Bayesian network predicts dropout risk in continuing education learners

Adult learners juggling jobs, families and finances are dropping out of online continuing education at rates that have long frustrated educators, and a new study argues that the problem is not a lack of data but a lack of models that can reason transparently about why students leave. Researchers have introduced BN-PIF, a Bayesian Network-Based Probabilistic Inference Framework, which models the conditional relationships among learner attributes in a directed acyclic graph and uses probabilistic inference to flag students at risk of dropping out before disengagement becomes permanent. The framework also maps inferred risk factors to individualized interventions such as academic counselling, peer mentoring or financial aid.

The motivation is straightforward. The study notes that dropout rates in MOOC-based continuing education are commonly reported at 40 to 70 percent, with some estimates exceeding 60 percent, and that traditional machine learning approaches are often opaque and fail to capture the probabilistic and causal interdependencies among the diverse factors influencing adult learners. Financial stress, course involvement and personal drive are among the influencing variables, but black-box classifiers make it difficult for administrators and teachers to understand when and how to intervene. Bayesian networks, by contrast, represent variables as nodes and conditional influences as directed edges, allowing both high-accuracy prediction and meaningful inferences about the structure of dropout risk.

BN-PIF draws on multiple data sources, including demographic surveys, financial records, learning management system activity logs and psychological assessments. The preprocessing pipeline matches datasets through a shared learner identifier, removes duplicates and inconsistent entries, imputes missing numerical values with mean or median methods depending on skewness, fills missing categorical values with the mode, and applies expectation-maximization imputation to psychological survey items with patterned missingness. Outliers in engagement duration and assessment scores are detected with the interquartile range rule. Continuous variables such as engagement duration, assessment scores and behavioral markers are then converted into discrete categories using a Minimum Description Length based supervised discretization approach, which automatically selects cut points from the data distribution rather than relying on manually adjusted thresholds.

Structure learning proceeds in two stages. Mutual information analysis and correlation-based filtering first narrow the search space to statistically significant learner-variable relationships, with a filtering threshold of 0.01. Several algorithms were then tested, including K2, hill-climbing and the constraint-based PC algorithm, and score-based hill-climbing with the K2 scoring metric proved the best compromise between predictive performance, structural stability and interpretability. Hyperparameters were tuned to a maximum of four parents per node, a maximum of 10,000 search steps, a score convergence tolerance of 10 to the power of minus 4, and a tabu list of 50 entries to prevent local cycling. The resulting directed acyclic graph places dropout as the objective node connected to engagement level, financial pressure, academic achievement and motivation, with engagement moderated by learning management system activity and prior performance, and financial stress linked to employment status and scholarship support. Conditional probability tables were estimated by maximum likelihood with Laplace smoothing, and inference runs through exact variable elimination, with posteriors consistent with belief propagation.

The empirical validation used the Kaggle Dropout or Academic Success dataset, which contains 4,424 student records from Portuguese higher education institutions across 33 features including age, marital status, scholarship status, admission score and attendance. The authors are explicit that this dataset is not an MOOC or continuing education program; it serves as a structured validation benchmark, while the framework is designed for adaptive, ongoing educational contexts. Under stratified 10-fold cross-validation, in which the directed acyclic graph was re-learned in each training fold to avoid information leakage, BN-PIF achieved a mean accuracy of 88.9 percent with a standard deviation of 1.2 across folds. Baseline models trailed behind: DPBN reached 82.4 percent, SMOTE-based classification 84.1 percent and MCD-VI 85.3 percent, each with higher variability. The 95 percent confidence interval for BN-PIF, spanning 87.9 to 89.9 percent, does not overlap with the baselines’ intervals, and paired t-tests on fold-wise accuracy confirmed the improvement was statistically significant at p below 0.05. BN-PIF also posted the best mean F1-score of 0.88, indicating balanced precision and recall for identifying dropouts.

Generalization was tested externally on the Open University Learning Analytics Dataset, a collection of anonymized student, course, assessment and virtual learning environment interaction data from the UK Open University covering more than 30,000 registered students, 22 module presentations and over 10.6 million interaction entries. No retraining or parameter estimation was performed on this dataset; it was used only for external testing. In an early two-week prediction task, BN-PIF achieved an F1-score of 0.85, an accuracy of 86.7 percent and a ROC-AUC of 0.91, with a fold standard deviation of 1.4 and a statistically significant gain of 10.9 percent over the best baseline. Across both datasets the framework recorded an area under the curve of 0.92 on the Kaggle benchmark and 0.90 on OULAD, and the fold-wise density and F1 distributions were narrow and closely aligned, which the authors read as evidence of stable performance across structurally different data sources.

Speed matters for early warning systems, and the study reports empirical latency benchmarks rather than theoretical estimates. Running on a Python 3.10 implementation with the pgmpy package on an Intel i7 workstation with 32 gigabytes of memory under Ubuntu 22.04, the inference engine averaged 0.35 milliseconds per single prediction and about 0.18 milliseconds per instance at a batch size of 10,000, with a median response time of 0.32 milliseconds, a 95th percentile of 0.58 milliseconds and a 99th percentile of 0.82 milliseconds across 10,000 repeated calls. Throughput reached roughly 8,500 predictions per second at batch sizes between 5,000 and 10,000, and under simulated concurrent load with 500 users the mean latency stayed below 2 milliseconds, with 95 percent of requests under 5 milliseconds and 99 percent under 8 milliseconds. The authors describe real-time here as the capability to update posterior probabilities with near-zero latency as new evidence arrives, without structural retraining, and note the architecture can be extended to a dynamic Bayesian network if finer temporal slices are needed.

On the intervention side, BN-PIF converts inferred risk into recommendations through a utility-based layer that weighs each candidate action against the causal factors detected for a given learner. High financial pressure may trigger financial assistance, low motivation may prompt academic counselling or peer advice, and a feedback loop updates intervention-effectiveness weights based on whether dropout probabilities actually fall after support is delivered. Simulated intervention analyses, computed as posterior Bayesian updates within the cross-validation scheme rather than live deployments, showed statistically significant reductions in predicted dropout probability for counselling, financial aid, peer mentoring and structured study planning. A sensitivity analysis estimated that financial stress raises dropout likelihood by 0.23 with a 95 percent confidence interval of 0.19 to 0.27, while higher engagement lowers it by 0.31 with an interval of 0.27 to 0.35, and conditional independence tests found a mean conditional mutual information of 0.42 for retained edges against 0.08 for permuted networks.

The authors are careful about causal language. The directed structure of the network allows a causally interpretable representation, they write, but it does not determine exact interventional causal consequences, because the learned relationships are observational and domain-specific rather than derived from randomized treatments or complete structural causal models. Intervention effects are interpreted as probability-changing outcomes under modeled conditions, not as causal effects. They also acknowledge limitations: the framework presupposes conditional independence relations that may not hold amid latent behavioral dynamics, discretization of continuous variables can lose information, structure learning is computationally costly in high-dimensional data and can settle in local optima, and conditional probability tables become unstable with sparse or heavily imbalanced data. The model performs less well on high-frequency clickstream sequence modeling, where deep sequential models such as LSTMs or Transformers may be better suited.

Fairness receives attention as well. Because predictive models trained on historical educational data can encode socioeconomic or demographic bias, the transparent graphical structure allows direct inspection of how protected attributes such as age and gender enter the network. The authors report that engagement and academic performance sit higher in the causal chain leading to dropout, with demographic variables playing a smaller conditional role, and that subgroup comparisons using disparate impact testing showed no consistent differences in predicted risk scores after controlling for engagement patterns, with standardized mean differences below 0.1. Intervention recommendations rest on changeable behavioral and contextual factors rather than demographic proxies, and intervention decisions are not automated. For educators and platform designers, the study’s pitch is that an interpretable, probabilistic early-warning system, validated across two structurally different datasets and fast enough for production use, could shift dropout management from reactive damage control to proactive, targeted support.

Subject of Research: Bayesian network modeling for predicting and intervening on dropout risk among continuing education learners

Article Title: Bayesian network modeling-based prediction and intervention for dropout risk in continuing education learners

Article References: Sun, Y. (2026). Bayesian network modeling-based prediction and intervention for dropout risk in continuing education learners. Discover Internet of Things, 6(1), Article 254. https://doi.org/10.1007/s43926-026-00493-6

Image Credits: AI Generated

DOI: 10.1007/s43926-026-00493-6

Keywords: Bayesian networks, dropout prediction, continuing education, MOOCs, machine learning, probabilistic inference, learner analytics, educational data mining, personalized intervention, interpretability, OULAD, cross-validation