Implementation of Long Short-Term Memory (LSTM) Models for Engagement Estimation in Online Learning

Considering that engagement is a dynamic inner state, understanding learners' engagement is an essential component for educators to provide personalized pedagogy in any learning setting. In online learning settings, reliable engagement estimation of learners is crucial for monitoring the quality of learning. In this paper, we discuss sequence-based engagement estimations using Long Short-Term Memory (LSTM) models with The Dataset for the Affective States in E-Environment (DAiSEE). The engagement states were classified as Not Engaged, Normally Engaged, and Very Engaged. We aimed to identify the most effective combination of pre-processing methods for an imbalanced distribution dataset by applying pre-processed methods: undersampling, oversampling, normalization, dimensional reduction, and the combination scenarios where the order of the pre-processed methods. The processed data were then used as the input to test four different LSTM models: single-LSTM, stacked-LSTM, Bidirectional LSTM (Bi-LSTM), and Multilayer Bi-LSTM. Overall, the Multilayer Bi-LSTM in the pre-processing scenario 4 yield the best performance, with a validation accuracy of 0.902.