Predicting Assessment Item Difficulty Levels Using a Gaussian Mixture Model

Abstract

The difficulty level of an assessment item plays an important role in ensuring well qualified evaluation process as well as helping in the generation of appropriate assessments for personalized learning. AMrita Personalized Learning and Evaluation (AMPLE) platform adopts an automatic calibration methodology using Gaussian Mixture Models for difficulty level assignment. This methodology uses performance features derived from the test-takers responses recorded in the assessment engine. Verification of this model, carried out on a diverse data set of assessment items spread over six subjects and 6000 students achieved about 91% accuracy by comparing the model-generated output with teacher-supplied difficulty levels.