Klasifikasi Berita Hoaks Berbahasa Indonesia Menggunakan TF-IDF dengan Jumlah Fitur Tinggi dan Algoritma K-Nearest Neighbors sebagai Model Baseline

Abstract

This study aims to develop an automatic classification model to distinguish fake news and valid news in Indonesian-language texts using the Term Frequency–Inverse Document Frequency (TF-IDF) method and the K-Nearest Neighbors (KNN) algorithm. This research is positioned as a baseline study to evaluate the effectiveness of using TF-IDF representations with a relatively large number of features in the KNN algorithm, as studies that specifically address this aspect in Indonesian fake news classification remain limited. The dataset consists of 500 news documents that undergo preprocessing stages, including text cleaning, character normalization, punctuation removal, and stopword elimination. Numerical document representations are generated using TF-IDF with a maximum of 5,000 features. The dataset is then divided into 400 training samples and 100 testing samples before training the KNN model using standard parameters. The evaluation results show that the proposed model achieves an accuracy of 78%, with balanced precision, recall, and f1-score across both classes. The confusion matrix indicates that most documents are correctly classified, although some misclassifications still occur due to lexical and structural similarities between fake news and valid news texts. In the context of fake news detection, this level of accuracy is considered sufficiently effective as an early detection support system, as the model is able to limit False Negative errors to a relatively low level, allowing the majority of fake news to be successfully identified. These findings demonstrate that the combination of TF-IDF and KNN represents a simple, efficient, and viable baseline approach for developing Indonesian fake news detection systems.