fairadapt: Causal Reasoning for Fair Data Pre-processing

Causal reasoning can help identify and remove discrimination in machine learning: The fairadapt method uses causal graphical models to understand how sensitive attributes like gender or race influence outcomes, enabling more targeted interventions than traditional fairness approaches.

Counterfactual questions reveal hidden bias: By asking "What would my salary have been if I were a different gender/race?", the method can quantify individual-level discrimination and help organizations understand whether their algorithms are making fair decisions based on legitimate factors versus protected characteristics.

Pre-processing data is a practical fairness strategy: Rather than modifying algorithms themselves, fairadapt adjusts the input data before machine learning models are trained, making it a flexible approach that works with existing ML systems and is accessible through an open-source R package.

Not all causal pathways from sensitive attributes are discriminatory: The method allows organizations to specify which connections between protected characteristics and outcomes should be considered legitimate (for example, if education level mediates the relationship between background and salary), helping distinguish between correlation and actual discrimination.

This approach bridges fairness and explainability: By making causal relationships explicit and enabling counterfactual reasoning, fairadapt helps organizations both mitigate bias and justify their decisions to affected individuals and stakeholders.