Implementasi Model Hybrid IndoBERT-LinearSVC untuk Deteksi Spam Judol Obfuscated pada Komentar YouTube

Danang Budiman Hidayat (Universitas Muria Kudus)
Ahmad Abdul Chamid (Universitas Muria Kudus)
Ahmad Jazuli (Universitas Muria Kudus.)

Progresif: Jurnal Ilmiah Komputer
Vol. 22 No. 3 (2026): Juli

Online gambling (judol) spam comments on Indonesian YouTube are disguised using obfuscated text techniques, substituting non-standard Unicode characters designed to bypass string-matching moderation systems. Vocabulary-based Natural Language Processing (NLP) models such as IndoBERT risk representation degradation as tokenizers map non-standard characters to unknown ([UNK]) tokens. This study implemented a hybrid model using a Dual-Track Preprocessing architecture combining IndoBERT [CLS] vectors from normalized text with TF-IDF character N-Gram LinearSVC features from raw Unicode text via sparse matrix concatenation. Experiments on 6,000 YouTube judol comments show the hybrid model achieving Accuracy=0.9978, Precision=0.9956, Recall=1.0000, F1-Score=0.9978 on the test set, identical to standalone IndoBERT and outperforming standalone LinearSVC (F1=0.9944). The hybrid model preserved IndoBERT peak performance while fully closing the 4-comment obfuscated-spam gap missed by LinearSVC, with no performance penalty from fusion. Keywords: Character N-Gram; Gambling Spam Detection; Hybrid Model; IndoBERT; LinearSVC Abstrak Komentar spam judi online (judol) di YouTube Indonesia disamarkan menggunakan teknik obfuscated text, yakni substitusi karakter Unicode non-standar yang dirancang untuk menghindari sistem moderasi berbasis pencocokan string. Model Natural Language Processing (NLP) berbasis kosakata seperti IndoBERT berisiko mengalami degradasi representasi karena tokenizer memetakan karakter non-standar ke token tidak dikenal ([UNK]). Penelitian ini mengimplementasikan model hybrid dengan arsitektur Dual-Track Preprocessing yang menggabungkan vektor [CLS] IndoBERT dari teks ternormalisasi dan fitur TF-IDF karakter N-Gram LinearSVC dari teks Unicode mentah melalui konkatenasi sparse matrix. Eksperimen pada 6.000 sampel komentar judol YouTube menunjukkan model hybrid mencapai Accuracy=0,9978, Precision=0,9956, Recall=1,0000, F1-Score=0,9978 pada test set, identik dengan IndoBERT standalone dan unggul atas LinearSVC standalone (F1=0,9944). Model hybrid mempertahankan kinerja puncak IndoBERT sekaligus menutup seluruh celah 4 komentar obfuscated yang terlewat oleh LinearSVC, tanpa penalti performa dari proses fusi.