Recurrent Neural Network-Fitnets: Improving Early Prediction of Student Performanceby Time-Series Knowledge Distillation

Autor/inn/en	Murata, Ryusuke; Okubo, Fumiya; Minematsu, Tsubasa; Taniguchi, Yuta; Shimada, Atsushi
Titel	Recurrent Neural Network-Fitnets: Improving Early Prediction of Student Performanceby Time-Series Knowledge Distillation
Quelle	In: Journal of Educational Computing Research, 61 (2023) 3, S.639-670 (32 Seiten)Infoseite zur Zeitschrift PDF als Volltext Verfügbarkeit
Zusatzinformation	ORCID (Okubo, Fumiya)
Sprache	englisch
Dokumenttyp	gedruckt; online; Zeitschriftenaufsatz
ISSN	0735-6331
DOI	10.1177/07356331221129765
Schlagwörter	College Students; Academic Achievement; Prediction; Neurology; Models; Knowledge Level; At Risk Students + Suchen Sie Ihr Suchwort? Collegestudent; Schulleistung; Vorhersage; Neurologie; Analogiemodell; Wissensbasis
Abstract	This study helps improve the early prediction of student performance by RNN-FitNets, which applies knowledge distillation (KD) to the time series direction of the recurrent neural network (RNN) model. The RNN-FitNets replaces the teacher model in KD with "an RNN model with a long-term time-series in which the features during the entire course are inputted" and the student model in KD with "an RNN model with a short-term time-series in which only the features during the early stages are inputted." As a result, the RNN model in the early stage was trained to output the same results as the more accurate RNN model in the later stages. The experiment compared RNN-FitNets with a normal RNN model on a dataset of 296 university students in total. The results showed that RNN-FitNets can improve early prediction. Moreover, the SHAP value was employed to explain the contribution of the input features to the prediction results by RNN-FitNets. It was shown that RNN-FitNets can consider the future effects of the input features from the early stages of the course. (As Provided).
Anmerkungen	SAGE Publications. 2455 Teller Road, Thousand Oaks, CA 91320. Tel: 800-818-7243; Tel: 805-499-9774; Fax: 800-583-2665; e-mail: journals@sagepub.com; Web site: https://sagepub.com
Erfasst von	ERIC (Education Resources Information Center), Washington, DC
Update	2024/1/01