Indonesian Online News Topic Classification Using Naive Bayes and Support Vector Machine

Abstract

Indonesian online news portals publish large volumes of articles every day, making manual topic grouping time-consuming, inconsistent, and difficult to maintain when articles contain overlapping contexts. Objective: This study develops and compares Naive Bayes and support vector machine models for Indonesian online news topic classification and identifies the model that can be used as the basis for an automatic newsroom support tool. Methods: A dataset of 1,631 Kompas.com articles was collected through web scraping from six channels: economy, automotive, technology, lifestyle, politics, and sport. The article body was used as input, while the source channel became the class label. The text was prepared through duplicate checking, text normalization, case folding, tokenizing, filtering with Indonesian and custom stopwords, and stemming with Sastrawi. The processed texts were transformed with TF-IDF using a maximum of 10,000 unigram and bigram features, min_df=2, and max_df=0.9. The data were split into 80% training data and 20% testing data with stratification. GridSearchCV with five-fold cross validation was applied to tune Multinomial Naive Bayes and support vector machine parameters. Results: Naive Bayes achieved 97.25% testing accuracy, while support vector machine with a linear kernel, C=10, and gamma=scale achieved 98.17%. SVM also produced higher macro precision, recall, and F1-score. Remaining errors mainly appeared in technology articles because their vocabulary overlapped with lifestyle, automotive, and politics topics. Conclusion: TF-IDF with linear SVM effectively classifies Indonesian online news topics and supports automated content organization workflows.