Debiased Multimodal Personality Understanding through Dual Causal Intervention
Researchers from Capital Normal University and partner institutions have introduced a novel approach to mitigate bias in multimodal personality understanding, a critical component of human-centered artificial intelligence. Addressing the issue of subject bias stemming from diverse demographic backgrounds, the team constructed a Structural Causal Model (SCM) to analyze these impacts. They proposed the Dual Causal Adjustment Network (DCAN), which utilizes a Back-door Adjustment Causal Learning module to block spurious correlations from observable factors and a Front-door Adjustment module for latent biases. To support this research, they created the Demographic-annotated Multimodal Student Personality (DMSP) dataset. Experiments on the CFI-V2 benchmark and the new DMSP dataset demonstrated that DCAN significantly improves prediction accuracy, reaching over 92%, while also enhancing fairness metrics like equal opportunity and demographic parity by substantial margins. This work aims to ensure fairer and more accurate personality trait analysis in AI systems by achieving causal disentanglement of representations.
Wire timeline
Debiased Multimodal Personality Understanding through Dual Causal Intervention
Researchers from Capital Normal University and partner institutions have introduced a novel approach to mitigate bias in multimodal personality understanding, a critical component of human-centered artificial intelligence. Addressing the issue of subject bias stemming from diverse demographic backgrounds, the team constructed a Structural Causal Model (SCM) to analyze these impacts. They proposed the Dual Causal Adjustment Network (DCAN), which utilizes a Back-door Adjustment Causal Learning module to block spurious correlations from observable factors and a Front-door Adjustment module for latent biases. To support this research, they created the Demographic-annotated Multimodal Student Personality (DMSP) dataset. Experiments on the CFI-V2 benchmark and the new DMSP dataset demonstrated that DCAN significantly improves prediction accuracy, reaching over 92%, while also enhancing fairness metrics like equal opportunity and demographic parity by substantial margins. This work aims to ensure fairer and more accurate personality trait analysis in AI systems by achieving causal disentanglement of representations.
cs.AI updates on arXiv.org