Comparison of Methods of Affect Transition Analysis: An Example of SimInClass Dataset

Abstract

Studies of emotional-cognitive sequences are the growing body of research area in educational context. These studies focus on how emotions change during the learning-teaching process due to their dynamic nature. In affect transition studies, the change of emotion, depending on the event and time, is usually analyzed by using (a) lag sequential analysis (LSA), (b) L metric, (c) L* metric, and (d) Yule's Q metric. Yet, various methodological criticisms exist in the literature while utilizing these sequential analysis methods. In this study, it is aimed to comparatively examine lag analysis, L metric, L* metric, and Yule’s Q in terms of proportion of invalid values, maximum transition metrics, minimum transition metrics, and analysis results. For this reason, the emotional states of the fifteen prospective teachers were collected and their emotions were labeled every 0.5 seconds with EEG (Electroencephalogram), GSR (Galvanic Skin Response), and Microsoft Kinect in a teacher training simulator (SimInClass). The dataset contained 17570 emotions, and the data were analyzed by utilizing lag analysis, L, L* and Yule's Q. The results showed that LSA had yielded the most proportion of invalid results. In addition, it was observed that the number of invalid values increased as the segment length became shorter in all analysis methods. When the maximum and minimum transition metric values were examined, it was found that as the sequence length increased in L and L* analyses, the value of the metrics approached 1, which is the largest value they can reach. However, it was noted that the lag analysis maximum-minimum transition metrics fluctuate independently from the sequence length. It was concluded that there were differences in the analysis results produced by the four sequential analysis methods with the same functions. It was thought that this situation might be due to the different invalid results produced by the analyses. When the results were compared with the studies in the literature, it was thought that it would be beneficial to pay attention to the nature of the data (emotional or behavioral), the data type (singe modality or multimodal modality), the amount of data (short sequences or long sequences), the environment in which the dataset was created (computer-based or not), and the sampling rate (automated data collection tool or observation) when choosing sequential analysis methods.