Analysts and decision makers seek to understand their data at a subset level as this can reveal insights that are not apparent at global (all records) or local (single record) levels. Unfortunately, there are exponentially-many subsets to consider, and undisciplined, manual inspection covers a tiny fraction of the possibilities. This talk will step through multiple use cases of Subset Scanning and use it to efficiently detect vulnerable/privileged subgroups and detect bias in predictive classification models at the subset level. The code is part of the Linux Foundation’s AI Fairness 360 Toolkit ai-fairness-36....
21 авг 2024
