IBM announced an upgrade to its artificial intelligence (AI) Fairness 360 toolkit on Wednesday. This latest release adds compatibility with scikit-learn data science library and R, and expands accessibility for a larger range of developers, according to a blog post.

SEE: Managing AI and ML in the enterprise 2020: Tech leaders increase project development and implementation (TechRepublic Premium)

The AI Fairness 360 (AIF360) toolkit was initially released by IBM on Github in 2018. The open source package aimed to help developers examine, report, and mitigate bias and discrimination within their machine learning models and throughout the AI application lifecycle.

The toolkit itself contains more than 70 fairness metrics and 11 unique bias mitigation algorithms developed within the research community, designed to translate algorithmic research from the lab into real-life practices throughout industries including finance, human capital management, healthcare, and education, per the blog post.

IBM’s update, however, makes bias detection even more accessible by opening compatibility with R users and scikit-learn.

IBM AI Fairness 360 updates

  • Availability for R users

AI fairness is critical as machine learning models are increasingly used for high-stakes and high-risk decisions. Machine learning analyzes and generalizes patterns within high volumes of data and can inadvertently create hidden biases toward more privileged groups, according to an IBM blog post.

AI Fairness 360 aims to ensure fairness within these processes, and with the update, this fairness can also be applied to those using R programming language.

The AI Fairness 360 R package is an open source library that contains a comprehensive accumulation of metrics for datasets and models to test for discrimination. R users can use the Fairness 360 algorithms to mitigate bias within that data.

  • Compatibility with scikit-learn

Scikit-learn, a data science library, is typically used for training established machine algorithm models, computing basic metrics, and building model pipelines.

While many IBM notebooks use scikit-learn classifiers with pre- or post-processing workflows, switching between AI Fairness 360 algorithms and scikit-learn algorithms previously disrupted the workflow, making the user convert data structures back and forth, according to a blog post.

The latest AI Fairness 360 edition introduces an aif360.sklearn module. In that module, users can find the currently complete scikit-learn-compatible AIF360 functionalities. Not all functionality has been migrated yet, but the hope is to make AIF360 capabilities interchangeable with that of scikit-learn, according to the post.

Algorithms and metrics within scikit-learn can be swapped with debiasing algorithms and metrics. An example the post provided was “instead of a simple LogisticRegression classifier, you can use an AdversarialDebiasing classifier instead of just the recall_score, you can measure the equal_opportunity_difference or difference in recall between protected groups. All of this should be as easy as swapping a line of code.”

IBM warned that it can’t ensure the full compatibility of AIF360 with scikit-learn. Certain scikit-learn preprocessors, like sklearn.decomposition.pca, will omit the sample properties such as protected attributes and cause errors in the AIF360 algorithms later down the road, according to the post.

The old API will still be operational as IBM continues replicating its functionality to the API. Once that is done, IBM may discontinue support for the old API, but that will be communicated before the time comes, as stated in the post.

A recent report from DataRobot found that nearly half (42%) of AI professionals in the US and UK are “very” or “extremely” concerned about AI bias. This bias can compromise a brand’s reputation or lose customer trust. These new capabilities from AIF360 can help more professionals avoid this harmful discrimination.

For more, check out How IBM Watson’s new natural language processing capabilities helps business users on TechRepublic.

Also see

Image: IBM