Academic self-efficacy is an important condition to ensure learning effectiveness and improve the sustainability of online learning quality. The present study intended to examine the relationship among online teacher support, academic procrastination, and academic self-efficacy by employing a questionnaire survey and structural equation modeling in the online learning environment. We report on survey data from N = 827 college students from China. The results showed the following: that online teacher support and its emotional, cognitive, and autonomy support dimensions have significant positive effects on academic self-efficacy, and it can predict the academic self-efficacy of college students through the mediating effect of academic procrastination. Therefore, in order to enhance the academic self-efficacy of online learners, it is crucial to increase online teacher support, provide guidance and motivation to online learners, reduce academic procrastination, and improve the emotional, cognitive, and autonomy support perceived by students in online courses and learning activities. The findings contribute to the exploration of the impact and mechanism of online teacher support regarding the academic self-efficacy of learners, providing theoretical and empirical support for enhancing academic self-efficacy in online learning and the development of a sustainable quality of teaching.

Cognitive diagnosis is one of the essential components in intelligent education and aims to diagnose student’s skill or knowledge mastery based on their responses. Recently, with the development of artificial intelligence, some researchers have applied neural network methods to cognitive diagnosis. Although they achieved some success, they seemed to lack a certain basis for designing network structures and could not obtain a unified method for designing network structures. We propose a neural network method for cognitive diagnosis based on Q-matrix constraints, introducing the Q-matrix from traditional cognitive diagnosis to enhance the reliability and interpretability of the network structure. Specifically, our method is highly consistent with generalized deterministic inputs, the noisy “and” gate model (GDINA), and the network structure reflects the direct contribution of skills to answering questions correctly, as well as the indirect contribution of interactions between skills to answering questions correctly. Finally, extensive experiments on both simulated and real datasets demonstrated that our method achieved high accuracy and reliability, with a particularly notable performance on low-quality datasets. As the number of questions and skills increased, our approach exhibited greater robustness compared to the classical methods, highlighting its potential for broad applicability in cognitive diagnosis tasks.