Model-free feature selection to facilitate automatic discovery of divergent subgroups in tabular data

Model-free approach reduces bias: Unlike existing feature selection methods that rely on training specific models, this paper's SAFS framework avoids "model-bias and spurious correlations" by selecting features without fitting a particular model first, making the process more objective and trustworthy.

Addresses scalability in data understanding: The research tackles a practical problem in responsible AI—manual data stratification (dividing data by characteristics like gender) doesn't scale well as datasets grow. Automatic discovery of divergent subgroups helps organizations understand how their data breaks down across multiple features simultaneously, which is essential for identifying potential fairness issues.

Computational efficiency gains: SAFS demonstrated significant time reduction in feature selection compared to six existing methods when tested on real healthcare (MIMIC-III) and insurance (Allstate Claims) datasets, making responsible AI practices more feasible for organizations with limited computational resources.

Supports data-centric AI principles: By enabling easier extraction of "data-centric insights" without requiring extensive model design and tuning, this work aligns with the growing recognition that understanding data quality and characteristics is foundational to building trustworthy AI systems.

Practical for identifying subgroup disparities: The method facilitates automatic discovery of how different subgroups in data may experience different outcomes, which is critical for detecting and mitigating algorithmic bias and discrimination in AI applications.