Domain-Independent Proximity Measures in Intelligent Tutoring Systems

Abstract

Intelligent tutoring systems (ITSs) typically analyze student solutions to provide feedback to students for a given learning task. Machine learning (ML) tools can help to reduce the necessary effort of tailoring ITSs to a specific task or domain.\nFor example, training a classification model can facilitate feedback provision by revealing discriminative characteristics in the solutions. In many ML methods, the notion of proximity in the investigated data plays an important role, e.g. to evaluate classification boundaries. For this purpose, solutions need to be represented in an appropriate form, so their (dis-)similarity can be calculated. We discuss options for domain- and task-independent proximity measures in the context of ITSs, which are based on the ample premise that solutions can be represented as formal graphs. We propose to identify and match meaningful contextual components in the solutions, and present first evaluation results for artificial as well as real student solutions.