MUSE: Multi-Scale Temporal Features Evolution for Knowledge Tracing
Project Overview
The document explores the integration of generative AI in education, particularly highlighting the MUSE model, a transformer-based knowledge tracing system that enhances the prediction of student performance over time by incorporating multi-scale temporal features. Traditional models often struggle with long temporal features and fixed window sizes, but MUSE overcomes these limitations through its innovative dual structure, which includes both local and global modules. This approach not only improves prediction accuracy but also showcases its effectiveness in applications, as evidenced by its performance in the AIEd Challenge 2020. The findings suggest that leveraging such advanced AI models can significantly enhance personalized learning experiences and support educators in identifying students' learning trajectories, ultimately leading to improved educational outcomes.
Key Applications
MUSE: Multi-Scale Temporal Features Evolution for Knowledge Tracing
Context: Online education and knowledge tracing for students during and after COVID-19 school closures.
Implementation: The model integrates local and global temporal features through a transformer-based architecture and uses techniques like adversarial training and random answer masking.
Outcomes: Achieved 5th place in the Riiid AIEd Challenge 2020, demonstrating improved accuracy in predicting student performance.
Challenges: Handling extremely long temporal features and model complexity due to the self-attention mechanism's scaling with sequence length.
Implementation Barriers
Technical barrier
Complexity of self-attention mechanism limits the handling of extremely long temporal features.
Proposed Solutions: Implementation of multi-scale temporal sensor units to capture both local and global features effectively.
Implementation challenge
Existing approaches track knowledge drifts under a fixed window size, which limits adaptability.
Proposed Solutions: Developing a model that can consider different temporal ranges and adapt dynamically.
Project Team
Chengwei Zhang
Researcher
Yangzhou Jiang
Researcher
Wei Zhang
Researcher
Chengyu Gu
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Chengwei Zhang, Yangzhou Jiang, Wei Zhang, Chengyu Gu
Source Publication: View Original PaperLink opens in a new window
Project Contact: Dr. Jianhua Yang
LLM Model Version: gpt-4o-mini-2024-07-18
Analysis Provider: Openai