Analisis Komparatif MLP dan GraphSAGE dalam Deteksi Bot Twitter/X pada Benchmark TwiBot-22

Mochammad Fikri Chaerul Chalik Ramdhan (Sekolah Tinggi Teknologi Terpadu Nurul Fikri, Depok)
Sigit Puspito Wigati Jarot (Sekolah Tinggi Teknologi Terpadu Nurul Fikri, Depok)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 6 No 12 (2026): May 2026

Bot accounts on Twitter/X remain a significant challenge because they affect information integrity, distort public discourse, and complicate platform moderation. This article evaluates two bot detection approaches on the TwiBot-22 benchmark: a profile-feature-based Multilayer Perceptron (MLP) and a graph-based GraphSAGE model, using a 12-Stage Evaluation Framework that covers data validation, feature engineering, model training, threshold analysis, feature ablation, and multi-seed evaluation. The study is limited to an offline benchmark setting with 1,000,000 labeled accounts, 13.99% bots and 86.01% humans, and a fixed split of 70% training, 20% validation, and 10% testing. In the single-seed 15-feature comparison, MLP achieved F1(bot) of 0.53 and PR-AUC of 0.48, while GraphSAGE reached F1(bot) of 0.53 and PR-AUC of 0.46. In the confirmatory three-seed evaluation, the user_only_8 configuration produced F1(bot) of 0.53 and PR-AUC of 0.49 with lower variance, whereas all_15 produced F1(bot) of 0.53 and PR-AUC of 0.47 with higher variance. These findings indicate that the more economical profile-only configuration preserves practically identical binary-decision quality, offers better probability ranking quality, and shows lower variance. The main contribution of this article is a feature-economy argument: on TwiBot-22, added graph and feature complexity does not automatically yield proportionate practical gains.