Question bank calibration using unsupervised learning of assessment performance metrics

Abstract

In traditional e-assessments as well as in personalized learning, question bank calibration plays an important role in ensuring high-quality assessment outcomes. We propose an unsupervised learning approach that uses performance metrics derived from test-taker responses for precise calibration of question banks. We show partitioning test-takers into three groups using their scores is an effective feature extraction method. This approach, enables us to accurately capture performance variations in test-taker sub-groups thereby effectively capturing inter-item correlations. Integration of such a Machine Learning (ML) approach into the Question Bank module of a cloud-based e-assessment system enables seamless automatic question calibration. We validate the effectiveness of our approach, using a large dataset, collected from a two thousand student university-level proctored assessment.