Towards Context-Aware Emotion Recognition Debiasing From a Causal Demystification Perspective via De-Confounded...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Context-Aware Emotion Recognition (CAER) leverages contextual information for improved emotion understanding.
• Existing CAER models suffer from context bias in datasets, leading to unbalanced emotional state distributions and biased learning.
• This bias acts as a confounder, causing models to learn spurious correlations and limiting performance.

Purpose of the Study:

• To address the context bias dilemma in CAER datasets.
• To disentangle emotion recognition models from the impact of biased learning using causal inference.
• To develop a module that mitigates confounding effects in visual representation learning.

Main Methods:

• Formulated causalities in the CAER task using a customized causal graph.
• Developed a Contextual Causal Intervention Module (CCIM) based on backdoor adjustment theory.
• Integrated CCIM as a plug-and-play component to de-confound models during training.

Main Results:

• CCIM effectively de-confounds the learning process by addressing context bias.
• The module facilitates the seeking of approximate causal effects, improving model robustness.
• Systematic experiments on three datasets demonstrated significant performance improvements with CCIM integration.

Conclusions:

• Causal inference offers a powerful approach to mitigate bias in CAER.
• The proposed CCIM module is an effective and versatile solution for improving context-aware emotion recognition.
• This work advances the field by providing a method to overcome dataset limitations and enhance model generalization.