Adjusting CNN emotion-class probabilities by subtracting fitted gender coefficients, one class at a time, reduced male/female true-positive-rate gaps on a FairFace/DeepFace test set without lowering accuracy.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Mitigating Algorithmic Bias in Multiclass CNN Classifications Using Causal Modeling
Adjusting CNN emotion-class probabilities by subtracting fitted gender coefficients, one class at a time, reduced male/female true-positive-rate gaps on a FairFace/DeepFace test set without lowering accuracy.