Edulab is looking for interested students for projects regarding student knowledge modelling. Edulab is the largest provider of math education in Denmark and get over 1 million student answers per day. In this project you would get access to working on this huge amount of data, trying to solve real world problems.
The overall problem of knowledge tracing is trying to predict how well a student can answer a given question, based on the student’s history of previously answered questions. This problem has received a lot of attention, using a large set of different mathematical models to solve it [1-3]. Especially deep neural network has been shown to perform very well on Edulabs data, but there is still a lot of interesting work to do. The project can take on many different forms based, and we will just propose 3 different research questions that could be of interest, but are also open for other ideas.
1. The current problem of knowledge tracing is defined as predicting how well the student will do on the next question given the students history. The temporal aspect is extremely important when prediction how well a student will do (meaning, if they do well in the very recent questions, there is a large possibility they will do well on the next question). It would be interesting to change the current problem of only focusing on how the student do on the next question, to also predict how the student will do later in the future. This would ideally require a better student knowledge model, as the temporal aspect can’t be “abused” as much when modelling.
2. Current deep learning approaches for knowledge tracing, can be considered as black boxes, and deriving qualitative information useable by teachers or similar is very hard. Developing models which are both expressive and more interpretable than current neural models is of great interest.
3. Recently Zhang et al.  used memory networks to great success for knowledge tracing. This was on a much smaller dataset, for students in a limited age interval. Reproducing and extending this work using Edulabdataset, would provide challenges with how to handle much more diverse students and different level of activities. In addition the number of different topics at Edulab is much greater, opening up for extending with hierarchical models.
If none of these ideas sounds interesting to you, but you have a good idea you would like to test out using a dataset that is much larger than what is normally available for research in this domain, you can contact us as well.
 Piech, Chris, et al. "Deep knowledge tracing." Advances in Neural Information Processing Systems. 2015.
 Khajah, Mohammad M., et al. "Integrating knowledge tracing and item response theory: A tale of two frameworks." CEUR Workshop Proceedings. Vol. 1181. University of Pittsburgh, 2014.
 Zhang, Jiani, et al. "Dynamic key-value memory networks for knowledge tracing." Proceedings of the 26th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2017.
I samarbejde medEduLab