Implement individual fairness metrics about aequitas HOT 4 OPEN

The sample datasets can be used for implementing this.

from aequitas.

I'm unclear on what the output should be.

For example, Dwork et al. focus on using the Lipshitz condition as a constraint on the optimization function. We want to take the predictions from an unconstrained algorithm and determine how far from the LIpshitz condition it is.

As a first step, would we want two matrices one with distance measured in the feature space and a second with distance measured in the outcome space (from which we could determine a score such as what fraction of the pairs of points fail the conditions)?

Similarly, would the input be the whole feature space for a prediction? This seems necessary, but undesirable, since we would require all the data.

from aequitas.

@saleiro I've outlined how I think this should work in the issue ticket. Does that outline of functions seem reasonable? I'll start with the clustering?

from aequitas.

@anisfeld I suggest that you create a README file within the individual module and outline exactly what you are going to implement. Let's abstract if the features are passed in the df or not. It's up to the final user to decide what she wants to use as representation. They can even pass different dfs based on different representations, train/test splits over time etc...

from aequitas.