of 0.8891, macro precision of 0.7909, macro recall of 0.8679, and macro F1-Score of 0.8178. IndoBERTweet achieved an accuracy of 0.8663 and a macro F1-Score of 0.7842, whereas SVM achieved an accuracy of 0.8407 and a macro F1-Score of 0.7349. These results indicate that IndoBERT and IndoBERTweet provided better classification performance than SVM on the Indonesian-language Roblox review dataset used in this study. IndoBERT was identified as the best-performing model based on its accuracy and macro F1-Score, although all models still encountered difficulties in classifying neutral reviews due to class imbalance and the ambiguous characteristics of neutral expressions.