Evaluating Student Performance in E-learning Systems: A Two-step Robust Bayesian Multiclass Procedure

This paper addresses a computational method to evaluate student performance through convolutional neural network. Image recognition and processing are fundamentals and current trends in deep learning systems, mainly with the outbreak in coronavirus infection. A two-step system is developed combining a first-step robust Bayesian model averaging for selecting potential candidate predictors in multiple model classes with a frequentist second-step procedure for estimating the parameters of a multinomial logistic regression. Methodologically, parametric conjugate informative priors are used to deal with model uncertainty and overfitting, and Markov Chains algorithms are designed to construct exact posterior distributions. An empirical example to the use of e-learning systems on student performance analysis describes the model's functioning and estimation performance. Potential prevention policies and strategies to address key technology factors affecting e-learning tools are also discussed.