# Domain Adaption for Knowledge Tracing

**Authors:** Song Cheng, Qi Liu, Enhong Chen  
**arXiv:** [2001.04841](https://arxiv.org/abs/2001.04841)  
**Date:** 2020-01-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2001.04841v1/

## Key points

- The rise of online education systems has made knowledge tracing a crucial aspect of personalized education
- Existing methods for knowledge tracing are domain-specific, which creates a problem when there is a lack of data in some domains
- This problem is referred to as domain adaptation for knowledge tracing (DAKT)
- To address this issue, the authors propose an adaptable framework called adaptable knowledge tracing (AKT)
- AKT incorporates educational characteristics such as slip, guess and question texts based on deep knowledge tracing (DKT) to obtain a well-performing model
- Three domain adaptation processes are proposed and adopted: pre-training an auto-encoder to select useful source instances for target model training; minimizing the domain-specific knowledge state distribution discrepancy under maximum mean discrepancy (MMD) measurement to achieve domain adaptation; and adopting fine-tuning to deal with output dimension differences between source and target domains.
- Extensive experimental results demonstrate AKT's effectiveness in achieving great knowledge tracing performance and superior transferable ability across various domains with different levels of data availability.

## Summary

The rise of online education systems has made knowledge tracing, the task of predicting students' knowledge state, a crucial aspect of personalized education. However, existing methods are domain-specific and with a larger number of domains in the real world and a lack of data in some domains, it is increasingly important to utilize information from other domains to train knowledge tracing models for target domains. This problem is referred to as domain adaptation for knowledge tracing (DAKT), which involves achieving great knowledge tracing performance in each domain and transferring well-performed models between domains. To address this issue, the authors propose an adaptable framework called adaptable knowledge tracing (AKT). For the first aspect of DAKT, AKT incorporates educational characteristics such as slip, guess and question texts based on deep knowledge tracing (DKT) to obtain a well-performing model. For the second aspect, three domain adaptation processes are proposed and adopted: pre-training an auto-encoder to select useful source instances for target model training; minimizing the domain-specific knowledge state distribution discrepancy under maximum mean discrepancy (MMD) measurement to achieve domain adaptation; and adopting fine-tuning to deal with output dimension differences between source and target domains. Extensive experimental results on two private datasets and seven public datasets demonstrate AKT's effectiveness in achieving great knowledge tracing performance and superior transferable ability. Overall, this paper highlights the importance of domain adaptation for effective personalized education through knowledge tracing. The proposed AKT framework provides a promising solution that can be applied across various domains with different levels of data availability.