Modelling Learners’ Behaviour: A Novel Approach Using GARCH with Multimodal Data

Abstract

Most of the contemporary approaches in learner behaviour modelling either quantize continuous data into discrete states/events (e.g., HMM), or assume that the patterns in the data are distributed homogeneously in time (e.g., auto-regression). This paper proposes a novel approach that overcomes the above mentioned issues and models learner behaviour using Generalized Auto-Regressive Conditional Heteroskedasticity (GARCH). GARCH uses continuous time-series data, without the need for information quantization, and it considers the heterogeneity of event distribution in the time-series. A study was conducted to demonstrate how GARCH can be configured in an adaptive assessment setting. Specifically, GARCH was applied on six different constructs from eye-tracking and electroencephalogram (EEG) data, and compared with existing methods of modelling time-series data, such as Markov Models and models having auto-regressive components. The comparison shows that the models having a GARCH component outperform other models for most of the students, for all the variables considered. The results are encouraging towards building accurate learner behaviour models, adequate to drive the design and development of adaptive feedback tools (e.g., an early alert system), but further investigation is required.