Abstract:

Student success has been a concerning issue in higher education institutions. The lack of success paves the way for several personal and social problems. It has also financial and reputational implications for institutions. Thus, techniques that foster student success in higher education attracted much attention.

Methods based on learning analytics are often used as a tool to improve student success outcomes. They are mainly based on student and/or learning data and predictive modelling techniques. They are applied to inform proactive interventions. For instance, identifying at-risk students early in the course, and target them with an intervention. To do so, predictive models are trained using legacy data of previous students and applied to present students’ own data.

These so-called ‘first-order’ techniques do not account for the effect that an intervention has on a cohort of students. Consequently, further research was needed to conduct ‘second-order’ techniques. These models aim to optimize targeted interventions in terms of selected at-risk students as well as the design of the interventions. Historical knowledge of the effects of and on what interventions were performed are thus included in the model itself.

Though an improvement, these second-order models do not allow for the optimal impact of interventions to improve student success outcomes. For instance, it has been reported that at-risk students dropout more quickly if they had received an intervention. This is because of that, they fail to distinguish between at-risk students who respond favorably because of intervention and students who respond favorably on their own accord when not targeted.

To address this limitation, we suggest moving from predictive to ‘uplift modelling’ of intervention. Uplift modelling is a novel approach in machine learning, which aims to predict the net effect of performing an action on a certain outcome. It is used in direct marketing and personalized medicine for identifying individuals who are most likely to be positively influenced by an intervention.

The aim of this position paper is to provide an overview of uplift modelling, with particular emphasis on the optimal targeting of at-risk students.

We will discuss (i) predictive modelling & its limitation
(ii) concepts & applications of uplift modelling,
(iii) and the use of and challenges of uplift modelling within the specific context.

To explore the applicability, an experimental design was set up by identifying a control and treatment group. Thus, we can model the difference in behavior between the intervention and control groups. This allows us to predict the change in each at-risk student’s success probability when intervened. The introduction of a control group adds another layer of complexity into uplift modelling, unlike predictive modelling that only involves a treatment group.

In general, uplift modelling aims at identifying and targeting the persuadable. Although it has application to various settings, most of the current developments are in direct marketing and personalized medicine. In our context, applying uplift modelling maximizes the likelihood of identifying at-risk students who are most favorably influenced by an intervention. Therefore, employing intervention policy based on uplift modelling allows educators to ensure efficient use of limited resources while increasing student success outcomes.

Keywords:

Student success, Educational Intervention, Learning Analytics, Predictive Modelling, Uplift Modelling.