Informing predictive models against Students Dropout

Students’ dropout is a complex widespread phenomenon which often lead to conditions of social, educational and professional exclusion. The design of Early Predictive Analytic Models can be a valid tool to counteract this phenomenon, which can be further enhanced by using Machine Learning. In this position paper we aim to contribute with two main points. First of all, we introduce the prominent position of the skills assessment, considered both as a target or as input data for the model, as essential integration to demographic and of economic, social and cultural status variables, often used as predictors for dropout risk. This leads us also to give a definition of implicit dropout, i.e. failure to achieve the expected skills, applicable in different educational contexts. Furthermore, we highlight the importance of integrate the predictive models in a broader framework described as a sequence of phases. The framework stresses the need to make the model “informed” at three levels: a reference pedagogical theory (a theory-laden dimension in a data-intensive approach); the persistence of the initial information and their integration together with the life cycle of the model (its creation, use and update); the guidelines to enable the explainability and transparency of the model outcomes, in accordance with the principles of Trustworthy AI. These contributions are presented both through an abstract description and an undergoing case study in Italian school system.