PEMANFAATAN DAN OPTIMASI DEEP LEARNING UNTUK MEMPREDIKSI CAPAIAN KINERJA PUBLIKASI KARYA ILMIAH SUATU PERGURUAN TINGGI
Bagaskara Eka Nugraha, Prof. Ir. P. Insap Santosa, M.Sc., Ph.D., IPU.; Dr.Eng. Ir. Sunu Wibirama, S.T., M.Eng., IPM.
2025 | Tesis | S2 Teknologi Informasi
A good reputation is highly sought after by every Higher education institution. In Indonesia, an institution’s accreditation status serves as a benchmark for its reputation and is often a major consideration for prospective students when choosing a Higher education institution. One of the factors in determining an institution’s “Accredited” status is the academic publication achievements of its faculty. To secure an “Accredited” status in the next period, the ability to predict future research publication performance becomes a crucial aspect of strategic planning for Higher education institution. Although a study has attempted to use deep learning LSTM models to predict individual publication counts, it still has not been able to determine the best deep learning model, particularly for predicting publication counts as a measure of performance achievement.
This sutdy focuses on predicting the research publication performance of a Higher education institution faculty using deep learning machine learning techniques. Three deep learning models, known for their effectiveness in forecasting sequential time series data, are utilized and compared: Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Convolutional LSTM (ConvLSTM). To enhance the accuracy of these deep learning models, this study compares manual hyperparameter tuning methods with Bayesian optimization for each model.
This study uses data on the number of research publications by a particular Higher education institution from 2010 to 2023, which serves as one of the key performance indicators in predicting accreditation status. This data was obtained from SINTA, which aggregates publication data from various databases. The publication data is used to train the deep learning models to predict the number of faculty research publications in the following year. The trained models are then evaluated using Mean Absolute Error (MAE) to determine the prediction accuracy.
The results indicate that the LSTM and BiLSTM models provide the most accurate predictions, with average MAEs of 294 for LSTM and 235 for BiLSTM. Meanwhile, Bayesian optimization for hyperparameter tuning does not significantly improve model performance for BiLSTM and ConvLSTM compared to manual hyperparameter tuning. In contrast to BiLSTM and ConvLSTM, using Bayesian optimization on LSTM results in higher/worse MAEs than manual tuning of LSTM. Among the three models and the hyperparameter tuning methods applied, BiLSTM with Bayesian optimization emerges as the best option for predicting the research publication performance of a Higher education institution.
Kata Kunci : Publika karya ilmiah, Perguruan Tinggi, Deep learning, LSTM, Bayesian optimation