Decentralized AI Accurately Identifies Student Dropout Risks While Preserving Data Locality

Educational researchers evaluating predictive artificial intelligence frameworks have demonstrated that universities can detect students at risk of withdrawing or failing without pooling sensitive personal records into centralized databases. Centralized machine learning systems built to identify academic difficulties traditionally rely on aggregating large volumes of student interaction data from different courses and campuses into a single database. While this aggregated architecture creates accurate warning tools, it introduces acute compliance risks under international privacy statutes such as the Family Educational Rights and Privacy Act, the General Data Protection Regulation, and the Digital Personal Data Protection Act. By conducting systematic comparative experiments across multiple model families on real educational activity logs, investigators revealed that training predictive models strictly on an institution’s local data achieves accuracy and discriminative capability comparable to fully centralized systems, demonstrating that retaining data locality does not inherently degrade early warning performance.

The study addressed an operational friction within learning analytics where modern predictive models demand rich behavioral histories to forecast grades and course non-completion. Educational institutions increasingly track granular student engagement metrics, including digital clickstream actions, learning management system interactions, online attendance, and early assessment submissions, to detect academic struggles before terms conclude. When stored in central cloud repositories, these comprehensive digital footprints become vulnerable to external hacking, unauthorized institutional access, and sophisticated data inference attacks such as membership inference and model inversion. Internal institutional ethics boards and administrative governance frameworks frequently restrict the cross-institutional sharing of raw academic transcripts and activity records. Consequently, university administrators have faced an unresolved dilemma regarding whether retaining data locally within individual institutional boundaries causes a substantial penalty in machine learning accuracy compared to large-scale data centralization.

To rigorously quantify this trade-off, researchers Srinivas Jadhav and Ashok Tadamalle designed an empirical evaluation framework using the publicly accessible Open University Learning Analytics Dataset. Rather than assuming that federated architectures automatically deliver privacy or compliance, the authors established clear architectural boundaries that treated each course module-presentation combination as an independent client entity. Across the examined dataset of 32,593 student enrolments spanning 22 distinct module-presentation clients, each client maintained its own natural student population with varying class sizes and demographic compositions. This configuration simulated the non-independent and identically distributed data environments characteristic of multi-institutional deployments. Throughout the training pipelines, each institutional client retained its underlying records locally, exchanging only mathematical model parameters or localized predictions with a central coordinating server that acted strictly as an orchestration processor rather than a centralized data warehouse.

The experimental setup focused on an early warning milestone fixed at day 70 of course engagement, establishing a practical operational window where advisors still have sufficient instructional time to implement supportive tutoring and academic interventions. Features extracted for each student incorporated three distinct analytical categories: static demographic variables and prior academic qualifications, cumulative engagement metrics capturing Virtual Learning Environment interactions up to the day-70 threshold, and temporal sequence arrays recording daily interaction clicks. The prediction objective was formulated as a binary classification problem where the positive class represented at-risk outcomes, specifically students who formally withdrew or failed to complete their course. To prevent statistical data leakage across simulated client borders, all preprocessing operations, including the one-hot encoding of categorical attributes and numerical feature standardization, were fitted exclusively on a designated 70 percent training split before being applied to the 15 percent validation and 15 percent held-out test sets.

The investigation tested five distinct model families across diverse algorithmic approaches, spanning transparent traditional statistical baselines to complex deep neural networks and gradient-boosted decision trees. Evaluated architectures included regularized logistic regression, multilayer perceptrons incorporating dropout regularization to mitigate overfitting on tabular fields, one-dimensional convolutional neural networks joined with multilayer perceptrons to capture localized engagement bursts, and transformer encoders paired with multilayer perceptrons to process long-range chronological dependencies in daily click activity. For tree-based algorithms, the researchers assessed XGBoost, LightGBM, and CatBoost ensembles. Additionally, the investigators assembled a centralized soft-voting ensemble averaging predictions from CatBoost, the convolutional hybrid, and the transformer hybrid, establishing an empirical upper bound for predictive performance on the benchmark dataset.

To evaluate decentralized deployment methods, the authors deployed Federated Averaging across all gradient-trained neural networks and the logistic regression model, while developing a specialized client ensemble approach for decision tree models. Under standard Federated Averaging, client sites trained localized models over their private student data across several local epochs before transmitting updated parameter weights to the coordinating server, which computed sample-size-weighted parameter aggregations. Conversely, because standard decision tree ensembles do not naturally accommodate gradient averaging across disparate split thresholds, tree models were trained independently within each local client site. Under this client ensemble configuration, student inference was evaluated against localized client models, and aggregated multi-institutional predictions were formed by computing client-size-weighted averages of output probabilities, providing a structured comparison against traditional federated parameter updating.

Evaluation of centralized models confirmed strong discriminative capability across all examined architectures, with performance metrics remaining consistent across five independent experimental runs with randomized seeds. Within the centralized baseline environment, logistic regression reached an Area Under the Receiver Operating Characteristic Curve of 0.862, while tree-based models and neural hybrids achieved AUC values ranging from 0.885 to 0.891. The centralized soft-voting ensemble achieved the strongest overall performance on the test benchmark, attaining an accuracy of 0.808, an at-risk F1 score of 0.804, and an AUC of 0.893. Tree models under centralized conditions consistently produced an accuracy of approximately 0.81, an at-risk F1 score of 0.81, and an AUC of 0.89, illustrating that gradient-boosted architectures handle tabular demographic and behavioral education indicators with high predictive stability.

A critical outcome emerged when contrasting centralized performance against strictly localized institutional training. Models trained purely on their respective local client data without sharing parameters or aggregating updates matched or exceeded centralized counterparts. Most notably, a simple local-only logistic regression model achieved an accuracy of 0.814, an at-risk F1 score of 0.811, and an AUC of 0.892, effectively matching the performance of the complex centralized multi-model ensemble. Tree models trained locally likewise maintained identical metric boundaries around 0.81 accuracy, 0.81 F1 score, and 0.89 AUC. These empirical measurements demonstrate that for structured learning analytics, keeping data strictly inside local institutional boundaries incurs no inherent accuracy loss, challenging the prevailing operational assumption that high-accuracy machine learning always requires pooling institutional records into massive central warehouses.

In contrast, models coordinated through standard Federated Averaging exhibited noticeable performance degradation compared to both centralized and local-only configurations. When trained via Federated Averaging, federated logistic regression dropped to an accuracy of 0.753, an at-risk F1 score of 0.765, and an AUC of 0.805, alongside higher run-to-run variance, including a precision-recall AUC standard deviation of 0.066. Complex neural hybrids similarly saw their AUC drop to between 0.845 and 0.848 under federated updating, while decision tree client ensembles achieved intermediate AUC levels between 0.857 and 0.859. The researchers concluded that the performance deficit observed in federated settings stems not from retaining data locally, but from the mathematical limitations of the FedAvg aggregation mechanism when dealing with skewed, heterogeneous, and non-IID student distributions across course presentations.

The study also examined algorithmic equity across student demographic subgroups, identifying modest but persistent disparities in error rates using the centralized CatBoost model as an analytical benchmark. When analyzing performance across gender classifications, the model yielded a higher at-risk recall of 0.81 for male students compared to 0.70 for female students, even though precision remained nearly identical at 0.87 for both groups. Larger disparities emerged when evaluating students with declared disabilities: while recall was balanced at 0.76 across both cohorts, the false-positive rate for students with declared disabilities reached 0.21, nearly double the 0.12 rate observed for students without declared disabilities. Similarly, students residing in the most deprived Index of Multiple Deprivation bands exhibited higher recall alongside elevated false-positive rates of 0.20, compared to 0.04 in the least deprived bands, indicating that automated flags could subject vulnerable student groups to disproportionate or potentially stigmatizing interventions if decision thresholds are left unmonitored.

The authors outlined clear operational limitations, cautioning that their multi-institutional environment was simulated from a single university dataset by partitioning module offerings, which cannot fully replicate real-world cross-institutional differences in educational policies, computing infrastructure, or curriculum structures. Furthermore, the experiments relied on standard Federated Averaging and did not incorporate formal mathematical privacy guarantees such as differential privacy or secure multiparty aggregation, meaning parameter exchanges could still remain vulnerable to adversarial reconstruction. The governance framework represents an architectural blueprint rather than a deployed compliance solution. Future investigations are planned to explore heterogeneity-resilient federated aggregation algorithms such as FedProx and SCAFFOLD, evaluate genuine cross-institutional consortia, and incorporate post-hoc explainability frameworks such as SHAP to provide transparent, interpretable guidance for institutional advisors.

Subject of Research: Governance-aligned artificial intelligence and federated learning models for predicting student dropout in higher education

Article Title: AI, governance and federated models for student dropout prediction in education

Article References: Jadhav, S., & Tadamalle, A. (2026). AI, governance and federated models for student dropout prediction in education. Discover Artificial Intelligence, 6(1), Article 1787. https://doi.org/10.1007/s44163-026-02724-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02724-9

Keywords: learning analytics, student dropout prediction, federated learning, data governance, logistic regression, machine learning, higher education, Open University Learning Analytics Dataset, FedAvg, privacy compliance, artificial intelligence, CatBoost, neural networks, educational technology, algorithmic fairness, precision education, multimodal AI, predictive modeling