Recently, I am doing knowledge tracking. I read two Chinese reviews and summarized the context of knowledge tracking. After that, I will mainly read papers on cognitive tracking based on deep learning. Everyone is welcome to discuss. I will continue to update the papers and thoughts on in-depth knowledge tracking later...
1. 2020 Research Progress on Cognitive Tracking Models in Educational Big Data_Hu Xuegang
Q: How to mine and analyze massive educational data?
A: Cognitive tracking progress. First, two methods of cognitive tracking progress are introduced, combined with application scenario analysis, and finally the advantages and disadvantages of the two models are analyzed and future research directions are discussed.
Based on the data of students answering exercises during the learning process, track the change of students' mastery of knowledge points over time, and predict students' performance in answering exercises in the future.
- Cognitive Tracking Model Based on Bayesian Method-BKT
The relationship between observed variables and knowledge points is known, and the knowledge points associated with exercises are grouped to form a sequence
Hypothesis: Each exercise is only related to one knowledge, forgetting factors are ignored, and the transition probability remains unchanged.
Algorithm: From the answering situation at time t --> the mastery situation at t-1 (conditions) --> the mastery situation at t.
Expansion: Educational data characteristics (similar test questions,)
Features: It has strong interpretability, but loses the characteristics of student data, and the predictive ability is average
First-order HMM
- Cognitive Tracking Model-DKT Based on Deep Learning Method
Build a model for all knowledge points, the relationship between observations and knowledge points is unknown, and a student answers all knowledge points as a sequence
RNNinput, hidden layer, output
One-hot notation, why Q+q,
Extension: Educational data characteristics (forgetting curve, student clustering)
Features: strong expressive ability and predictive performance, unable to explicitly obtain the cognitive state of students for each specific knowledge point, general interpretability
-
Analysis and Outlook
Bayesian parameter optimization, the trade-off between accuracy and interpretability, and the inconsistency with actual problems
2. 2021 Research Progress on Knowledge Tracking Based on Deep Learning_Liu Tieyuan
Knowledge tracking: By establishing a model of students' knowledge status changing over time, it can judge the degree of students' mastery of knowledge and dig out potential learning rules from the learning trajectory, so as to provide personalized guidance and realize artificial intelligence-assisted education.
DLKT: The DKT model is based on RNN
Three major technical improvement issues:
- Improvements to interpretability issues
- Ante-hoc: Insufficient resolution, both the attention mechanism and the self-explanatory model know to improve the interpretability of a certain part
- Post-hoc: high requirements for some conditions (LRP requires a high gradient of the prediction function)
- long-term dependence
- LSTM-based extended model: Extending the length of RNN sequence learning has not been fundamentally resolved
- Model based on self-attention mechanism: Transformer model does not rely on RNN, which solves the long-term dependency problem, but loses the ability of RNN to model sequences
- Missing Learning Feature Problem
- Embedding method: Add learning features to the input embedding vector of the model, or embed it into the calculation process as an additional calculation factor
- Loss function restriction: the way to encode additional learned features into the loss function as a restriction
- New structure: Using a new model structure, the way to incorporate additional learned features into the model calculation process
Application of DLKT model:
- Topological sorting for discovery topics
- for subjective questions
- Validation of educational theory
- Recommended learning path
- career forecast
- generate test paper