A Novel Partitioned Random Forest Method-Based Facial Emotion Recognition

Facial emotion recognition (FER) has been applied to various sectors, including e-learning, marketing, humanoid robot design, HMI/HCI applications, and medicine. The rapid development of intelligent technologies has led researchers to strive to improve facial emotion recognition techniques. A range of machine learning (ML) methods can be used to recognize facial expressions based on data from small to large datasets. Random Forest (RF) is simpler and more efficient than other ML algorithms. Some modified versions of RF have been developed to improve classification accuracy in the literature. Most improved RF versions modify attribute selection processes or combine them with other machine learning algorithms, increasing their complexity. Identifying an appropriate training dataset and determining its size remain open questions. The partitioned random forests (PRFs) approach is proposed as a modified strategy for improving FER. The proposed method divides multiple regions (different data lengths) into the feature space, allowing the algorithm to learn more complex decision boundaries. Using three statistical measures Lyapunov exponents (LE), Correlation Dimension (CD), and approximate entropy (AE), we evaluated the performance of machine learning algorithms over different data lengths. A crucial role for classification accuracy is played by the Lyapunov exponent or LE and the size of the dataset. A PRF is more effective on smaller datasets and with higher LE values. The proposed method for partitioning the datasets has been successfully tested on the FER dataset to classify six basic emotions (sadness, anger, fear, surprise, disgust, and happiness). Based on our numerical results, PRF performed better than traditional RF and other ML methods for FER, providing 98.37% mean absolute accuracy. Thus, a robust and useful method for improving classification rates is proposed for both small and large datasets.