Deep Knowledge Tracing (DKT), along with other machine learning approaches, are biased toward data they have seen during the training step. Thus for problems where we have few amount of data for a certain class, the models will tend to give good results on classes where there are many examples, and poor results on those with few examples. This problem generally occurs when the classes to predict are imbalanced and this is frequent in educational data where for example, there are skills that are very difficult or very easy to master. There will be less data on students that correctly answered questions related to difficult knowledge and that incorrectly answered knowledge easy to master. In that case, DKT is unable to correctly predict the current student’s knowledge on those skills. In order to improve DKT in that sense, we penalized the model using a cost-sensitive technique. In other words, we have augmented the loss function with the same loss where we have masked certain skills. We also included in the DKT, a Bayesian Network (built from domain experts) by using the attention mechanism. The resulting model is able to accurately track knowledge of students in Logic-Muse Intelligent Tutoring System (ITS), compare to the BKT or the original DKT.
There are 2 main files : DKT.pyand DKT_BN.py.
DKT.py: takes the data and run the original DKT + cost sensitive on selected skills (with limited amount of data). The combination of DKT and BN is also possoble using attention mechanism but we put that part in comments.DKT_BN.pyisDKT.pybut trained multiple times.