Intelligent Generative AI-Assisted Academic Advisement System Using Learning Analytics and Transformer Models

Abstract

A lack of personalisation, delayed intervention, and high student-to-advisor ratios can limit conventional academic advising. This study combines an empirical analysis of student-performance factors with a proposed AI-assisted academic-advisement architecture. The implemented component comprised descriptive and bivariate statistical analyses of a publicly available synthetic dataset containing 6,607 student records and 20 variables. Attendance showed the strongest bivariate linear association with examination score ((r=0.581)), followed by hours studied ((r=0.446)). These results describe relationships encoded in the synthetic data and do not establish causal effects or real-world predictive performance. The proposed architecture comprises five layers: data input; preprocessing and feature engineering; learning analytics and prediction; FT-Transformer and generative AI; and explainability and advisement output. The FT-Transformer design uses feature-specific embeddings and multi-head self-attention to model relationships among numerical and categorical attributes. Huber loss and regularisation are specified as design choices intended to reduce sensitivity to large residuals and mitigate overfitting; their benefits have not yet been experimentally evaluated. Course recommendation, language-based advice generation, explainability, and advisor-feedback mechanisms remain proposed components requiring implementation and validation. The framework is intended for future evaluation in GCC higher education institutions and subsequent adaptation to Arabic-language advising. Prospective studies should assess predictive performance, fairness, explanation quality, usability, and effects on student outcomes.