Cognitive diagnosis method via neural networks with transfer learning and Q-matrix constraints

Cognitive diagnostic assessment (CDA) is an important educational assessment method that identifies the strengths and weaknesses of students in specific cognitive skills or attributes. Artificial neural networks (ANNs) can learn complex, nonlinear relationships from data and have become one of the most widely used machine learning methods in CDA. However, most existing ANN-based CDA methods require users to design the network structure manually, which is a challenging task for education professionals without AI expertise. Moreover, neural network training often encounters scarce labeled data, which limits their usability and applicability in cognitive diagnostic practice. Therefore, a simple and easy-to-use general neural network cognitive diagnosis method that can automatically adapt to different datasets and learning tasks is still lacking. In this paper, we propose a neural network cognitive diagnosis method (Bi-QNN) that is constrained by the Q-matrix and an attribute interaction matrix and uses transfer learning for training. Our method has the following advantages: (1) Its network structure can be automatically constructed according to the Q-matrix and interaction matrix corresponding to any dataset, eliminating the need for manual design of the neural network. (2) The network structure design of the new model is inspired by the GDINA model, which can better express and capture the main and interaction effects of attributes. (3) The model training scheme based on transfer learning helps address the scarcity of labeled data, thereby improving the usability and wider applicability of the model. To evaluate the performance of Bi-QNN, we conduct extensive experiments on simulated and real datasets covering various scenarios of CDA. Experimental results show that Bi-QNN has lower prediction errors on the simulated datasets than the parametricmethodsGDINA and DINA, indicating a better fit to the data. Our model is robust to the number of attributes and maintains high classification accuracy as this number increases, demonstrating that Bi-QNN can handle complex problems with more attributes in CDA. The training method based on transfer learning enables Bi-QNN to adapt effectively to datasets with varying sample sizes, maintaining superior performance compared with other models across multiple conditions in simulated and empirical datasets. Bi-QNN generally outperforms other models, suggesting that it can benefit from knowledge transfer and can generalize to new domains. Bi-QNN is a simple, easy-to-use general neural network cognitive diagnosis method with good expressiveness and adaptability. It can provide more accurate and reliable diagnostic feedback for students and teachers and facilitate personalized and adaptive learning. The improvement in model performance is limited by the reliance on simulated data, and the model remains slightly sensitive to the quality of the test items. These issues need to be verified and improved on more datasets.