TLCD: Deep Transfer Learning Framework Enables Cross-Disciplinary Cognitive Diagnosis with Improved Data Efficiency

The increasing prevalence of online education generates vast amounts of student data, creating opportunities to accurately assess individual learning needs through cognitive diagnosis, but also presenting challenges in extracting meaningful insights. Zhifeng Wang, Meixin Su, and Yang Yang from Central China Normal University, alongside Chunyan Zeng from Hubei University of Technology and Lizhi Ye, address this problem by introducing a novel framework, TLCD, for cross-disciplinary cognitive diagnosis. This research overcomes the difficulties of limited data and differing knowledge structures between subjects by combining deep learning with transfer learning, allowing the model to leverage common features across disciplines. The team demonstrates that TLCD significantly improves the accuracy of assessing student learning in new subjects, representing a substantial advance in personalised education and intelligent tutoring systems.

The study pioneers a technique that captures complex patterns in student learning processes by employing deep learning’s robust feature learning capabilities, allowing for detailed insights into knowledge mastery and learning strategies. Researchers trained deep learning models to extract intricate relationships from student data, moving beyond traditional methods that rely on simpler, macro-level assessments of ability. To overcome the challenge of limited data within specific disciplines, the team implemented transfer learning strategies, utilizing common features identified in a primary discipline to enhance model performance in a target discipline.

The method addresses data disparities by efficiently utilizing resources across subjects, reducing the need for extensive data collection and processing. Extensive experiments validated the effectiveness of TLCD, utilizing a comprehensive dataset to evaluate its performance. The study demonstrates that the deep learning-based model outperforms basic cognitive diagnostic models in cross-disciplinary tasks, providing a more accurate evaluation of student learning situations. Researchers rigorously tested the model’s ability to generalize across different subjects, confirming its reliability and potential for widespread application in personalized education.

Deep Learning Diagnoses Student Learning Processes

This research delivers a breakthrough in cognitive diagnosis, introducing a cross-disciplinary method based on deep learning and transfer learning to accurately assess student learning. Scientists developed a model that effectively captures complex patterns in individual learning processes, moving beyond traditional methods that rely on manual annotation and struggle with high-dimensional data. The team successfully utilized deep learning’s robust feature learning capabilities to gain deeper insights into student learning behaviors, knowledge mastery, and strategies. To address the challenge of data scarcity across different disciplines, the researchers implemented transfer learning strategies, enhancing model performance in a target discipline by leveraging common features identified in a main discipline.

Experiments were conducted using a dataset comprising students from YNEG high school, spanning eight different subjects. The core of this work lies in a neural network cognitive diagnostic model, NeuralCD, which models complex student interactions during problem-solving, considering student traits, question relevance, and interaction functions to calculate the probability of a correct answer and rate student proficiency in key knowledge concepts. This model, implemented with continuous vector representation, represents a significant advancement in accurately gauging student understanding across multiple subjects and provides strong support for personalized and targeted education.

Transfer Learning Improves Cognitive Diagnosis Accuracy

This research presents a novel cross-disciplinary cognitive diagnosis method, TLCD, which leverages deep learning and transfer learning strategies to improve the assessment of student knowledge across different academic subjects. By utilizing common features between disciplines, the team successfully enhanced model performance in evaluating student understanding, demonstrating improved accuracy in diagnosing knowledge mastery compared to basic models. Specifically, the approach achieved accuracy rates of 63. 08% for English, 62. 5% for history, 70.

83% for politics, and 63. 33% for another political subject, indicating effective diagnostic capabilities. The study successfully applied this method to a dataset of high school student exam results, providing a comprehensive diagnosis of knowledge across multiple disciplines. While the results demonstrate a clear improvement in cross-disciplinary cognitive diagnosis, the authors acknowledge that further research could explore the impact of different transfer learning strategies to identify the most effective approach for optimizing model performance and enhancing its applicability in diverse educational settings.