<![CDATA[Student dropout prediction supports early warning and institutional intervention, but class imbalance can obscure errors affecting students at risk. This review aims to synthesize imbalance-handling strategies and identify gaps in their representation, validation, and evaluation in educational prediction. A PRISMA-guided systematic review of Scopus-indexed studies published from 2020 to 2026 combined narrative synthesis with complementary bibliometric mapping using VOSviewer. A total of 49 studies were included. Conventional resampling was the most frequently coded imbalance-handling theme, while algorithm-level balancing and threshold optimization received less attention. Temporal and sequential approaches represented academic histories and learning activities, and generative methods were explored for student-data augmentation. The included studies did not establish a consistently superior technique across datasets and evaluation protocols. Reporting gaps concerned quantitative class distributions, training-only preprocessing and resampling, advanced minority-class metrics, reproducibility, and validation of synthetic records. The bibliometric map contextualizes these themes without establishing comparative effectiveness. The synthesis supports clearer outcome definitions, conventional baselines, validation matched to the prediction setting, and minority-class evaluation. Where synthetic data are used, fidelity, predictive utility, privacy, and domain consistency should be assessed together. For informatics and computer science, the review contributes a structured account of how data representation, imbalance handling, and validation jointly affect the assessment of predictive systems, helping researchers distinguish credible methodological improvement from gains attributable to incomplete evaluation.]]>
