Causethical ML - Seminar

Dear all,

Tomorrow, you will be participating in the causal fairness block panel. Remember, there will be no presentation, so we will start directly with the panel. For your reference, we will be discussing the following questions/points, time permitting:

• Path specific counterfactual fairness and counterfactual fairness are two different metrics. What are benefits and limitations of each fairness metric? What questions do they answer and how useful are the answers in the fairness debate?
• What defines a sensitive attribute? Is a sensitive attribute always a root node? Would there be another kind of attribute that might cause unfairness in ML other than a demographic attribute?
• What biases are being removed with the presented approaches? How can historical biases be tackled?
• Apart from the need to know the full causal model if the data, what other limitations do these techniques have that might prevent their application to real data?
• In the first paper: Can a resolving variable and a proxy variable be the same in any scenario?
• To the last paper: Could the measure of counterfactual unfairness in combination with accuracy be used to find the best (true?) SCM for a given problem?

Best wishes,
Your CausethicalML Team