Application of Multimodal Emotion Recognition Technology in Recommendation Systems

Abstract

Traditional recommender systems, relying mainly on users' historical behavioral data to predict preferences, often overlook the importance of real-time emotions in decision-making. As a result, they fail to meet the emotional needs of users. This study focuses on integrating multimodal emotion recognition technology with recommender systems to address this issue. It clarifies the necessity of such integration by analyzing the shift from correlation mining to causal understanding in recommendation systems and the value of multimodal emotion recognition. The research systematically analyzes five cutting-edge technologies: empathetic recommendation via large language models, robustness enhancement through causal inference, generative recommendation using diffusion models, emotion alignment via cross-modal contrastive learning, and privacy-preserving recommendation based on federated learning. Then, the research explores core challenges like efficiency, interpretability, and short-term homogenization, along with solutions such as knowledge distillation and neuro-symbolic methods. These technologies effectively tackle issues like emotional understanding, data bias, sparsity, modal alignment, and privacy protection. The study concludes that these technologies are converging, with future directions focusing on causal explanation, efficiency improvement, and long-term user well-being, driving recommender systems toward greater intelligence, robustness, reliability, and human-centricity.