Vol.:(0123456789) SN Computer Science (2020) 1:70 https://doi.org/10.1007/s42979-020-0075-z SN Computer Science ORIGINAL RESEARCH Applying the Deep Learning Method for Simulating Outcomes of Educational Interventions Hyemin Han 1  · Kangwook Lee 2  · Firat Soylu 1 Received: 12 December 2019 / Accepted: 17 February 2020 © Springer Nature Singapore Pte Ltd 2020 Abstract Predicting outcomes of educational interventions before investing in large-scale implementation eforts in school settings is essential for educational policy-making. However, due to time and resource limitations, conducting longitudinal, large-scale experiments testing outcomes of interventions in authentic settings is difcult. Here, we introduce the deep learning method as a way to address this issue and illustrate the use of the deep learning method for the prediction of intervention outcomes through a MATLAB implementation. The presented deep learning method extracts predictable patterns from an empirical dataset to simulate large-scale intervention outcomes. Findings from our simulations suggest that the deep learning applied simulation model can predict intervention outcomes signifcantly more accurately compared to the traditional regression analysis methods. Keywords Deep learning · Machine learning · Computer simulation · Neural network · Educational intervention · Outcome prediction · Policy-making Introduction Diferent educational interventions based on fndings from psychological studies have been developed to promote aca- demic motivation and social adjustment among children and adolescents [1–5]. Some of these interventions have suc- cessfully produced positive long-term, large-scale behavioral efects among diverse populations [6, 7]. Given the invest- ment required for a long-term, large-scale intervention, edu- cators should carefully predict such efects before applying the interventions in the real world. To do so, frst, fndings from laboratory studies need to be replicated in authentic contexts, to examine whether similar intervention efects can be found in real-life contexts. However, limited time and resources make it difcult to conduct such replication stud- ies. Moreover, research ethics is also an issue, since inter- ventions with null or negative efects can signifcantly afect students’ long-term development [8, 9]. Analyzing educational intervention data and interpret- ing results for educational applications present unique challenges. In an educational intervention, there are usu- ally many independent variables across diferent levels. For example, in a school intervention, student characteristics (e.g., academic success measures, demographics, interests, and attitudes), teacher characteristics (e.g., education, expe- rience), and school-related factors (e.g., location, school type, infrastructure) might all be factors afecting imple- mentation outcomes. Given this complexity, mainstream, particularly parametric, data analysis approaches can be prone to statistical errors. To address the aforementioned issues, we introduce a computational method for educators to predict longitudinal This article is part of the topical collection “Deep learning approaches for data analysis: A practical perspective” guest-edited by D. Jude Hemanth, Lipo Wang and Anastasia Angelopoulou. Kangwook Lee and Firat Soylu have contributed equally to this work. * Hyemin Han hyemin.han@ua.edu Kangwook Lee kw1jjang@gmail.com Firat Soylu fsoylu@ua.edu 1 Educational Psychology Program, University of Alabama, Box 870231, Tuscaloosa, AL 35478, USA 2 The School of Electrical Engineering, Korea Advanced Institute of Science and Technology, 291 Daehak-ro, Yeseong-gu, Daejeon 34141, South Korea