Title: Interpretable machine learning techniques for the analysis of process data from computer-based assessments
Abstract: This work is inspired by the "Machine Learning in Education" workshop for our master students that took place last Fall. Here, we investigate the relationship between process data indicators extracted from log files and students’ performance by aligning the explorative nature of data mining techniques and state-of-research on how students interact with one problem-solving item from PISA 2012. Using data from Scandinavian students, the analyses were conducted in three steps: (1) extraction of process indicators from log files, (2) modeling of students' performance via tree-based classifiers, and (3) evaluation of the importance of extracted indicators into the modeling. These steps stand on the foundation of interpretable machine learning to generate new knowledge by offering important insights into the abundance of data that can be derived from log files and modeled through predictive analytics. We also present an R package that has been developed which aims to add a flexible tool to the data analyst’s arsenal.