When Probabilities Are Not Enough - A Framework for Causal Explanations of Student Success Models

Student success and drop-out predictions have gained increased attention in recent years, connected to
the hope that by identifying struggling students, it is possible to intervene and provide early help and
design programs based on patterns discovered by the models. Though by now many models exist achieving
remarkable accuracy-values, models outputting simple probabilities are not enough to achieve these
ambitious goals. In this paper, we argue that they can be a first exploratory step of a pipeline aiming to
be capable of reaching the mentioned goals. By using Explainable Artificial Intelligence (XAI) methods,
such as SHAP and LIME, we can understand what features matter for the model and make the assumption
that features important for successful models are also important in real life. By then additionally
connecting this with an analysis of counterfactuals and a theory-driven causal analysis, we can begin to
reasonably understand not just if a student will struggle but why and provide fitting help. We evaluate
the pipeline on an artificial dataset to show that it can, indeed, recover complex causal mechanisms and
on a real-life dataset showing the method’s applicability. We further argue that collaborations with social
scientists are mutually beneficial in this area but also discuss the potential negative effects of personal
intervention systems and call for careful designs.