作者：Kejiang Xiao*（肖克江）, Wenqi Yang, Guangjie Zhu, Jiefan Qiu, Shiyan Pang & Chongming Zhao

出版刊物：Journal of Super computing

出版时间：2026年

内容摘要：

Multimodal emotion recognition (MER) relies on effective fusion of heterogeneous data but is challenged by uncertain data and high-conflict evidence. Dempster–Shafer (D–S) evidence theory is advantageous for fusing uncertain information; yet, its performance is limited by two key issues: unreliable basic probability assignment (BPA) generation and counterintuitive results under high-conflict scenarios. In this paper, we propose a decision fusion framework that addresses these challenges by designing a BPA allocation method and improving the evidence theory fusion algorithm. The proposed method utilizes Possibilistic C-Means (PCM) clustering technology to construct noise-resistant interval models, enabling robust BPA generation even with small sample sizes. It also calculates evidence weights by considering both direct and indirect relationships between evidence, thereby improving the fusion accuracy of highly conflicting data. Furthermore, its modular architecture facilitates parallel execution across modalities and evidence sources, ensuring low computational cost and scalability for distributed and high-performance computing environments. We validated our method using the CMU-MOSI and CMU-MOSEI dataset and demonstrated its superior performance compared to existing methods. We also designed BPA experiments with noisy datasets and high-conflict data fusion experiments to verify the effectiveness of the proposed method.

多模态情感识别（MER）依赖于异构数据的有效融合，但面临数据不确定性与高冲突证据的挑战。Dempster–Shafer（D–S）证据理论在融合不确定信息方面具有优势；然而，其性能受限于两个关键问题：不可靠的基本概率分配（BPA）生成，以及高冲突场景下的反直觉结果。本文提出一种决策融合框架，通过设计BPA分配方法并改进证据理论融合算法来应对上述挑战。所提方法利用可能性C均值（PCM）聚类技术构建抗噪区间模型，即使在小样本条件下也能实现稳健的BPA生成。该方法还通过同时考虑证据之间的直接与间接关系来计算证据权重，从而提升高冲突数据的融合精度。此外，其模块化架构便于跨模态与跨证据源的并行执行，确保低计算成本，并具备面向分布式与高性能计算环境的可扩展性。我们使用CMU-MOSI和CMU-MOSEI数据集对所提方法进行了验证，并证明其性能优于现有方法。我们还设计了含噪声数据集的BPA实验以及高冲突数据融合实验，以验证所提方法的有效性。