Dis4DR: Disentangled Affective and Biometric Features for Multimodal Depression Recognition

Area of Science:

• Computational linguistics
• Affective computing
• Machine learning

Background:

• Multimodal depression recognition (DR) integrates facial, audio, and text data, outperforming single-modal methods.
• Identity and gender information in signals can complicate DR accuracy.
• Data collection challenges and missing modalities hinder existing DR models.

Purpose of the Study:

• To introduce Dis4DR, a multimodal framework enhancing DR through feature disentanglement and privileged knowledge distillation.
• To address challenges of ambiguous identity/gender information and incomplete multimodal data in DR.
• To improve the robustness and generalization of depression recognition models.

Main Methods:

• Feature Disentanglement Network for DR (Dis4DR) separates homogeneous and heterogeneous features.
• Disentangles identity and gender features from depression-related features.
• Employs privileged knowledge distillation to transfer information from complete to incomplete multimodal inputs.

Main Results:

• Dis4DR consistently outperforms existing multimodal DR approaches across AVEC datasets.
• Achieves superior performance even when only a single modality is available.
• Demonstrates up to a 3.27% relative reduction in prediction error compared to Dis2DR.

Conclusions:

• Dis4DR effectively enhances multimodal depression recognition by disentangling relevant features.
• The framework shows significant robustness and generalization capabilities, especially with incomplete data.
• Dis4DR represents a substantial advancement in automated depression recognition technology.