Recommendation systems for learning materials are a common approach to
personalizing learning environments. In this Bachelor's Thesis we focus on
recommending learning exercises using student's performance data. Many
approaches using skill modeling to profile students and exercises have been
developed in this context. Additionally, factorization techniques have been
shown to perform well in predicting students' future performance. We describe
some of these models and focus specifically on tensor and weighted
factorization for dynamic performance prediction. We then evaluate some
factorization approaches on data gathered from an online learning platform.
Finally, we specify criteria for successful recommending in educational systems
and assess the appropriateness of factorization for this purpose. Nonnegative
matrix and tensor factorization did not significantly improve on
predicting with item averages. However, temporally weighted matrix factorization
was found to be a viable alternative to tensor factorization. For
recommending purposes, factorization seems appropriate only under some
circumstances and with enough consideration for its weaknesses. From this
we conclude that the choice of a recommendation model and strategy must
take into account the goal and context of the learning environment.
|
