MLP-BranchNet: A Deep Learning Architecture with SHAP Explainability for Predicting Student Dropout Rates

Abstract

Student dropout in higher education is a persistent challenge that impacts academic outcomes and institutional performance. Timely identification of at-risk students is crucial for enabling effective interventions. This study introduces MLP-BranchNet: A deep learning architecture with SHAP Explain-ability for predicting student dropout rates. The model utilizes a hybrid branching structure based on Multi-Layer Perceptrons (MLPs), where parallel branches independently learn feature abstractions before being fused into a unified representation. This design enables the model to capture complex, non-linear relationships within institutional datasets effectively. Following comprehensive preprocessing, including data cleaning, normalization, and exploratory data analysis, MLP-BranchNet was trained and evaluated. The model achieved a test accuracy of 68.5% and a loss of 0.6323, indicating good overall performance. While performance metrics such as precision, recall, and F1-score were promising, they also revealed limitations in detecting actual dropout cases, largely due to class imbalance. SHAP (SHap-ley Additive exPlanations) was employed to interpret feature contributions, achieving a confidence level of 74.28%, thereby enhancing the model’s transparency and trustworthiness. The ROC curve analysis yielded an AUC score of 0.51, suggesting the need for further optimization. An interactive dashboard was also developed to provide real-time visualization of predictions, feature attributions, and risk profiling for institutional stakeholders. This work highlights how interpretable deep learning models, such as MLPBranchNet, can support data-driven strategies in education. Future enhancements will include the integration of resampling techniques, ensemble methods, and domain-aware architectures to better address minority class prediction challenges.