AI‑Driven Student Performance Prediction in Higher Education Using Multi‑Source Educational Data

Abstract

Predicting how students will perform remains a central challenge in educational data mining and learning analytics. Anticipating difficulties early allows instructors to refine feedback and plan interventions to keep learners on track. Our study describes an AI driven framework for predicting academic success using a multi layered dataset collected at the University of Petra's IT College in the 2022 2023 year. The data contain 952 anonymised records and integrate five complementary data layers: registration details, aggregated course-level intended learning outcomes (CILOs), Learning Management System (LMS) behaviours data, placement test scores and program-level intended learning outcomes (PILOs). We outline a rigorous preprocessing and feature engineering pipeline and benchmark eight traditional machine learning classifiers against two neural network models. Class imbalance is mitigated by random oversampling, and hyper parameters are tuned systematically. Gradient boosted tree ensembles deliver the strongest results, achieving around 85 % accuracy and an area under the curve of approximately 0.96, while long short term memory and convolutional models perform slightly worse. TreeSHAP analyses reveal that course outcomes and placement tests are the most influential predictors. We situate these results within the literature and discuss implications for timely interventions and fair, interpretable analytics. The study’s contributions include (A) the integration of five complementary educational data layers and (B) the explicit modelling of CILOs and PILOs in student-performance prediction.