<![CDATA[The rapid growth of online news consumption in Indonesia has intensified the challenge of combating fake news, which undermines public trust and threatens social stability. Conventional approaches, including manual verification, are increasingly inadequate to address the scale and speed of digital information dissemination. This study aims to develop an automatic Indonesian fake news classification system using a deep learning framework that integrates Bidirectional Long Short-Term Memory (Bi-LSTM) with Word2Vec embeddings. Unlike many existing fake news detection models that rely on limited validation settings or focus predominantly on English-language data, this work explicitly addresses the linguistic characteristics and practical constraints of the Indonesian context, thereby strengthening model relevance for real-world deployment. The dataset comprises 6,000 balanced news articles, including 3,000 valid items from Detik.com and 3,000 hoax items from Turnbackhoax.id, collected between January and October 2024. Text preprocessing involved cleaning, stopword removal, tokenization, and padding. A 300-dimensional Word2Vec embedding model was employed, and the classifier was trained using stratified 3-fold cross-validation to ensure robust performance estimation. An ensemble inference strategy was further applied to reduce inter-fold variance and enhance generalization on unseen data, directly addressing a common limitation of prior single-model approaches. Experimental results show that the proposed model achieves an accuracy of 86.43% and an F1-score of 86.28%, alongside a high mean Average Precision of 0.927 during validation. Compared with previously reported deep learning baselines, this framework demonstrates competitive yet more stable performance under realistic evaluation settings, supporting scalable deployment.]]>
Copyrights © 2026
