Cognitive diagnostic assessment: A Q-matrix constraint-based neural network method.

Area of Science:

• Educational Psychology
• Artificial Intelligence
• Machine Learning

Background:

• Cognitive diagnosis models assess student skills but are sensitive to model choice and sample size.
• Developing adaptable models for diverse assumptions and data sizes is challenging.
• Artificial neural networks show promise for cognitive diagnosis.

Purpose of the Study:

• To propose a Q-matrix constrained neural network (QNN) for cognitive diagnosis.
• To develop an unsupervised training framework (SOM-NN) for the QNN to reduce human effort.
• To evaluate the QNN and SOM-NN's effectiveness and interpretability.

Main Methods:

• Developed a QNN using a Q-matrix to define neural network architecture.
• Implemented a self-organizing map-based neural network (SOM-NN) for unsupervised QNN training.
• Conducted experiments on simulated and real datasets to assess performance.

Main Results:

• The proposed QNN and SOM-NN demonstrated effectiveness in accuracy and interpretability.
• Unsupervised QNN achieved significant advantages on small sample datasets with high guessing/slipping rates.
• The approach showed particular strength in pattern-wise agreement rates.

Conclusions:

• The QNN and SOM-NN offer a robust solution for cognitive diagnosis, bridging psychometrics and machine learning.
• This work provides a practical reference for classroom assessment and adaptive learning systems.
• The unsupervised approach is particularly beneficial for challenging datasets and the cold start of personalized systems.