A Comparison of Evaluation Measures for Emotion Recognition in Dimensional Space

Emotion recognition from physiological signals like electroencephalography (EEG) can be performed using different underlying emotion models. While dimensional emotion models have recently gained attention, measures to evaluate recognition methods that are based on these models differ from study to study. This paper offers an analysis of proposed evaluation measures by comparing recognition results achieved on a self recorded dataset. Emotions are estimated using ridge regression and estimation results are compared using different evaluation measures. Additionally, three different baselines are studied, two types of random regression as well as naive estimation. Among the investigated evaluation measures, bandwidth accuracy was found to have many desirable characteristics.