Protocol and Results Clause Samples

Protocol and Results. As in the challenge tasks, we evaluate performance in terms of unweighted average recall (UAR). In addition, we use the original challenge feature sets for the tasks of emotion and likability recognition in our experiments. Thus, for emotion recognition, we use 384 features resulting from a systematic combination of 16 low-level-descriptors (LLDs) and corresponding first order delta coefficients with 12 functionals (▇▇▇▇▇▇▇▇ et al., 2009); for likability recogni- tion, we utilize 6 125 features by brute-forcing based on 64 LLDs and 61 functionals (▇▇▇▇▇▇▇▇ et al., 2012a) – all features are extracted with the open-source toolkit openS- MILE (Eyben et al., 2010). In the same vein, we keep the classifiers, their implementations, and parameters as in Challenges: for emotion recognition, Support Vector Ma- chines (SVMs) trained by Sequential Minimal Optimiza- tion (SMO) with polynomial kernel (degree 1) and a com- plexity constant of 0.05; for likability recognition, Random Forests (RF) with a number of trees N = 1 000 and a fea- ture subspace size of P = .02. The Weka toolkit (Hall et al., 2009) is used in both cases. Note, that the instance se- lection algorithm is only applied on the training set. The test set is not modified and kept the same as in the original Challenge setup in order to allow for a direct comparison.