Sequential Knowledge Tracing with Transformer Models

Sammanfattning

Transformer models, delivering big improvement in AI text-models (NLP), are now being applied in Knowledge Tracing to track the knowledge of students over time. One of the first, SAINT, showed quite some improvement over the then SOTA results on the public EdNet dataset and caused an increase in research based on transformer-based models. In this paper, we firstly aim to reproduce the SAINT results on the EdNet dataset but are unable to report a similar performance as the original paper. This might be due to implementation details, which we were not able to completely reconstruct. We hope to pave the road for further reproducibility, as an increasingly important part of AI research. Furthermore, we apply the model to a company dataset much larger than any public dataset (more interactions, more exercises and more skills). Such a dataset is on the one hand more challenging (more skills mixed), and on the other hand, provides much more data (which should help our models). We compare the SAINT model and the seminal IRT model, and find that the SAINT model performance is 4% better in AUC but 1.7% worse in RMSE. Our experiments on window size suggest that transformer models still struggle with modelling beyond recent performance, and do not yet deliver the step-change observed in NLP.