Analisis Sentimen Publik Polemik LPDP di Media Sosial X: Perbandingan Machine Learning dan Deep Learning

Iqbal Ananda Rizqy Nasution (Universitas Bina Sarana Informatika)
Muhammad Iqbal (Universitas Bina Sarana Informatika)
Mia Rosmiati (Universitas Bina Sarana Informatika)

Jurnal Nasional Komputasi dan Teknologi Informasi
Vol. 9 No. 3 (2026): Juni, 2026

Abstrak – Program beasiswa Lembaga Pengelola Dana Pendidikan (LPDP) menjadi topik perbincangan masif di platform media sosial X pada periode Februari hingga Mei 2026, dipicu oleh polemik terkait kewajiban pengabdian penerima beasiswa. Tingginya volume percakapan publik yang dihasilkan memerlukan pendekatan otomatis untuk memahami persepsi masyarakat secara efisien dan objektif. Penelitian ini membandingkan performa empat metode yang mewakili tiga generasi pendekatan analisis sentimen, yaitu Logistic Regression dan Support Vector Machine sebagai representasi machine learning klasik, Bidirectional Long Short-Term Memory dengan FastText sebagai representasi deep learning, serta IndoBERTweet sebagai representasi transformer. Sebanyak 1.217 tweet berbahasa Indonesia dikumpulkan melalui Tweet Harvest dengan empat ronde pencarian menggunakan kata kunci terkait polemik LPDP, dan setelah proses audit menghasilkan 1.195 tweet valid yang digunakan dalam pemodelan, kemudian dilabeli secara otomatis menggunakan InSet Lexicon dengan threshold ±3 menghasilkan distribusi 69,68% negatif, 17,17% netral, dan 13,15% positif, serta diseimbangkan menggunakan Random Oversampling. Hasil evaluasi menunjukkan bahwa Bidirectional LSTM dengan FastText mencapai performa terbaik dengan accuracy 79,17% dan F1-score macro 62,77%, diikuti Logistic Regression dengan accuracy 75,73%, SVM dengan accuracy 75,31%, dan IndoBERTweet dengan accuracy 72,08%. Temuan ini mengindikasikan bahwa model deep learning berbasis sekuensial lebih unggul dibandingkan machine learning klasik maupun transformer pada dataset informal berukuran terbatas, kemungkinan disebabkan oleh kemampuan FastText dalam menangani variasi penulisan bahasa informal Twitter berbahasa Indonesia. Kata kunci : Analisis Sentimen; LPDP; Bidirectional LSTM; IndoBERTweet; Media Sosial X;   Abstract – The LPDP scholarship program became a massive topic of discussion on the X social media platform during the period of February to May 2026, triggered by controversy surrounding scholarship recipients' service obligations. The high volume of public conversations generated requires an automated approach to understand public perception efficiently and objectively. This study compares the performance of four methods representing three generations of sentiment analysis approaches, namely Logistic Regression and Support Vector Machine as classical machine learning representatives, Bidirectional Long Short-Term Memory with FastText as a deep learning representative, and IndoBERTweet as a transformer representative. A total of 1,217 Indonesian-language tweets were collected via Tweet Harvest using four rounds of searches with keywords related to the LPDP controversy, then automatically labeled using InSet Lexicon with a threshold of ±3, resulting in a distribution of 69.68% negative, 17.17% neutral, and 13.15% positive, and balanced using Random Oversampling. Evaluation results show that Bidirectional LSTM with FastText achieved the best performance with an accuracy of 79.17% and a macro F1-score of 62.77%, followed by Logistic Regression with 75.73% accuracy, SVM with 75.31% accuracy, and IndoBERTweet with 72.08% accuracy. These findings indicate that sequential-based deep learning outperforms both classical machine learning and transformer models on informal datasets of limited size, likely due to FastText's ability to handle writing variations in informal Indonesian Twitter Language. Keywords: Sentiment Analysis; LPDP; Bidirectional LSTM; IndoBERTweet; Social Media X;