Academic achievement critical factors and the bias and variance decomposition

Abstract

This study is centered on the sources of machine learning bias in the prediction of students’ grades. The dataset comprises 29,788 Portuguese high school teacher final grades corresponding to 10,364 public high school students’ academic paths (from the 10th to the 11th grades). We use an artificial neural network to perform the tasks. In the experimental phase, we undertake a bias and variance decomposition when predicting the 11th year students’ grades. Two different implementations are used, a critical implementation that comprises only academic achievement critical factors and a lagged implementation where the preceding teacher grade is appended. The critical implementation has a higher machine learning bias, notwithstanding the higher critical factors’ contribution. The lagged implementation, on the other hand, has a smaller bias, but a smaller critical factors’ contribution. It is possible for a machine learning model to have a reduced bias and simultaneously a little critical factors’ contribution, simply by accessing information about the historical value of the target variable. The education stakeholders should therefore be aware of the critical quality of the model in use. In defining policies and choosing the variables to influence, predictive models with low biases and built upon the critical factors information are indispensable. A machine learning model based on the critical factors produces more consistent estimates of their effects on AA. They are therefore suitable models to assist in policymaking. On the other hand, if the goal is to obtain a simple set of predictions, the use of target variable historical values is appropriate.