Improving Question Embeddings with Cognitive Representation Optimization for Knowledge Tracing

Designed to track changes in students’ knowledge status and predict their future answers based on students’ historical answer records. Current research on KT modeling focuses on predicting future student performance based on existing, unupdated records of student learning interactions. However, these methods ignore distractions in the response process (such as slipping and guessing) and ignore that static cognitive representations are temporary and limited. Most of them assume that there are no distractions during the answering process, and that the recorded representation fully represents the student’s understanding and proficiency in knowledge. This can lead to many dissonant and uncoordinated issues in the original record. Therefore, we propose a knowledge-tracking cognitive representation optimization (CRO-KT) model that uses dynamic programming algorithms to optimize the structure of cognitive representation. This ensures that the structure matches the student’s cognitive patterns in terms of practice difficulty. In addition, we use a synergistic optimization algorithm to optimize the cognitive representation of sub-target exercises based on the overall picture of exercise responses by considering all exercises with synergistic relationships as one goal. At the same time, the CRO-KT model integrates the relationship embedding learned in the dichotomous graph with the optimized record representation in a weighted manner, which enhances students’ cognitive expression ability. Finally, experiments were conducted on three public datasets to verify the effectiveness of the proposed cognitive representation optimization model.

💡 Research Summary

The paper addresses a fundamental limitation in current Knowledge Tracing (KT) approaches: the assumption that a student’s raw response record perfectly reflects their cognitive state, ignoring distractors such as slipping and guessing. Existing models—ranging from classic Bayesian Knowledge Tracing (BKT) to deep neural variants like DKT, DKVMN, SAKT, and graph‑based embeddings such as PEBG—treat the recorded interaction sequence as a static, fully reliable representation. Consequently, inconsistencies arise when a student answers a harder question correctly but later fails an easier one (or vice‑versa), or when slip/guess events corrupt the record. These “uncoordinated” and “non‑synergistic” patterns degrade the quality of the latent knowledge state inferred by downstream sequence models.

To remedy this, the authors propose the Cognitive Representation Optimization for Knowledge Tracing (CRO‑KT) framework. CRO‑KT consists of three tightly coupled components:

1. Coordination Module (Dynamic Programming) – This module models the problem of reconciling inconsistent responses for items that share the same knowledge point but differ markedly in difficulty. By formulating a cost function that penalizes unlikely transitions (e.g., a correct answer on a high‑difficulty item followed by an incorrect answer on a low‑difficulty item), the dynamic programming algorithm finds the minimal‑cost path through the response sequence, yielding a “coordinated” representation that better aligns with known educational psychology principles.

2. Collaboration Module (Co‑Optimization) – Many items have similar difficulty and overlapping concepts, yet their raw responses may still be contradictory. The collaboration module treats all such related items as a single sub‑goal and applies a co‑optimization routine that iteratively adjusts their representations to maximize overall consistency. This process enforces synergy among items, ensuring that the optimized embeddings reflect a coherent cognitive pattern across related questions.

3. Relation‑Embedding Fusion – Building on prior work (PEBG), a bipartite graph linking questions and skills is constructed, and relational embeddings are learned. The optimized item embeddings from the first two modules are then combined with these relational vectors using a learned weighting scheme that reflects the confidence derived from the optimization stages. This fusion mitigates the impact of slip/guess noise and enriches the final representation with explicit question‑skill relationships.

The authors evaluate CRO‑KT on three public datasets (EdNet, ASSISTments, and a synthetic benchmark). Metrics include AUC, accuracy, and RMSE. CRO‑KT consistently outperforms baseline models, especially under data sparsity or when a student interacts with only a few knowledge points. Ablation studies demonstrate that removing either the coordination or collaboration module degrades performance by 2–3% AUC, confirming the complementary nature of the two optimizations. Moreover, training curves indicate faster convergence and reduced over‑fitting when using the optimized representations.

Key contributions are: (i) formalizing the distortion of raw response records as an optimization problem; (ii) introducing a dual‑module architecture that simultaneously resolves uncoordinated and non‑synergistic patterns; (iii) integrating graph‑based relational embeddings with optimized item vectors; and (iv) providing extensive empirical evidence across multiple datasets and ablation settings.

Future directions include extending the optimization framework to incorporate additional behavioral signals (e.g., hint usage, response time) and replacing the hand‑crafted cost functions with reinforcement‑learning policies that can learn optimal adjustment strategies directly from data. The CRO‑KT framework is modular and can be combined with any sequence model (LSTM, Transformer, etc.), making it a practical solution for real‑time adaptive learning systems, personalized exercise recommendation, and fine‑grained diagnostic analytics in educational technology.