University of Tübingen Researchers Open-Source ‘CARLA’, A Python Library for Benchmarking Counterfactual Explanation Methods Across Data Sets and Machine Learning Models

10453
Source: https://arxiv.org/pdf/2108.00783.pdf

Machine learning methods are applied to everyday life in various ways, from disease diagnostics, criminal justice and credit risk scoring. Machine learning models have great potential to provide effective support in human decision-making processes but often come with unintended consequences for an end-user–their predictions may be favorable depending on how different organizations employ them. For example, the denial of a loan or parole might have adverse effects on future development.

Counterfactual explanations (CE) provide means for prescriptive model explanations by suggesting actionable feature changes to allow individuals to achieve favorable outcomes in the future. Choosing an appropriate method is a crucial aspect of meaningful counterfactual explanation. It must be decided on with care as many different types are available, all designed to achieve specific goals.

Researchers at the University of Tübingen are unveiling a new python library CARLA (Counterfactual And Recourse LibrAry) which will allow for benchmarking counterfactual explanation methods across different data sets and machines.

CARLA is an innovative open-source library that allows researchers to benchmark predictive counterfactual explanation and recourse CE methods for managing uncertainty in their analyses.

CARLA is a tool with more than ten counterfactual explanation methods combined. It can be easily integrated into your project to gather a real-time analysis of what would happen if something changed. This is possible because CARLO has built-in evaluation measures that check how different solutions compare against one another across data sets. Researchers can also use CARLA to benchmark various counterfactual methods on popular data sets across various ML models.

CARLA is a versatile tool that can be used to solve countless problems. It has been optimized for popular frameworks Tensorflow and PyTorch, making it easier to use these types of programs in projects with custom implementations. It also supports features like immutable data sets and explicitly specified hyperparameters so users can make sure they’re using the best possible strategy from start to finish.

Paper: https://arxiv.org/pdf/2108.00783.pdf

Github: https://github.com/carla-recourse/CARLA