AN EXPLAINABLE DEEP LEARNING FRAMEWORK FOR PERSONALIZED UNIVERSITY MAJOR RECOMMENDATION BASED ON STUDENTS' INTERESTS AND APTITUDES

Abstract

Selecting an appropriate university major significantly influences students’ academic achievement, motivation, and career development. However, conventional academic guidance frequently relies on subjective assessments without adequately integrating students’ interests, aptitudes, and academic performance. This study proposes an explainable deep learning framework for personalized university major recommendations using multidimensional student profiles collected from senior high schools across 20 subdistricts in Tanggamus Regency, Indonesia. The framework integrates Holland’s RIASEC vocational interest theory, psychometric aptitude assessments, and academic achievement indicators. Twenty neural network configurations involving Feedforward Neural Networks (FFNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Bidirectional LSTM (Bi-LSTM) were evaluated using classification metrics. Comparative results identified Bi-LSTM as the best-performing architecture, achieving 91.7% accuracy, followed by LSTM (91.2%), GRU (90.5%), and dropout-based FFNN (90.1%). Furthermore, 84% agreement was observed between students’ selected majors and system recommendations. Nevertheless, learning curves indicated substantial overfitting and limited validation performance, highlighting the need for independent verification. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were incorporated to enhance recommendation transparency. This study contributes an integrated framework combining vocational interests, deep learning, and explainable artificial intelligence to support personalized academic guidance and evidence-based university major selection.