Vocational education generates information in several different forms. A student may be associated with a particular course, complete a sequence of activities, receive scores on assessments and display patterns of engagement over time. In a conventional spreadsheet, these elements are usually stored in separate columns or tables. That format preserves individual values but can obscure the relationships among them. A knowledge graph instead represents information as entities and links: a student is enrolled in a course, a course contains activities, an activity contributes to an assessment and an assessment records a result. Such graphs are especially useful when the meaning of a data point depends on its context. A low score, for example, may mean something different when it follows repeated difficulty with prerequisite activities than when it appears as an isolated result.

The researchers’ approach was designed to address a persistent weakness in educational prediction systems: the tendency to treat tabular variables and contextual relationships as independent sources of information. Their architecture first constructed a knowledge graph from multimodal vocational-education data, with nodes representing students, courses, activities and assessments. It then applied Node2Vec, an embedding method that translates graph structure into compact vectors. Node2Vec performs simulated walks through a network, recording which nodes tend to appear near one another and using that information to position related entities close together in a lower-dimensional mathematical space. The resulting vector does not simply describe a student’s raw score or a course’s title; it encodes patterns of connectivity that may reveal how learners interact with particular content and activities.

Those graph embeddings were combined with tabular attributes before classification by XGBoost, or eXtreme Gradient Boosting. XGBoost builds an ensemble of decision trees sequentially, with each new tree concentrating on errors made by the previous ones. The method is effective for nonlinear relationships and mixed data types, and it can capture interactions that would be difficult to specify manually. In this framework, a learner’s structural position in the knowledge graph could be considered alongside conventional attributes such as assessment-related information. The model’s task was not merely to retrieve existing relationships, but to use the combined representation to predict student outcomes or performance within the educational network.

The reported evaluation metrics were striking across several dimensions. The proposed system reached 0.9898 for accuracy, recall and F1-score, while precision was 0.9899. Accuracy measures the proportion of predictions that are correct overall; precision measures how often positive predictions are correct; recall measures how many of the relevant positive cases the system identifies. The F1-score is the harmonic mean of precision and recall, making it useful when both missed cases and false alarms matter. The model also produced a Matthews correlation coefficient of 0.9848 and a Kappa score of 0.9847. These statistics are intended to account more carefully for agreement and class balance than accuracy alone. An AUC of 0.9999 indicates that the classifier almost perfectly separated the evaluated outcome categories across decision thresholds.

The comparison with baseline algorithms was much less flattering for the conventional approaches tested. K-nearest neighbours achieved an accuracy of 0.5903, while Naïve Bayes reached 0.5344 and logistic regression recorded 0.9466. K-nearest neighbours classifies a case according to the labels of nearby examples in feature space, but it may struggle when the representation does not preserve the underlying educational context. Naïve Bayes assumes conditional independence among features, an assumption that is often unrealistic when courses, activities and assessments are tightly connected. Logistic regression performed considerably better, but its linear decision structure may not capture the more complex patterns represented by graph embeddings and gradient-boosted trees. The results therefore suggest that the advantage came from the hybrid representation as much as from the choice of classifier.

If developed further, such a system could support adaptive learning platforms that modify recommendations as a learner’s progress changes. A graph-based model might help identify activities associated with a concept a student has not mastered, connect that weakness to prerequisite material, or flag a learner whose assessment pattern resembles previous cases requiring additional support. Instructors could potentially receive earlier indications that a student is falling behind, while course designers could examine links between activities and performance across a programme. Because knowledge graphs preserve explicit entities and relationships, they may also offer a more interpretable basis for educational decisions than a prediction generated solely from an opaque numerical model. A teacher could inspect the network paths or associated features contributing to a classification rather than receiving only a probability.

The researchers caution, however, that the headline numbers should not be mistaken for proof that AI can reliably forecast every student’s future. The study’s stated limitations include the scope of the dataset and the generalizability of the results across institutions. A model trained on one collection of vocational-education records may encounter very different curricula, grading practices, student populations or patterns of technology use elsewhere. Near-perfect test performance can also arise when data contain strong signals specific to the evaluation set, or when related records are divided between training and testing in ways that make prediction easier than it would be in real deployment. Independent validation, institution-level testing and prospective studies would be needed to determine whether the model can maintain its performance on genuinely unseen learners and programmes.

The work also highlights a broader technical and ethical challenge in educational AI: prediction is only useful when it leads to appropriate action. A system that identifies a student as high risk could help direct tutoring and resources, but an incorrect classification could influence expectations or limit opportunities. Graphs built from educational data must be carefully governed because they can connect personal performance with courses, behaviours and institutional records in ways that expose sensitive patterns. The authors report that their study did not involve human participants, patient data, animals or experimental procedures requiring ethical approval, and that it received no external funding. For now, the framework is best understood as a promising research architecture rather than a ready-made replacement for teachers. Its most important contribution may be the demonstration that educational prediction can benefit when AI learns not only from what students score, but also from how learners, knowledge and assessment experiences are connected.