Exploring Algorithmic Recourse

  • How do these systems note which variables can or can’t change for an individual? If I remove variables such as age and race from recourse analysis, how do I investigate fairness issues?
  • What type of cost function represents which variables are easier to change than others?
  • How does this all work in the era of GPT-3, Stable Diffusion, DALL-E?

the CARLA paper and benchmark

I started with the CARLA paper on OpenReview. As expected, it’s important to set immutable features and a cost/distance function. The CARLA library includes eleven methods for developing counterfactuals (LIME being a familiar one), settings for hyperparameters, and TensorFlow / PyTorch classifier support. The researchers were skeptical of causal methods because they are based on assumption you have the true causal graph and correlations between variables (classic example being income and savings — people with higher income will find it easier to have higher savings). The team finds no one method that excels in all metrics and datasets.
After the paper’s release, the GitHub repo has a few additional methods and a new dataset: Home Equity Line of Credit (HELOC).

Recourse does not grant equity

I watched a video on algorithmic recourse from December 2021.

What about generative models?

The bread and butter of the data science / modeling / ML commercial work is tabular data, non-neural methods, and ‘boring’ data (trading stocks, maximizing reach of advertisement budget, credit and insurance, etc.) Recourse methods on these models is important. But the headlines are large language models, image generators, translators, etc. etc. so let’s go.

DALL-E generations for ‘professor teaching a lecture on algorithmic recourse’, Sept 15

Additional Reading

CARLA cites Wachter’s 2017 paper for popularizing counterfactual explanations for black box models:

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store