Explainable Prediction of Academic Failure Using Bayesian Networks

Currently, academic dropout is a crucial problem of higher education institutions in Chile due to the high social and economic costs that it entails. Usually, dropout prediction is performed using computational methods analyzing the student's personal information as well as their academic indicators during their studies, which require a period of time to know them. Moreover, this event is highly correlated to the academic failure of several key courses. An attractive alternative is to apply an initial test to achieve this prediction without requiring completion of the typical study periods. In this work we propose the use of Bayesian networks to make the explainable predictions of academic failure from an initial test considering the incorporation of knowledge from experts in educational management in the computational model. In particular, different configurations of Bayesian network models are applied to first-year engineering students at a Chilean university during the year 2019. The results indicate that this approach generally obtains 85% accuracy where the Bayesian networks show complex relationships between the variables. This work shows that the Bayesian network model can eventually detect the students most likely to drop out at the beginning of their studies. As future work, we plan to incorporate external variables as the student socio-economic data.