Feature Extraction of Students and Problems via Exam Result Analysis using Variational Autoencoder

Abstract

In this paper, we propose a novel examination-result analysis method based on latent variables gained from Variational AutoEncoder (VAE) specially designed for this purpose. We train our VAE so that the range of latent variables are within 0 and 1 and also monotonical concerning output of VAE’s decoder, while minimizing reconstruction loss between input and output like existing VAEs. Using the latent variables, we report a detailed analysis of both the problems and the examinee.