Structural Neural Networks Meet Piecewise Exponential Models for Interpretable College Dropout Prediction

Abstract

We propose a novel approach to address the issue of college student attrition by developing a hybrid model
that combines a structural neural network with a piecewise exponential model. This hybrid model not only
shows the potential to robustly identify students who are at high risk of dropout, but also provides insights
into which factors are most influential in dropout prediction. To evaluate its effectiveness, we compared the
predictive performance of our hybrid model with two other survival analysis models: the piecewise
exponential model and a hybrid model combining a fully-connected neural network with a piecewise
exponential model. Additionally, we compared it to five other cross-sectional models: Ridge Logistic
Regression, Lasso Logistic Regression, CART decision tree, Random Forest, and XGBoost decision tree. Our
findings demonstrate that the hybrid model outperforms or performs comparably to the other models when
predicting dropout among students at the University of Delaware in Spring 2020, Spring 2021, and Spring
2022. Moreover, by categorizing predictors into three distinct groups—academic, economic, and socialdemographic—
we discovered that academic predictors play a prominent role in distinguishing between
dropout and retained students, while other predictors may indirectly influence predictions by impacting
academic variables. Consequently, we recommend implementing an intervention program aimed at
identifying at-risk students based on their academic performance and activities, with additional consideration
for economic and social-demographic factors in customized intervention plans.