Cognitive Diagnostic Model Based on Knowledge Relationships and Exercise Value Weights
-
-
Abstract
Existing cognitive diagnosis methods overlooked the explicit and implicit correlations between knowledge concepts in exercises and the value weights of the exercises. To address these issues, a neural network-based diagnostic model (QENCD) was proposed in the paper. The model first employed the Jaccard algorithm to explore the similarity between knowledge concepts, identified the explicit knowledge concepts in exercises, and computed their weights using the feature vector centrality algorithm. Subsequently, it utilized the information from explicit know-ledge concepts as prior information to infer the implicit knowledge concepts in exercises. It then introduced factors of student potential and exercise value weights to reconstruct the interaction function. Finally, combined with the guessing factors, it was predicted whether the students could answer the questions correctly, and their actual levels were assessed. In comparative experiments conducted on four real-world datasets (ASSIST0910, MathEC, Junyi, and Math1), QENCD was compared with three statistical cognitive diagnosis models (IRT, MIRT, DINA) and three neural network cognitive diagnosis models (NCD, CDGK, ICD). Additionally, to validate the effectiveness of each component, ablation experiments were performed by removing different parts of the model (such as the know-ledge matrix and exercise value weights) to assess the impact of each part on the model's performance. In addition, interpretability experiments based on the Degree of Agreement (DOA) metric were conducted to assess the explanatory power of the models. The experimental results show that, across four real-world datasets, the QENCD model outperforms the current best baseline model (ICD) with an average improvement of 0.84% in ACC, 1.12% in AUC, and a reduction of 0.38% in RMSE. The results of the ablation experiments show that different components of the model contribute significantly to its accuracy and interpretability. Based on the DOA values of four models (DINA, NCDM, CDGK, and QENCD) on four real datasets, it is known that the QENCD model can more accurately identify and interpret students' knowledge states.
-
-