Enhanced Learning Behaviors and Ability Knowledge Tracing
Knowledge tracing (KT) aims to understand the evolution of students’ knowledge states during learning using machine learning techniques. While KT has made significant strides with deep learning techniques, a gap remains reflecting students’ actual knowledge level—the significant effects of students’...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-01-01
|
| Series: | Applied Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3417/15/2/883 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Knowledge tracing (KT) aims to understand the evolution of students’ knowledge states during learning using machine learning techniques. While KT has made significant strides with deep learning techniques, a gap remains reflecting students’ actual knowledge level—the significant effects of students’ learning behaviors and abilities are omitted, which can reflect their knowledge acquisition more deeply and ensure the reliability of the response. This paper introduces the Enhanced Learning Behaviors and Ability Knowledge Tracing model (LBAKT), addressing this gap by incorporating both learning behavior and ability into its framework. The LBAKT model is structured into four layers: embedding, learning, forgetting, and predicting. The embedding layer enriches exercise information through an exercise embedding unit (capturing knowledge skills, exercises, and difficulty) and a learning ability embedding unit (considering answer time, answers, and hint times), forming a fundamental learning embedding. The learning layer quantifies learning gain by analyzing differences between current and previous learning embedding, interval times, and relative knowledge states, employing a learning gate to assess students’ capacity to absorb knowledge. The forgetting layer models forgetting behavior with forgotten erasing and retention functions, based on previous knowledge states, current learning gains, and interval times. Experimental results demonstrate that LBAKT outperforms existing models in terms of Area Under the Curve (AUC) and Accuracy (ACC) metrics when predicting students’ knowledge states. Specifically, on the longer sequence datasets ASSIST2012 and ASSISTChall, LBAKT achieved an AUC that is 8.8% higher than that of the worst-performing model, demonstrating superior performance. However, the model’s performance on the shorter sequence Algebra0607 was less satisfactory, with an AUC 3.16% lower and an ACC 5.0% lower than that of the top-performing model. Furthermore, our model enhances interpretability by integrating learning behavior and ability, a crucial factor for a deeper understanding of how students acquire and retain knowledge over time. |
|---|---|
| ISSN: | 2076-3417 |