CARAT: Contrastive Feature Reconstruction and Aggregation for Multi-Modal Multi-Label Emotion Recognition

TL;DR

This work tackles multi-modal multi-label emotion recognition (MMER) by addressing modality specificity and label–modality dependencies. It introduces CARAT, a three-stage framework combining label-wise attention, reconstruction-based fusion with supervised contrastive learning, and shuffle-based aggregation to exploit both modality complementarity and label co-occurrences. A two-level reconstruction pipeline (FRR and SRR) guided by a contrastive latent space with prototypes enables modality-aware, label-specific representations, while a max- pooling-like mechanism selects the most informative modality per label. Empirical results on CMU-MOSEI and M3ED show CARAT achieving state-of-the-art performance and robust gains under aligned and unaligned settings, with ablations confirming the contribution of each component. The approach advances MMER by effectively modeling modality-to-label dependencies and cross-modal co-occurrences, with practical implications for robust, real-world emotion analysis across diverse data collections.

Abstract

Multi-modal multi-label emotion recognition (MMER) aims to identify relevant emotions from multiple modalities. The challenge of MMER is how to effectively capture discriminative features for multiple labels from heterogeneous data. Recent studies are mainly devoted to exploring various fusion strategies to integrate multi-modal information into a unified representation for all labels. However, such a learning scheme not only overlooks the specificity of each modality but also fails to capture individual discriminative features for different labels. Moreover, dependencies of labels and modalities cannot be effectively modeled. To address these issues, this paper presents ContrAstive feature Reconstruction and AggregaTion (CARAT) for the MMER task. Specifically, we devise a reconstruction-based fusion mechanism to better model fine-grained modality-to-label dependencies by contrastively learning modal-separated and label-specific features. To further exploit the modality complementarity, we introduce a shuffle-based aggregation strategy to enrich co-occurrence collaboration among labels. Experiments on two benchmark datasets CMU-MOSEI and M3ED demonstrate the effectiveness of CARAT over state-of-the-art methods. Code is available at https://github.com/chengzju/CARAT.

Figures (7)

• Figure 1: An example of MMER (left) and correlations between two relevant emotions and the video sequence (right).
• Figure 2: The overall structure of CARAT with three sequential steps (up). Detailed implementations of two-level feature reconstruction network, contrastive representation learning network, and Max Pooling-like network (bottom).
• Figure 3: The semantic diagram of shuffle. The shuffle is conducted in both batch and modality dimensions. Different colors represent different modalities. After shuffling, new co-occurrence relations can appear in the training set (the gray dotted boxes).
• Figure 4: (a) The visualization of modality-to-label dependencies, indicating the correlation of labels in each row to modality in each column, where darker colors indicate stronger correlations. (b) Two cases of emotion recognition by multiple methods.
• Figure 5: t-SNE visualization of embeddings without/with (left/right) CL. Different colors represent different modalities and different shades represent different emotions.