Temporal Intelligence and Algorithmic Equity: A Multi-Phase Framework for Predictive Student Success in Higher Education

Abstract

This literature review presents a systematic comparative analysis of machine learning models for predicting student dropout and academic success, leveraging the widely used Predict Students’ Dropout and Academic Success dataset from the Polytechnic Institute of Portalegre. The core novelty lies in synthesizing and critically evaluating four complementary research paradigms baseline ensemble modeling, phased temporal prediction, fairness-aware ranking under uncertainty, and multi-stakeholder learning analytics to advance a holistic framework for ethical and effective educational intervention. Unlike prior surveys that focus narrowly on algorithmic accuracy, this work emphasizes the interplay between predictive performance, temporal feasibility, class imbalance handling, algorithmic fairness, and institutional integration. A key contribution is the identification of the end of the first semester (S1) as the optimal intervention window, where early academic indicators (e.g., approved curricular units, semester grades) yield the highest predictive power (F1 = 0.745) before data attrition from dropout erodes model efficacy in later phases. The review further contributes by advocating for uncertainty-aware, randomized ranking systems that guarantee stability and multigroup fairness addressing critical ethical gaps in high-stakes educational decision-making. Finally, it calls for a shift from student-centric risk scoring toward multi-level analytics that simultaneously empower learners, instructors, and administrators. This integrative approach offers significant engineering and pedagogical value: it provides institutions with a principled, evidence-based roadmap for deploying robust, interpretable, and ethically grounded predictive systems that not only forecast outcomes but actively enhance students’ academic qualification skills through timely, personalized support.