Feature Selection Based on Causal Analysis Algorithm for Speech Emotion Recognition

Feature selection is an important step of pre-processing in machine learning and data analysis. However, existing feature selection algorithms rarely reveal the underlying causal mechanisms between features, ignores the interaction between them and lacking interpretability. In addition, how to use causal analysis methods to mine causal relationships between speech emotions and features, select the subset of speech features that are optimal for speech sentiment with the class variable. To address these issues, this article presents a causal feature selection algorithm called Interleaved Forward-Backward Markov Blanket (IFBMB) algorithm for speech emotion feature selection, aimed at finding causal relation-ships between speech emotion and features. In this paper, we extracted two feature sets named FP, MFCC on IEMOCAP and EMODB databases. Experimental results showed that the proposed IFBMB algorithm outperforms the traditional causal feature selection algorithm. IFBMB can effectively select the main features for speech emotion to reduce the dimensionality and improve the speech emotion recognition rate. The speech emotion recognition rate of the IFBMB algorithm on the MFCC and MFCC+FP of EMODB Database reaches 91.59% and 92.83%.