Explainable XGBoost for Indonesian Hoax Detection under the Electronic Transactions Law

Abstract

The rapid circulation of misleading information in digital spaces creates a need for screening tools that are accurate, transparent, and suitable for human review. This study develops a reproducible Indonesian hoax-detection pipeline and examines whether model explanations can support cautious legal review. The experiment uses a political-hoax text corpus with fixed training, validation, and test splits. The primary classifier combines word- and character-level TF-IDF features with XGBoost, while a frozen-encoder IndoBERT-Lite model is evaluated as a CPU pilot. TreeSHAP summarizes global and local feature contributions, and LIME is used to inspect borderline predictions. On the held-out test set, XGBoost achieved 0.9725 accuracy, 0.9724 macro-F1, 0.9957 ROC-AUC, and a 0.0235 Brier score; the constrained IndoBERT pilot reached a validation macro-F1 of 0.3343 and is not treated as a final benchmark. The most influential features included source and article-genre markers such as “baca juga,” “referensi,” “Kompas,” “Facebook,” and “foto hoaks,” indicating that the model may learn publisher or writing-style shortcuts in addition to claim-related signals. The audit also identified normalized duplicate overlap across the training-validation and training-test splits. The resulting system should therefore support triage, explanation, and documentation by trained reviewers, not serve as a standalone basis for determining the truth of a claim or establishing an Electronic Information and Transactions Law violation.