- focused on considering forgetting behavior in knowledge tracing
- extended the DKT model so that the information related to forgetting can be incorporated into the model
- achieved better predictive performance compared with the DKT model on two real-world datasets for knowledge tracing
- revealed that the combination of several types of information which affect the behavior of forgetting resulted in improved performance

University of Michigan - Ann Arbor

- Introduction
- Related Work
- Preliminaries
- Proposed Approach
- Experiments
- Conclusion

Augmenting Knowledge Tracing by Considering Forgetting Behavior

- CAEs that use KT are very widespread, even more so BKT
- However, many students have forgetting which is typically not accounted for in most BKT models
- In this study, they focus on modeling and predicting a student’s
knowledge considering the forgetting behavior from large amounts
of learning logs to support providing adaptive learning materials in
CAE systems.

Augmenting Knowledge Tracing by Considering Forgetting Behavior (Introduction)

Augmenting Knowledge Tracing by Considering Forgetting Behavior (Related Work)

Augmenting Knowledge Tracing by Considering Forgetting Behavior (Preliminaries)

- Utilize the DKT model as our base model because it is a deep
neural network that can easily incorporate multiple input sources
and capture the nonlinear dynamics among them
- Extends DKT so that it can consider forgetting, the model can adapt the
student performance to the student’s forgetting
- Consider the following three features: 1) repeated time gap: the lag time between an interaction and the previous interaction with the same skill id, 2) sequence time gap: the lag time between an interaction and the previous interaction in the sequence; the skill id of an interaction does not matter, and 3) past trial counts: the number of times a student answers questions with the same skill id.

Augmenting Knowledge Tracing by Considering Forgetting Behavior (Proposed Approach)

- two datasets: ASSISTments and slepemapy.cz
- five-fold cross validation in which each dataset was split based on students
- the combination of concatenation and multiplication (hybrid) achieved the best AUC on both the ASSISTments and slepemapy.cz datasets, with values of of 0.7309 and 0.8046, respectively
- more features used for modeling, the better the predictive performance

Learn Before

Related