| dc.description.abstract | Spam comments promoting online gambling are increasingly prevalent in YouTube live-stream chats, degrading interaction quality and contravening Article 27(2) of Indonesia’s Electronic Information and Transactions Law (UU ITE). This study develops an automatic detector for such spam by employing a Gated Recurrent Unit (GRU) architecture with IndoBERT word representations. The dataset comprises comments from YouTube (government and news live streams) and X (formerly Twitter) collected over March 2025 to September 2025. After going through preprocessing stages that include tokenization, case folding, stemming, and stopword removal, 42,628 rows of data have been collected that are ready for use. The model architecture includes three core layers (GRU, dropout, and dense) and was tuned via grid search across 216 hyperparameter combinations. The best configuration uses a single GRU layer with 100 units, dropout of 0.5, a learning rate of 0.0001, and 43 epochs. Evaluation on a test set of 4,263 samples using a confusion matrix framework yielded 97.65% accuracy, 98.15% precision, 97.13% recall, and a 97.6% F1-score. To validate practical utility, the detector was deployed as a YouTube-API-based moderation bot and integrated with a Next.js dashboard for results visualization. Live trials during streaming sessions demonstrated stable real-time removal of spam comments. The findings indicate a substantive contribution to improving the quality of live-stream interactions while supporting efforts to curb the dissemination of illegal content. | en_US |
