Distributed Machine Learning Frameworks for Emotion Classification Leveraging Facial Action Coding System Features

Facial Expression Recognition (FER) plays a pivotal role in emotion processing, as facial expressions are primary source for perceiving human emotions. Over the past two decades, research in affective computing has significantly advanced, particularly in areas such as FER, speech emotion recognition, emotion recognition from physiological signals, Human-Computer Interaction (HCI), and robotics. This study focuses on FER, emphasizing the importance of interpretability and transparency in its applications across healthcare, education, industry, and transportation. We propose distributed frameworks for classifying seven discrete emotions—anger, disgust, fear, happiness, sadness, neutral, and surprise—using Action Units (AUs) from the Facial Action Coding System (FACS). These AUs effectively describe facial muscle movements and capture subtle variations in facial expressions. Our approach leverages benchmark datasets, including AffectNet, FER2013 and CK+48, which contains millions of labeled facial images, and employs distributed machine learning to enable scalable data processing. Experimental results indicates that logistic regression outperforms other multiclass classification algorithms, achieving a classification accuracy of 99.88%. SHapley Additive exPlanations (SHAP) analysis was used, which helped to highlight the contribution of specific AUs to emotion predictions, enhancing model interpretability. The distributed framework further improves classification performance by training multiple models on distinct subsets of data or focusing on specific facial regions, resolving ambiguities in similar emotion classes, and reducing training time. This work underscores the significance of distributed and explainable approaches in FER, ensuring scalable, accurate, and transparent emotion recognition systems for diverse applications.