Assessment of At-Risk Students' Predictions From E-Book Activities Representations In Practical Applications

Abstract

The use of e-book reading systems such as Bookroll, and their ability to record readers' activities allows the design of predictive models capable of identifying at-risk students from their reading characteristics. Even though previous works have obtained promising results in this task, these results may not evidence the expected prediction performance in practical applications due to their selected assessment methods. Accordingly, in this paper, we assess this performance in two practical scenarios. The first is when we keep stored data from previous years of our course which can be used to train our model, and the second is when we only have data from a different course to use in this training process. In order to obtain a more accurate assessment, we collected 92,574 samples of predictive performances from different models under the above-mentioned conditions. We also considered different feature representations along with variational latent representations, which can leverage our previous data to automatically design general hidden features. From our results, we understand that in the first condition we can expect a relatively good predictive performance, especially when using variational latent representations. However, in the second condition we found that even when using them, the predictive performances are very limited resulting in an impractical solution.