- Alan Turing Institute (EPSRC)
This project transforms the concept of counterfactual explanations into a practically useful tool for explaining automated black-box decisions.
As we deploy decision-making systems in the real world, questions of accountability become increasingly important. For system builders questions such as “Is the system working as intended?”, “Do the decisions being made seem sensible?” or “Are we conforming to equality regulation and legislation?” are important, while a subject of the decision-making algorithm may be more concerned with questions such as “Am I being treated fairly?” or “What could I do differently to get a favourable outcome next time?”
These issues are not unique to computerised decision-making systems. But with the growth of machine learning based systems these questions have become even more important. What distinguishes machine learning is its use of arbitrary black box functions to make decisions. As such, the functions used to make decisions may well be too complex to comprehend; and it may not be possible to completely understand the full decision-making criteria.
Recently, the project investigators presented a proof of concept for counterfactual explanations, This a novel type of explanation of automated decisions that describes minimum conditions that would have led to an alternative decision.
With a proof of concept already in place, this project undertakes the essential further work to transform counterfactual explanations into a practically useful tool to generate explanations for users and parties affected by automated decision-making systems.