Tentang SmartLib Bantuan Pengguna

Skripsi S1

Record Data Skripsi S1

MhswID :

220401046

Judul :

Optimasi Deteksi Berita Hoaks Politik Berbahasa Indonesia Menggunakan Fine-tuning IndoBERT Dengan WeightedRandomSampler

Penulis :

ESTER MULIANA

Abstrak :

Penyebaran berita hoaks politik berbahasa Indonesia meningkat seiring pe-satnya distribusi informasi melalui media digital, sehingga berpotensi memengaruhi opini publik dan stabilitas politik. Penelitian ini bertujuan mengoptimalkan deteksi berita hoaks politik berbahasa Indonesia melalui fine-tuning IndoBERT yang dikombinasikan dengan WeightedRandomSampler untuk menangani ketidakseim-bangan data, dengan memanfaatkan dataset hasil text summarization tanpa mengha-pus data kelas mayoritas seperti pendekatan undersampling pada penelitian acuan. Dataset bersumber dari CNN Indonesia, Kompas, Tempo, dan TurnBackHoax dengan total 31.332 data sebelum preprocessing, dibagi menjadi data latih, validasi, dan uji menggunakan stratified split berkomposisi 70:15:15. Hasil pengujian pada 4.041 data uji menunjukkan performa tinggi, dengan accuracy 97,75%, F1-score macro 0,9683, precision macro 0,9654, recall macro 0,9713, dan ROC AUC 0,9960. Model juga menunjukkan recall 0,96 dan F1-score 0,95 pada kelas Hoaks meskipun proporsinya hanya 22,89% pada data latih, membuktikan WeightedRandomSampler efektif mencegah bias terhadap kelas mayoritas. Dibandingkan penelitian acuan yang menggunakan undersampling dengan akurasi 94,1% dan tingkat kesalahan 5,4% pada 390 data uji, penelitian ini menghasilkan tingkat kesalahan lebih rendah, yaitu 2,25%, pada data uji yang jauh lebih besar. Hasil ini membuktikan bahwa kombinasi fine-tuning IndoBERT dan WeightedRandomSampler mampu meningkat-kan performa deteksi berita hoaks politik berbahasa Indonesia dibandingkan pen-dekatan undersampling. Kata kunci : berita hoaks politik, IndoBERT, Fine-tunning, WeightedRandomSam-pler, ketidakseimbangan data

Abstrak Inggris :

The spread of Indonesian-language political hoax news has increased alongside the rapid distribution of information through digital media, potentially affecting public opinion and political stability. This study aims to optimize the detection of Indone-sian-language political hoax news through fine-tuning of IndoBERT combined with WeightedRandomSampler to handle data imbalance, utilizing a dataset derived from text summarization results without removing majority-class data, unlike the under-sampling approach used in the reference study. The dataset was sourced from CNN Indonesia, Kompas, Tempo, and TurnBackHoax, comprising 31,332 records before preprocessing, which were split into training, validation, and testing sets using a stratified split with a 70:15:15 ratio. Testing on 4,041 test records showed high mod-el performance, with an accuracy of 97.75%, macro F1-score of 0.9683, macro pre-cision of 0.9654, macro recall of 0.9713, and ROC AUC of 0.9960. The model also achieved a recall of 0.96 and an F1-score of 0.95 for the Hoax class, even though this class made up only 22.89% of the training data, proving that WeightedRandomSam-pler effectively prevented model bias toward the majority class. Compared to the reference study, which used undersampling and achieved an accuracy of 94.1% with a 5.4% error rate on 390 test samples, this study achieved a lower error rate of 2.25% on a much larger test set. These results demonstrate that combining IndoBERT fine-tuning with WeightedRandomSampler, using data processed through text summari-zation, improves the detection of Indonesian-language political hoax news compared to the undersampling approach. Keywords: political hoax news, IndoBERT, fine-tuning, WeightedRandomSampler, data imbalance

Tahun :

2026

Cover :

220401046-cover-20261104115340998.pdf

Bab 1 :

220401046-bab1-20261104115343384.pdf

Bab 2 :

220401046-bab2-20261104115343437.pdf

Bab 3 :

220401046-bab3-20261104115343495.pdf

Bab 4 :

220401046-bab4-20261104115343551.pdf