Discussing Solutions to the Data Imbalance Problem in Emotion Recognition

Emotion recognition technology has been widely used in human-computer interaction, medical health and other fields. However, in practical applications, emotion datasets often have class imbalance problems, which lead to the model being seriously biased towards the majority class, significantly reducing the recognition accuracy and reliability of minority emotion classes. This paper focuses on comparing and analyzing methods such as ESC-GAN generative data augmentation technology, DER-GCN dialogue and event relationship perception graph model, and MultiEMO multimodal fusion framework to solve the problem of imbalanced emotion recognition categories, and explores the innovations and limitations in multiple scenarios. These methods compensate for minority emotions from different angles: for example, MultiEMO significantly improves the ability to classify minority emotions through cross-modal attention mechanism and weighted contrast loss, which can not only be applied to detect the psychological emotions of patients in the medical health field, but also help to provide support for fine-grained emotion classification in security scenarios. Experimental results show that these solutions significantly improve the accuracy and F1 value of emotion recognition, especially in extremely unbalanced categories. This paper provides a systematic reference for the selection of technology for high-value scenarios such as medical monitoring and intelligent security, promotes the interdisciplinary collaborative development in the field of emotional computing, and accelerates the application transformation of this technology in practice.