MANILA

MANILA is a web-based, low-code application for benchmarking machine learning (ML) models and fairness-enhancing methods. Many methods exist to reduce bias in ML models, but each affects both fairness and predictive performance differently. Choosing the right one usually requires running many comparative experiments, which demands programming skills and can easily go wrong. MANILA helps users compare combinations of models and fairness methods and select the one that offers the best balance between fairness and effectiveness. As a low-code tool, it requires little or no programming.

MANILA is based on an Extended Feature Model, a technique from software engineering for describing the options in a family of related software products and the rules governing how they can be combined. In MANILA, the model describes a general fairness benchmarking workflow as a Software Product Line. Each option, such as a dataset, an ML model, a fairness method or an evaluation metric, is a feature that users can select.

The rules defined between these features guide users as they build their experiments. They prevent incompatible choices that would cause the experiment to fail when it runs. MANILA then runs the selected experiments and presents the results, allowing users to identify the best trade-off between fairness and performance. The authors evaluate the application on how many different experiments it can express and how reliably it produces correct results.

Auto-discovered on 2026-08-05 by OECD Catalogue Automation