Grouping Students by Behavior Helps AI Diagnose What They Know

Artificial intelligence systems that try to infer what a student actually knows face a stubborn obstacle: most students answer only a small slice of the questions available to them. In real learning scenarios, exercise selection is subjective, so the distribution of student response data is highly imbalanced. Learners gravitate toward simpler items or exercise types they already handle well, while online platforms design their item pools around key, fundamental concepts. The result is that most students engage with mainstream problems and only a minority attempt obscure or challenging ones, leaving some learners with so little response history that models cannot reliably infer their knowledge state.

A team of researchers — Xiaomeng Zhang, Ruxiang Liu, Piao Shi and Qun Ren — has proposed a way to attack this data sparsity problem directly. In a study published in Discover Computing on 27 October 2026, they introduce the Group-Aware Cognitive Diagnosis framework, or GACD, which groups students with similar answering patterns together and then artificially creates additional “peer interactions” for students with sparse data within each group. The idea is to give the model more information to learn from for the very students who have the least of it, without distorting the overall structure of the data.

Cognitive diagnosis itself is a foundational component of intelligent education. It aims to assess students’ cognitive states by analyzing their historical response data, and it serves as the basis for downstream tasks such as personalized exercise and course recommendations. Early models rested on educational measurement theories like Item Response Theory, which treats student ability as a unidimensional latent trait linked to item responses through logistic functions, and DINA, which models proficiency as binary mastery vectors over a Q-matrix that maps exercises to knowledge concepts. These approaches rely on manually designed interaction functions and often fall short at capturing complex, non-linear relationships between what a student knows and how they answer.

Deep learning changed that. Neural networks have progressively replaced handcrafted interaction functions, and later research has mined latent information from students’ daily learning behaviors to enrich learner representations. Graph learning methods have attracted particular attention: one line of work unifies interaction and structural relationships through a multi-layer graph of student-exercise-knowledge relations, another refines cognitive representations by decomposing interaction graphs into response-signal-based subgraphs, and a third streamlines neighborhood aggregation on interaction graphs to enable real-time diagnosis. Yet, as the authors note, most existing cognitive diagnosis models overlook the critical issue of imbalanced response data distribution — the very problem that hurts the students with the fewest recorded answers.

GACD’s architecture has four components. First comes response graph decomposition: the original interaction matrix is split into two disjoint matrices, one containing only correct responses and one containing only incorrect responses, yielding a Right Subgraph and a Wrong Subgraph. This separation matters because correct and incorrect responses reflect distinct behavioral patterns, and learning representations from each independently lets the model capture student mastery patterns from success and failure contexts separately. The same student may even be assigned to different groups in the correct and incorrect views.

The second component is clustering-aware structural augmentation, the heart of the method. The framework first obtains initial feature representations of students and applies the K-means algorithm to partition them into latent clusters, minimizing intra-cluster variance so that students with similar learning behaviors or cognitive states end up together. Within each cluster, it identifies students who lack sufficient diagnostic evidence — those ranked in the bottom 50 percent by interaction density or historical accuracy — and adds edges between these tail students, randomly sampling a peer from the same tail subset to establish a structural link. This conditional edge sampling happens exclusively among low-degree students within behaviorally homogeneous clusters, targeting the sparsity problem directly rather than reinforcing the head-tail imbalance prevalent in educational interaction graphs. The authors argue this is more principled than generic augmentation methods like random edge addition or node perturbation, which modify structures randomly without domain semantics and can introduce noisy connections; here, augmented connections always link behaviorally similar peers.

Third, the framework performs dual-stream representation learning. It adopts LightGCN, a lightweight graph convolutional network, as the backbone encoder, with two parallel encoding streams corresponding to the Right and Wrong views. For each view, embeddings are propagated on both the original graph and the augmented graph. To ensure the augmentation does not distort the underlying semantics, a self-supervised contrastive strategy aligns node representations between the original and augmented graphs by minimizing an InfoNCE loss for both streams. This objective enforces that the augmented graph preserves the intrinsic semantic properties of the original while enriching its structure. The implementation uses sparse adjacency tensors rather than dense matrices, which keeps memory costs manageable for typical educational datasets of thousands to tens of thousands of students — though the authors acknowledge that on ultra-large graphs the number of augmented edges grows linearly with the number of tail students, and neighbor sampling or subgraph batching would be needed, an open direction for future work.

Fourth, the two view-specific representations are fused. Because correct and incorrect interactions represent heterogeneous, and often contradictory, signals about mastery, simple summation or concatenation could cause information interference. GACD instead uses a parameterized fusion mechanism with trainable weight matrices that project the two embeddings into a unified semantic space before integration. A transformation layer then makes the framework model-agnostic: backbones that require inputs aligned with the number of knowledge concepts, such as NeuralCDM, get a linear projection to the right dimension, while latent-vector models like KaNCD receive an identity or dimension-alignment mapping. The transformed student representations are finally fed into the interaction function of the chosen cognitive diagnosis model to predict the probability that a student answers a given exercise correctly, and the whole system trains end-to-end on a joint objective combining prediction loss and the contrastive regularization term.

The evaluation covered three real-world public datasets: Assist2017, collected from the ASSISTments online tutoring platform with mathematics exercise logs; NeurIPS2020, released for the NeurIPS 2020 Education Challenge covering a broad range of K-12 mathematical problems; and Junyi, from the Junyi Academy online learning platform with mathematics interaction logs over a longer period. The datasets vary in scale, knowledge structure and sparsity, and were split 7:1:2 into training, validation and test sets. The authors integrated GACD with backbones from two categories — statistical models like IRT and multidimensional IRT, and neural models including NCD, MFKC and KaNCD — and compared against LightGCN and the ORCDF framework, reporting accuracy, AUC, F1-score and the Degree of Agreement interpretability metric, with all results averaged over five runs using different random seeds. Across every dataset, GACD consistently improved all baseline cognitive diagnosis models, which the authors say demonstrates more accurate cognitive diagnosis under imbalanced response distributions. Compared with ORCDF, whose random response-reversal augmentation strategy may unintentionally reinforce erroneous information in the original data, GACD’s group-homogeneity enhancement avoids random or uncertainty-based operations and more reliably supplements semantically relevant interaction information for data-scarce student nodes.

Ablation experiments using KaNCD as the backbone showed that removing either the clustering-aware augmentation or the response graph decomposition significantly lowered accuracy, AUC and F1-score, with the decomposition proving especially important for interpretability. Hyperparameter analysis found the optimal clustering grouped students into four clusters for Assist2017 and NeurIPS2020 and five for Junyi — a difference the authors attribute to Junyi’s larger student population and lower, more dispersed average node degree, meaning sparser datasets need more clusters so that added edges capture true similarities. They caution, however, that choosing K adaptively without validation data remains an open problem that limits out-of-the-box practicality. Perhaps most tellingly, a stratified evaluation dividing test students into head, middle and tail groups by historical interaction count showed GACD outperforming the strongest baseline with the largest benefits for tail students — precisely those with limited response histories the framework was designed to help.

Subject of Research: A graph augmentation framework that improves cognitive diagnosis of student knowledge states under sparse response data.

Article Title: Enhancing cognitive diagnosis with group-aware graph augmentation and self-supervised alignment

Article References: Zhang, X., Liu, R., Shi, P., & Ren, Q. (2026). Enhancing cognitive diagnosis with group-aware graph augmentation and self-supervised alignment. Discover Computing, 29(1), Article 815. https://doi.org/10.1007/s10791-026-10738-1

Image Credits: AI Generated

DOI: 10.1007/s10791-026-10738-1

Keywords: cognitive diagnosis, intelligent education, graph neural networks, data sparsity, self-supervised learning, latent class analysis, knowledge tracing, K-means clustering, contrastive learning, educational data mining, Enhancing, cognitive