Autoencoders For Educational Assessment

In educational assessment research, a common goal is to determine students' knowledge about some construct. This knowledge is latent and can be represented by continuous variables which influence the individual's performance on a test. Item response theory (IRT) models structure this relation, defining specific functions between the knowledge of the individual, and the probability of answering an item correctly. Previous research implies that neural networks can emulate these models, and, with a modification in its architecture, overcome some of the limitations concerned to "big data" analysis. In this work, we compare two different types of neural networks for this application: autoencoders (AE) and variational autoencoders (VAE). Not only can these neural networks be used as similar predictive models, but they can recover and interpret parameters in the same way as in the IRT approaches.