Penyebaran berita hoaks politik di Indonesia meningkat signifikan, dengan 237 konten
hoaks politik teridentifikasi pada tahun 2024. Penelitian ini mengembangkan model
klasifikasi berita politik berbahasa Indonesia menggunakan algoritma Support Vector
Machine dengan representasi fitur Term Frequency-Inverse Document Frequency dan
metode Local Interpretable Model-Agnostic Explanations untuk meningkatkan
interpretabilitas. Data penelitian menggunakan 4.155 berita dari CNN Indonesia dan
Turnbackhoax melalui web scraping tahun 2024. Metode Cross-Industry Standard
Process for Data Mining diterapkan dengan preprocessing meliputi case folding,
cleaning, tokenizing, stopword removal, dan stemming. Pengujian dilakukan pada empat
kernel SVM dengan tiga skenario pembagian data. Hasil menunjukkan kernel linear pada
pembagian data 80:20 menghasilkan akurasi 98,56%, precision 98,52% untuk kelas valid
dan 98,63% untuk kelas hoaks, serta recall 99,25% untuk kelas valid dan 97,30% untuk
kelas hoaks. Penerapan LIME berhasil mengidentifikasi kata pembeda, dimana berita
hoaks didominasi kata media sosial seperti telusur, unggah, youtube, dan tiktok,
sedangkan berita valid menggunakan istilah formal seperti persen, rabu, politik, dan
hadir.

The spread of political hoax news in Indonesia has increased significantly, with 237
political hoax contents identified in 2024. This research develops a classification model
for Indonesian political news using the Support Vector Machine algorithm with Term
Frequency-Inverse Document Frequency feature representation and Local Interpretable
Model-Agnostic Explanations method to improve interpretability. The research data used
4,155 news articles from CNN Indonesia and Turnbackhoax collected through web
scraping in 2024. The Cross-Industry Standard Process for Data Mining method was
applied with preprocessing stages including case folding, cleaning, tokenizing, stopword
removal, and stemming. Testing was conducted on four SVM kernels with three data split
scenarios. Results showed that the linear kernel with 80:20 data split achieved 98.56%
accuracy, 98.52% precision for valid class and 98.63% for hoax class, and 99.25% recall
for valid class and 97.30% for hoax class. LIME implementation successfully identified
distinguishing keywords, where hoax news was dominated by social media terms such as
telusur, unggah, youtube, and tiktok, while valid news used formal terms such as persen,
rabu, politik, and hadir.