学位论文简介

针对当前多模态情感计算在处理非平稳、多源异构交互时面临的拓扑时序固化、频域特征擦除、深层语义漂移以及底层梯度失衡等核心挑战，本文以图神经网络为基础，确立了“拓扑重构 → 频域解耦 → 时空演化 → 梯度寻优”的研究主线。主要研究内容归纳如下：

(1) 动态演化拓扑建模与自适应聚合。 为突破静态构图的拓扑固化局限，引入重叠滑动窗口机制构建局部演化子图，规避长程序列的过度平滑问题。同时，校准拉普拉斯算子的度数惩罚偏置，推导度感知相似度矩阵，提升关键情感枢纽节点的聚合权重，纠正特征同质化稀释现象。

(2) 双维上下文滤波与多频传播解耦。为过滤冗余交互节点以构建高信噪比表征，设计基于语义关联与信息熵的动态滤波机制。此外，提出高低频并行的图传播架构，同步保留跨模态情感共性，并精准刻画模态间的异质性情绪冲突，实现频域层面的深层解耦。

(3) 权重时空演化与多尺度拓扑协同。 为抑制瞬态噪声引发的语义漂移现象，构建时序图演化门控循环机制，通过增量式更新图聚合规则，平衡长程依赖捕获与显存开销之间的矛盾。结合度归一化算子与并行多尺度图卷积架构，在规避结构性拓扑盲区的同时，实现异构信号多尺度时空特征的高效聚合。

(4) 博弈幅度校准与多目标凸包投影。 为解决跨模态联合训练的底层优化困境，引入基于合作博弈论的夏普利值的估值机制，自适应校准梯度范数以缓解幅度失衡。同时，设计改进的多目标凸包投影算法，将参数更新严格约束于最优凸包内，确立了异构信号均衡演化的理论收敛下界。

主要学术成果

[1] Yingxue Gao, Huan Zhao*, Zixing Zhang*. Adaptive Speech Emotion Representation Learning Based On Dynamic Graph [C]. IEEE International Conference on Acoustics, Speech and Signal Processing, 2024: 1116-11120. (CCF B, 学生第一作者)

[2] Huan Zhao, Yingxue Gao*, Haijiao Chen, Bo Li, Guanghui Ye, Zixing Zhang*. Enhanced Multimodal Emotion Recognition in Conversations via Contextual Filtering and Multi-Frequency Graph Propagation [C]. IEEE International Conference on Acoustics, Speech and Signal Processing, 2025: 1-5. (CCF B, 导师第一作者，学生第二作者)

[3] Huan Zhao, Yi Ju, Yingxue Gao*. Bilevel Relational Graph Representation Learning-based Multimodal Emotion Recognition in Conversation [C]. IEEE International Conference on Multimedia and Expo. 2024: 1-6. (CCF B, 导师第一作者，学生唯一通讯作者)

[4] Huan Zhao, Gong Chen, Zhijie Yu, Yingxue Gao*. Graph-Based Emotion Consensus Perception Learning for Multimodal Emotion Recognition in Conversation [C]. IEEE International Conference on Acoustics, Speech and Signal Processing. 2026: 11642-11646. (CCF B, 导师第一作者，学生唯一通讯作者)

[5] Huan Zhao, Zhijie Yu, Yong Wei, Bo Li, Yingxue Gao*. DSSR: Decoupling Salient and Subtle Representations Under Missing Modalities for Multimodal Emotion Recognition [C]. IEEE International Conference on Acoustics, Speech and Signal Processing. 2026: 12077-12081. (CCF B, 导师第一作者，学生唯一通讯作者)

[6] Yingxue Gao, Huan Zhao, Yufeng Xiao, Zixing Zhang. GCFormer: A Graph Convolutional Transformer for Speech Emotion Recognition [C]. IEEE International Conference on Multimodal Interaction. 2023: 307-313. (CCF C, 学生第一作者)

[7] Yingxue Gao, Huan Zhao*, Zixing Zhang*, Keqin Li. Mitigating Modality Imbalance in Multimodal Emotion Recognition via Dynamic Gradient Shaping [J]. IEEE Transactions on Affective Computing. (SCI 一区，学生第一作者，评审中)

[8] Yingxue Gao, Huan Zhao*, Zixing Zhang*. DyGIN: Modality-Unified Dynamic Graph Inception Network for Context-Aware Emotion Recognition [J]. IEEE Internet of Things Journal. (SCI 二区，学生第一作者，大修修回)