Enhancing Neural Cognitive Diagnosis: Q-Matrix Recovery, Task Distraction Sensitivity, and Dynamic Attention Modeling from Synthetic Data

Neural Cognitive Diagnosis Models have shown potential for inferring fine-grained student skill profiles. We extend NeuroCDM to enhance its expressiveness and applicability. First, we address the limitations of idealized Q-matrices: enabling the model to operate under incomplete skill-task mappings, improving robustness to real-world instructional noise. Second, we introduce a task-level parameter (δ) capturing the susceptibility of tasks to performance decline under transient attention difficulties - learned from task features via underlying γ parameters. Third, we model student-attention dynamically through a latent parameter φ capturing temporal variations in focus during task execution. To support training without costly human-labeled data, a validated simulation framework is used that generates realistic assessment datasets. NeuroCDM, trained on this synthetic data, accurately recovers 62,3%-100% of the omitted skill dependencies from the Q-matrix, with less than 20% deviation from the original values, and achieves a mean absolute error (MAE) ranging from 0.0555 to 0.1156. We also estimated the γ parameters with MAE below 0.15 for each. Estimations of the ϕ parameter had MAE below 0.015, furthermore a binary AD prediction was made based on performance - the results are also remarkably correct. Together, these contributions advance the scalability and diagnostic power of neural CDMs in educational contexts.