gms | German Medical Science

Interpretable machine learning methods have become very important in recent years to explain the behavior of black box machine learning models. A useful method for explaining single predictions of a model are counterfactual explanations or short counterfactuals. They explain predictions of single data points in the form of 'what-if scenarios': “If these variables had different values, the model would have predicted your desired outcome”. For such explanations to be plausible, they should only suggest small changes in a few variables. Therefore, counterfactuals can be considered as close neighbours of an actual data point, but their predictions have to be sufficiently close to a (different) desired outcome.

Revealing counterfactuals in algorithmic decision systems is particularly valuable for black box machine learning models for medicine, where wrong decisions pose a serious risk of harm to individuals. Counterfactuals explain why a certain outcome was not reached, can offer potential reasons to object against an unreasonable outcome and give guidance on how the desired prediction could be reached in the future. Counterfactuals are also valuable for predictive modelers to investigate the pointwise robustness and the pointwise bias of their model.

Current approaches can compute counterfactuals only for certain model classes or variable types, or they generate counterfactuals that are not consistent with the observed data distribution. To overcome these limitations, we propose the Multi-Objective Counterfactuals (MOC) method, which translates the counterfactual search into a multi-objective optimization problem and solves it with a genetic algorithm based on the Non-dominated Sorting Genetic Algorithm II (NSGA-II). MOC produces multiple counterfactuals that propose various changes to variables and represent different trade-offs between our proposed objectives.

We demonstrate the usefulness of MOC to explain predictions on a real-world medical data set example. Furthermore, we show in a benchmark study with ten different data sets and five different machine learning models that our approach outperforms state-of-the-art methods for counterfactual explanations.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.