Educational researchers have introduced a multi-cognitive-level knowledge tracing framework that explicitly differentiates between a learner’s progressing cognitive state and the cognitive requirements imposed by specific academic exercises. Knowledge tracing has traditionally sought to estimate how student understanding evolves across sequential learning interactions and use those evolving states to forecast future performance. However, conventional formulations typically represent concept proficiency as an undifferentiated or binary latent mastery score. To capture how students engage with material across varying depths of understanding, the newly proposed Multi-Cognitive-Level Knowledge Tracing model, or MCLKT, formulates knowledge states across four ordered categories: not mastered, knowing, understanding, and applying. By structuring interactions within a heterogeneous graph framework and applying cumulative-link ordinal estimation, the model captures sequential learning trajectories while distinguishing what an exercise demands from what a learner has demonstrated.
The theoretical motivation for MCLKT addresses fundamental conceptual distinctions in educational measurement and cognitive architecture. In standard diagnostic environments, determining whether an exercise assesses a particular knowledge concept represents a structural incidence relationship. Conversely, the specific cognitive depth required to solve that exercise represents a distinct pedagogical demand. Prior attempts to model cognitive levels often faced reliability challenges, particularly when attempting to classify classroom items into overly granular, adjacent categories derived from frameworks such as Bloom’s taxonomy. To maintain both diagnostic validity and empirical reliability, the authors consolidated cognitive categories into four ordered tiers: not mastered, where basic recall is unverified; knowing, which reflects the ability to recall or recognize basic facts, terms, or procedures; understanding, which encompasses explaining relations and selecting procedures in familiar contexts; and applying, where students deploy concepts to solve concrete or transfer problems.
Architecturally, MCLKT decouples exercise incidence from cognitive demand through the coordinated use of two distinct matrices. A binary matrix, denoted as Q, determines whether an exercise evaluates a given knowledge concept and governs graph connectivity. In parallel, a graded matrix, denoted as P, assigns integer values from zero to three to specify the positive cognitive demand required for each active exercise-concept relationship. Crucially, a zero entry in the P matrix denotes a structural null condition rather than a student failure state, preventing unassessed concepts from being conflated with non-mastery. The model encodes observed student responses and exercise requirements as independent, additive embedding signals. Consequently, an incorrect response does not erase, overwrite, or mischaracterize the underlying cognitive requirement of an attempted problem during temporal feature updates.
To model dependencies across the learning environment, the framework establishes a Student-Exercise-Knowledge heterogeneous interaction graph, designated as the SEK graph. This network incorporates three distinct relational structures operating under different temporal constraints. The student-exercise relation is personalized and constructed strictly causally from historical interaction sequences observed up to the current time step, ensuring that subsequent interactions or target exercises cannot leak into historical representations. The exercise-concept relation is defined structurally by the binary incidence matrix Q, providing normalized pathways for bidirectional message propagation between exercises and concepts. Finally, directed concept dependencies, designated as the KK relation, capture source-to-target prerequisite linkages derived either from expert pedagogical annotation or empirical transition histories from training data.
Temporal state updates within the architecture are governed by a relation-specific propagation mechanism paired with recurrent gating. Upon observing a student interaction, the model aggregates interaction-aware representations across active incidence and concept dependency links. The resulting relation-specific graph messages are filtered through an Erase-Add Gate before being processed by a Gated Recurrent Unit. The Erase-Add Gate determines the proportion of previous memory to suppress or supplement using an unconstrained trainable scaling scalar alongside bounded sigmoid and hyperbolic tangent transformations. Crucially, this recurrent updating cycle executes selectively: only knowledge concepts and exercise nodes directly involved in the current interaction or reached via active graph propagation pathways update their hidden states, while unreferenced concepts retain their preexisting latent states intact.
To map high-dimensional concept hidden vectors into interpretable educational states, the model employs a cumulative-link ordinal state head rather than an unconstrained multi-class classification layer. The concept representation projects onto a one-dimensional latent score that is evaluated against three monotonically increasing trainable thresholds enforced via softplus parameterization. This formulation generates three cumulative probabilities representing the likelihood that a student has achieved at least the knowing, understanding, or applying level for each concept. The individual probability distribution across the four categorical states is directly derived from successive differences between these cumulative probabilities, mathematically guaranteeing that the intrinsic ordinal hierarchy is maintained without arbitrary post-hoc adjustments.
For predicting student performance on subsequent exercises, the framework computes a continuous requirement compatibility metric that directly bridges estimated cognitive states with the demands of the target task. For every concept evaluated by an upcoming exercise, the model selects the cumulative probability corresponding to the exact cognitive tier required by the graded P matrix. These cumulative values are averaged across all evaluated concepts to produce a differentiable requirement compatibility score. This compatibility term is bounded, logit-transformed, scaled by a strictly positive softplus-transformed parameter, and combined with target exercise embeddings, historical exercise trajectories, and individual exercise bias parameters to generate the final response probability. Because compatibility feeds directly into the response prediction equation, the ordinal state head is optimized end-to-end via the primary binary cross-entropy loss without requiring separate cognitive state supervision during training.
The empirical evaluation of MCLKT was conducted across three distinct educational benchmarks: the CCNUC dataset, comprising 50 students, 124 exercises, and 17 concepts from a university programming course; the T19 dataset, derived from the New Zealand fourth-grade TIMSS 2019 assessment and containing 5,003 students, 207 exercises, and 11 concepts; and the MR_sub dataset from MoocRadar, featuring 3,950 students, 213 exercises, and 242 concepts. All experimental trials utilized student-level sixty-twenty-twenty train, validation, and test splits across five matched random seeds to prevent student identity overlap. Baselines included thirteen established models, such as Deep Knowledge Tracing, Bayesian-inspired architectures, attention models, and graph formulations like GKT and GIKT, evaluated either in their original form or augmented with an adapted cognitive-level evaluation interface.
In response prediction benchmarks where concept-to-concept graph structures were matched, the full MCLKT model consistently surpassed the graph knowledge tracing baseline GKT across all three datasets in area under the receiver operating characteristic curve, accuracy, and root mean squared error. In settings where concept dependency graphs were excluded, a variant denoted as MCLKT-w/o-KK demonstrated performance statistically comparable to the strongest baselines on the CCNUC benchmark and across most metrics on T19. Specifically, paired two-sided t-tests against PEBG-KT on CCNUC yielded no statistically significant differences, with p-values of 0.2471 for AUC, 0.2493 for accuracy, and 0.4468 for RMSE. However, on the broader MR_sub dataset, removing explicit concept graphs led to pronounced performance drops relative to models like DKT+, highlighting that explicit relational graphs provide vital inductive bias in complex, high-dimensional concept spaces.
To confirm that the internal states reflected genuine educational progression rather than ungrounded statistical artifacts, the authors evaluated direct state accuracy on a held-out subset of 9,038 valid human annotations from the CCNUC course. Five independent educational experts evaluated historical student interaction records using a structured multi-stage rubric, achieving an inter-rater reliability of 0.849 mean pairwise quadratic weighted kappa. The full MCLKT model achieved a quadratic weighted kappa of 0.4505 ± 0.0496 against these frozen expert labels, substantially outperforming adapted baselines such as GKT-CL, which recorded a kappa of 0.1297 ± 0.0420. Additional behavioral validations, including State-Performance Agreement across all three datasets, demonstrated that higher internal cognitive state assignments systematically correlated with superior subsequent student accuracy under varying exercise demands.
Ablation investigations confirmed that both the structural components and the graded cognitive definitions provided distinct performance benefits. Systematically removing the student-exercise, exercise-concept, or concept-concept relational pathways demonstrated that each relation contributes complementary information, with historical student-exercise paths exhibiting an especially strong influence on larger datasets. Furthermore, replacing the graded P matrix with a binary incidence baseline or randomly shuffling cognitive requirement assignments caused consistent degradation across AUC, accuracy, and error metrics across all datasets. Comparative analyses between the cumulative-link head and an unconstrained four-way softmax classification head revealed that while softmax achieved lower mean absolute error, the ordinal parameterization delivered superior ordinal ranking agreement, stronger behavioral consistency, and enhanced next-response predictive accuracy.
The study establishes that multi-level cognitive states can be effectively inferred from sequential learning logs by integrating causal heterogeneous interaction graphs with ordinal state constraints. While the learned cognitive states showed consistent predictive and behavioral validity across diverse curricula and testing conditions, the authors noted that these model estimates should serve as descriptive diagnostic indicators rather than deterministic guarantees of individual student competence. Future investigations aim to incorporate exercise discrimination and difficulty parameters, transition from static structural graphs to dynamically learnable relational weights, and expand longitudinal instructional datasets to evaluate model generalizability across broader educational environments.
Subject of Research: Heterogeneous graph learning and multi-level cognitive knowledge tracing in educational modeling
Article Title: MCLKT: educational-objective-aware heterogeneous graph learning for multi-cognitive-level knowledge tracing
Article References: Wang, Z., Hou, Y., Liu, X., & Zeng, C. (2026). MCLKT: educational-objective-aware heterogeneous graph learning for multi-cognitive-level knowledge tracing. Journal of King Saud University Computer and Information Sciences, 38(10), Article 1042. https://doi.org/10.1007/s44443-026-01500-0
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01500-0
Keywords: knowledge tracing, heterogeneous graphs, graph neural networks, cognitive modeling, Bloom's taxonomy, educational data mining, machine learning, ordinal regression, student modeling, concept mastery, sequential prediction, Gated Recurrent Unit, TIMSS 2019, MoocRadar, response prediction, inter-rater reliability
News Source: Denise Maddox. (November 20, 2026). Heterogeneous Graph Learning Traces Student Knowledge Across Graded Cognitive Levels. Scienmag.



