Journal on Efficiency and Responsibility in Education and Science

A Systematic Approach to Predicting Students' Academic Performance

2026-02-06T21:43:46+01:00

The rapid expansion of digital learning has generated large volumes of educational data, creating new opportunities to apply machine learning (ML) and data mining techniques to predict student academic performance. This study synthesizes 58 empirical studies that used Decision Trees, Random Forests, Support Vector Machines, Logistic Regression, and Artificial Neural Networks to identify at-risk students and improve educational outcomes.

The review focuses on predictor variables, validation methods, accuracy rates, and performance metrics. Findings suggest that the most effective predictive models combine four categories of variables: demographic factors, academic indicators, digital behavioral features, and psychosocial attributes. Among the algorithms examined, Random Forest and Artificial Neural Networks demonstrated the strongest predictive performance, achieving accuracy rates of 85%–93% across k-fold cross-validation and train-test split validation.

Performance measures such as precision, recall, F1 score, and AUC further confirm the robustness and generalizability of these models. ML-based academic prediction systems can strengthen early warning systems, support data-driven policymaking, and enable personalized learning interventions. The study concludes that combining multidimensional predictors with explainable AI can improve equity, personalization, operational efficiency, and accountability in educational decision-making.

Actionable Learning Analytics

2026-03-25T17:20:38+01:00

Higher education institutions need timely, explainable tools to identify students at risk of low performance on large-scale examinations and to guide targeted academic support strategies. In response to this challenge, this study proposes an explainable machine learning framework to predict undergraduate students' performance levels in Colombia's SABER PRO examination. Using student background variables (e.g., gender, region, school type, parental education, and occupation) and SABER 11 standardised test scores (Critical Reading, Mathematics, Citizenship Skills, Science, and English), we formulate a binary classification problem that distinguishes desirable outcomes (levels 3–4) from non-desirable outcomes (levels 1–2). We benchmark baseline models against non-linear learners, including XGBoost, GLMNET, SVM, DT, and LDA, using a 10-fold cross-validation protocol with systematic hyperparameter tuning. Model performance is assessed through confusion matrices and AUC scores. To support educational decision-making, we complement predictive results with explainability analyses, including global feature importance and individual-level explanations via SHAP, enabling transparent identification of the key drivers behind performance levels. The proposed approach provides actionable learning analytics to guide early academic support, promote responsible and transparent educational decision-making, and improve the likelihood of desirable SABER PRO achievement.

Unpacking the Black Box

2026-03-09T15:27:51+01:00

High student dropout rates remain a significant impediment to achieving the United Nations SDG 4 (equitable education). While Artificial Intelligence (AI) offers robust early risk prediction, the intrinsic black-box nature of high-performing models constrains their transparency. This study designs and investigates a multi-layered Explainable AI (XAI)-based assessment framework to generate actionable insights for student retention. We utilized AutoGluon to construct high-performing multiclass classification models (Graduated, Dropout, or Enrolled) on a higher education dataset. To address the complexity of the AutoGluon-generated models, we employed a hybrid XAI framework that couples global interpretability via a decision tree surrogate model and local interpretability via LIME (Local Interpretable Model-agnostic Explanations). The analysis revealed that models from the Boosting family, particularly XGBoost with bagging level 2, achieved the highest predictive performance (exceeding 0.890 across all metrics). The global analysis demonstrated that academic factors were the primary drivers of prediction, but critical socio-economic factors, such as Tuition fees, also exerted significant influence. Local LIME analysis provided granular, case-specific insights, strongly linking dropout status to first-year academic challenges and to features such as age at enrollment. This integrated XAI approach transforms complex models into an interpretable system, supporting student retention and educational equity (SDG 4).

Machine Learning Predictions of Student Outcomes

2026-02-27T10:05:21+01:00

Persistent disparities in student learning outcomes across Czech municipalities highlight the challenge of ensuring equitable access to quality education. These disparities are not only associated with demographic and economic conditions but also with the responsibility of municipalities and institutions to address structural inequalities. This study applies machine learning and SHAP analysis to predict student learning outcomes across municipalities with extended jurisdiction (MEJs), using demographic, economic, social, and housing indicators. Results highlight the dominant role of educational structure, with the share of people without secondary education and the proportion of younger adults holding college degrees emerging as the most influential predictors. Social and housing stressors, including parental executions, poverty destabilization, and housing allowances, further moderate outcomes, revealing nonlinear threshold effects that refine the explanatory narrative. The combined model achieved an R² of 0.629, confirming that while demographic and educational indicators explain most of the variance, contextual vulnerabilities add interpretive richness by identifying vulnerable subgroups. These findings underscore the dual influence of structural educational attainment and social stressors on student performance, while emphasizing educational responsibility as a key dimension in promoting equity and sustainable development.

Measuring Academic Efficiency in High-Impact Scholarships

2026-03-17T21:24:52+01:00

Evaluating the effectiveness of social support programs in higher education requires moving beyond homogeneous assessments of student performance. This study integrates intersectionality with dynamic efficiency analysis to examine how academic efficiency evolves across diverse student profiles within the Líderes del Mañana full-scholarship program in Mexico. Using a longitudinal dataset of 1,796 students (22,718 student–term observations), we apply a two-stage approach. First, Window Data Envelopment Analysis (DEA) estimates relative academic efficiency over time. Second, Gaussian Mixture Modeling identifies intersectional student profiles based on efficiency trajectories and contextual characteristics. Results reveal five distinct efficiency trajectories. While most students converge toward high-efficiency levels, one cluster exhibits a clear negative efficiency slope, greater variability, and limited institutional alignment, indicating it is a priority for intervention. Other clusters display stable high performance, continuous improvement, or moderate but non-accelerating trajectories. Findings demonstrate that efficiency differences are not explained by single demographic factors but by configurations of social background and institutional context. This study provides a scalable, data-driven framework for aligning equity and efficiency objectives in higher education policy and scholarship programs.

Imbalanced Multi-class Prediction of Student Drop-out and Graduation

2026-03-12T19:26:59+01:00

Student study status prediction, including drop-out and graduation, is a widely studied topic in higher education. Yet, evidence across studies remains difficult to compare due to differences in targets, imbalance treatment, metrics, and validation strategies. This systematic literature review synthesizes 70 peer‑reviewed articles published between 2017 and 2025 that apply machine learning or deep learning to predict study outcomes under class imbalance. Results reveal a strong dominance of binary targets, while multi‑class experiments are relatively rare, though they better reflect institutional categories and expose larger performance gaps across classes. Reported imbalance handling includes data‑level resampling, algorithm‑level class weighting, and ensemble or hybrid designs, but many studies lack sufficient procedural detail. Evaluation practices vary considerably; studies reporting per-class measures and imbalance-aware metrics, such as macro F1 and balanced accuracy, provide more decision-relevant evidence than those relying mainly on accuracy. Validation strategies range from hold‑out and stratified cross‑validation to nested validation, temporal splits, and external testing, shaping the credibility of reported performance for deployment. We propose an integrative taxonomy linking target formulation, imbalance degree, handling strategy, and evaluation design to enhance intervention efficiency through capacity‑aware prioritization, while strengthening responsibility through transparent reporting, defensible validation, and explicit attention to minority class performance.

Academic Productivity Dynamics in Colombian Social Science Programs

2026-03-24T14:23:18+01:00

Assessing academic productivity in higher education is challenging because performance depends on multiple correlated competencies and evolves across heterogeneous regional contexts. This is especially relevant in Colombian Social Science programs, where territorial disparities may mask differences in academic efficiency. This study analyses academic productivity dynamics from 2020 to 2023 using a combined PCA–Malmquist Index approach based on 11,099 observations. First, PCA was used as an unsupervised learning technique to reduce dimensionality and identify latent performance profiles. Second, the Malmquist Index was used to estimate productivity change through technological change (TC), pure technical efficiency change (PECH), and scale efficiency change (SECH). The findings show that the strongest profile was associated with Critical Reading, English, and Written Communication, increasing from 20% to 27% in the final period. Technological change explained 76% of productivity improvements, with Magdalena showing the best performance, while Huila lagged due to lower TC and PECH levels. The results highlight that academic productivity in Colombian Social Science programs is shaped by both educational performance and unequal regional capacity for academic modernization.