Pengembangan Explainable Model Berbasis Transformer Untuk Deteksi Komentar Judi Online Dan Kripto Yang Lebih Robust

Kautsar, Nur Ghulam Musthafa Al (2026) Pengembangan Explainable Model Berbasis Transformer Untuk Deteksi Komentar Judi Online Dan Kripto Yang Lebih Robust. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221126-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Maraknya komentar promosi judi online dan aset kripto di kolom komentar YouTube sulit dikenali metode keyword filtering karena istilah dan nama situs yang terus berubah. Model klasifikasi deep learning belum teruji ketahanannya terhadap komentar hard negative, yaitu komentar normal bergaya bahasa promosi, serta bersifat black box. Penelitian ini merancang model klasifikasi transformer yang tahan terhadap hard negative, menerapkan Explainable AI (XAI) untuk transparansi model, dan mengimplementasikannya ke ekstensi peramban. Dataset dibangun dari data Kaggle, scraping enam video YouTube, dan data hard negative sintetis dari ChatGPT, menghasilkan 9.799 data. Tiga model transformer (IndoBERT, mBERT, XLM-RoBERTa) di-fine-tune dan dibandingkan pada data hold-out dan hard negative, dilanjutkan hyperparameter tuning dan integrasi LIME. Ketiga model kompetitif pada data hold-out (akurasi 0,9862–0,9923), namun IndoBERT paling tangguh pada hard negative (0,936), mengungguli mBERT (0,754) dan XLM-RoBERTa (0,77), sehingga dipilih sebagai model utama. Kombinasi hyperparameter optimal (learning rate 3e-05, batch size 16, epoch 5, dropout 0,1) menghasilkan akurasi normal 0,9883 dan hard negative 0,990. Integrasi LIME memvisualisasikan kata-kata yang memengaruhi keputusan klasifikasi, mengungkap kemampuan model mengenali konteks kalimat sekaligus bias baru akibat data hard negative. Model terbaik diimplementasikan ke ekstensi peramban Chromium yang memindai dan menyembunyikan komentar spam otomatis. Penelitian ini membuktikan model transformer dengan data hard negative dan XAI dapat mendeteksi komentar promosi judi online dan kripto secara akurat, transparan, dan dapat diimplementasikan pada lingkungan nyata.
===========================================================================================================================================
Online gambling and cryptocurrency promotion comments on YouTube evade keyword filtering due to constantly changing terms and site names. Deep learning models remain untested against hard negative comments, normal comments styled like promotional ones, and are black box. This research designs a transformer-based model robust to hard negative comments, applies Explainable AI (XAI) for transparency, and implements it as a browser extension. The dataset combines Kaggle data, scraping from six YouTube videos, and synthetic hard negative data from ChatGPT, totaling 9,799 entries. Three transformer models (IndoBERT, mBERT, XLM-RoBERTa) were fine-tuned and compared on hold-out and hard negative data, followed by hyperparameter tuning and LIME integration. All three performed competitively on hold-out data (0.9862–0.9923), while IndoBERT was most robust on hard negative data (0.936), outperforming mBERT (0.754) and XLM-RoBERTa (0.77), and was selected as the main model. The optimal combination (learning rate 3e-05, batch size 16, epoch 5, dropout 0.1) yielded 0.9883 normal accuracy and 0.990 hard negative accuracy. LIME integration visualized words influencing classification decisions, revealing both contextual understanding and new biases from hard negative data. The best model was implemented into a Chromium browser extension that automatically scans and hides spam comments. This research shows that a transformer model trained with hard negative data and equipped with XAI can accurately and transparently detect gambling and crypto comments in real-world settings.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Transformer, Klasifikasi, Hard Negative, Explainable AI, Ekstensi Peramban. Transformer, Classsification, Hard Negative, Explainable AI, Browser Extension
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Nur Ghulam Musthafa Al Kautsar
Date Deposited:	30 Jul 2026 01:55
Last Modified:	30 Jul 2026 01:55
URI:	http://repository.its.ac.id/id/eprint/140149

Actions (login required)

View Item