Continuous Emotion Recognition: Another Look at the Regression Problem

Various regression models are used to predict the continuous emotional contents of social signals. The common trend to train those models is by minimizing a sense of prediction error or maximizing the likelihood of the training data. According to those optimization criteria, among two models, the one which results in a lower prediction error, or higher likelihood, should be favored. However, that might not be the case, since to compare the prediction quality of different models, the correlation coefficient of their prediction with the actual values is prevalently used. Hence, given the fact that a lower prediction error does not imply a higher correlation coefficient, we might need to reconsider the optimization criteria that we undertake in order to learn the regression coefficients, in order to synchronize it with the hypothesis testing criteria. Motivated by this reasoning, in this work we suggest to maximize a sense of correlation for learning regression coefficients. Two senses of correlation, namely Pearson's correlation coefficient and Hilbert-Schmidt independence criterion, are seen for this purpose. We have chosen the continuous audio/visual emotion challenge 2012 as the framework of our experiments. The numerical results of this study show that compared to support vector regression, the suggested learning algorithms offer higher correlation coefficient and lower prediction error.