Not Another Hardcoded Solution to the Student Dropout Prediction Problem: A Novel Approach Using Genetic Algorithms for Feature Selection.

Preventing student dropout is a challenge for higher education institutions (HEIs) that have worsened with COVID-19 and online classes. Despite several research attempts to understand and reduce dropout rates in HEIs, the solutions found in the literature are often hardcoded, making reuse difficult and therefore slowing progress in the area. In an effort to advance the area, this paper introduces a novel portable approach based on genetic algorithms to automatically select the optimal subset of features for dropout prediction in HEIs. Our approach is validated on a dataset containing approx. 248k student records from a Brazilian university. The results show that the proposed approach significantly increases the accuracy in dropout prediction, outperforming previous work in the literature. Our contributions in this paper are fourfold: the implementation of a (i) novel efficient and accurate automatic feature selector that does not require expert knowledge; (ii) an adaptive deep learning model for dropout prediction in sequential data sets; (iii) a portable solution that can be applied to other data sets/degrees; and, (iv) an analysis and discussion of the performance of feature selection and predictive models for dropout prediction.