Improving Emotion Recognition Performance by Random-Forest-Based Feature Selection

Towards affect‐aware vehicles for increasing safety and comfort: recognising driver emotions from audio recordings in a realistic driving study

“…When regarding the original data set, the SVM‐classifier is unable to reach the performance of its baseline classifier. In [69] it was stated that a performance increase of the SVM‐classifier could be achieved by conducting a feature selection based on an RF‐approach. Their results show that depending on the chosen data set, a feature selection can lead to a UAR above baseline, when using over 40% of the original features.…”

“…The largest feature set evaluated in our study comprises 98 features (10% of the original features). From the results presented in [69] it can be assumed that the number of features needed to achieve a performance above baseline is not yet reached. Furthermore, the presented feature selection is a so‐called ‘wrapper’‐method, which identifies optimal features based on the performance of the chosen learning algorithm (in this case RF).…”

“…By only selecting those features with high feature importance, irrelevant features were excluded from the feature set. This reduces the computing power for a later real‐time implementation dramatically, as only the relevant features will need to be extracted from the speech signal, and can also have a positive impact on the performance of the classifier [69]. As presented in [69], we assume that the selected features using permutation of OOB‐observations from RF are also the most relevant features when applied to SVM‐classifiers.…”