Multi-Level Relation-Aware Knowledge Distillation With Hierarchical Fusion for Incomplete Multimodal Sentiment Analysis

In real-world applications, multimodal sentiment analysis (MSA) often suffers from performance degradation due to incomplete data—arising from equipment failure, signal loss, etc. Recently, knowledge distillation has been verified as an effective paradigm for incomplete MSA. However, most existing works rely on instance-level knowledge while ignoring valuable relational information, leading to suboptimal results. To this end, in this paper, we propose a novel multi-level relation-aware knowledge distillation with hierarchical fusion framework, namely Multi-RKD, for incomplete MSA. Specifically, we introduce a Relational Consistency Distillation (RCD) mechanism that transfers inter-sample relationship knowledge to maintain structural consistency between complete and incomplete data. Furthermore, we introduce an Adaptive Response Distillation (ARD) mechanism that dynamically modulates knowledge transfer based on the teacher's relative reliability during training. An attention-based Hierarchical Multimodal Integrator (HMMI) is further developed to progressively enhance each modality by leveraging complementary information from available modalities, so as to fully exploit the shared latent semantics. The experimental results on three well-known datasets show that our method achieves the state-of-the-art performance under incomplete conditions, with average gains ranging from 1% to 4% in most accuracy metrics.