Implementasi Automated Text Classification Pada Repositori Digital: Studi Kasus Fakultas Komunikasi Dan Informatika Universitas Muhammadiyah Surakarta

Digital repositories in universities such as the University of Muhammadiyah Surakarta (UMS) store thousands of documents of students' scientific works that need to be classified so that they can be mapped effectively and more accessible. The classification process, which has been carried out manually, has caused low objectivity and inconsistency in subject labeling due to differences in uploader perception. This study develops a text classification model for determining the subject of thesis documents in the archives of scientific papers of the Faculty of Communication and Informatics UMS, which has the characteristics of multilabel, uneven data distribution, and the use of mixed languages (Indonesian and English). The research dataset consisted of 2,810 documents divided into 30 main subjects with data sharing of 60% training, 20% validation, and 20% testing. The study compared three algorithms, namely Support Vector Machine (SVM) with TF-IDF, Complement Naïve Bayes (CNB), and Bidirectional Long Short-Term Memory (Bi-LSTM) with GloVe. The experimental results showed that SVMs with LinearSVC gave the best performance after threshold tuning (T=-0.330), achieving F1-Macro scores of 0.2038, F1-Micro 0.4100, Precision of 0.1372, Recall of 0.4627, and Hamming Loss of 0.1040 in the test set. The best models have been integrated into Streamlit-based prototypes to provide subject recommendations in real-time. This study proves that the traditional SVM model is more effective than deep learning in handling medium-sized academic datasets with extreme data imbalances.