CIRCUS: A Causal Intervention-Based Framework for Enhancing Counterfactual Fairness in Trained Classifiers.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Ensuring fairness and mitigating bias in AI models is critical for societal acceptance, especially in sensitive applications.
• Counterfactual fairness, grounded in causal inference, is a prominent fairness concept requiring consistent predictions across modified sensitive attributes.
• Existing methods often lack fidelity in structural causal models (SCMs) due to simultaneous generation processes.

Purpose of the Study:

• To mitigate counterfactual biases in AI models through causal intervention.
• To propose a novel framework, CIRCUS, for enhancing counterfactual fairness in classifiers.
• To develop an effective causal intervention and counterfactual generation method.

Main Methods:

• Proposed the causal inference tabular generative adversarial network (CITGAN) for causal intervention and counterfactual generation, enforcing causal consistency via an end-to-end topological process.
• Integrated exogenous variable inference with sequential generation in CITGAN to preserve structural functional dependencies.
• Developed the CIRCUS framework, which generates counterfactually discriminatory samples (CDSs) using causal intervention and applies label preprocessing for bias correction.

Main Results:

• The CIRCUS framework effectively enhances counterfactual fairness while maintaining robust classification performance.
• For Deep Neural Network (DNN) models, CIRCUS reduced MMD_L and MMD_K values by an average of 39.7% and 40.4%, respectively.
• For Residual Network (ResNet) models, CIRCUS achieved reductions of 56.7% and 54.5% in MMD_L and MMD_K values, respectively.

Conclusions:

• The proposed CITGAN architecture and CIRCUS framework offer a robust solution for mitigating counterfactual biases in AI.
• Causal intervention is an effective strategy for enhancing counterfactual fairness in machine learning classifiers.
• The results demonstrate significant improvements in fairness metrics without compromising classification accuracy.