Interpretable Deep Learning for University Dropout Prediction

Abstract

The early identification of college students at risk of dropout is of great interest and importance all over the world, since the early leaving of higher education is associated with considerable personal and social costs. In Hungary, especially in STEM undergraduate programs, the dropout rate is particularly high, much higher than the EU average. In this work, using advanced machine learning models such as deep neural networks and gradient boosted trees, we aim to predict the final academic performance of students at the Budapest University of Technology and Economics. The dropout prediction is based on the data that are available at the time of enrollment. In addition to the predictions, we also interpret our machine learning models with the help of state-of-the-art interpretable machine learning techniques such as permutation importance and SHAP values. The accuracy and AUC of the best-performing deep learning model are 72.4% and 0.771, respectively that slightly outperforms XGBoost, the cutting-edge benchmark model for tabular data.

Modeling Learners to Early Predict Their Performance in Educational Computer Games

“…Consequently, based on all three performance measures and their rather low standard deviation (computed from the 10-model accuracy), it can be said that the deep learning model can robustly predict the final score of learners when they complete 30% to 70% of their solutions. Our findings are consistent with those reported by Liu et al [34], Hajra et al [35], and Baranyi et al [36] in that deep neural networks could outperform traditional machine learning techniques in prediction tasks of learner performance. While like the work reported by Chen and Cui [37] and Okubo et al [38], we found that longer sequences of learner's actions could improve the prediction power of deep learning models, our results show that they may not necessarily be able to provide the optimal predictive performance.…”

Fairness Testing: A Comprehensive Survey and Analysis of Trends

“…On the one hand, DL becomes increasingly pervasive, being used in a wide range of software applications and thus attracting interest from the research community. On the other hand, compared to traditional ML algorithms (e.g., regression and decision tree), DL is less interpretable [228], making it harder to reason about fairness in a direct manner.…”

Predição da evasão estudantil: uma análise comparativa de diferentes representações de treino na aprendizagem de modelos genéricos

“…Como já mencionado, muitos trabalhos direcionam a predic ¸ão de evasão a um estágio específico da trajetória acadêmica dos alunos, em geral, ao momento da primeira matrícula, sem considerar indícios do comportamento do estudante no curso atual, como em [Nagy and Molontay 2018] e [Baranyi et al 2020]; ou após um número fixo de períodos cursados, considerando, neste caso, dados que caracterizem o desempenho inicial do aluno, como em [Costa et al 2020] e [Yu et al 2021]. Em comum, estes trabalhos costumam representar os dados de cada estudante a partir de um único registro, já que o momento de predic ¸ão é fixo e, assim, sabe-se exatamente a estrutura de atributos a ser considerada.…”