Feature-weighted maximum representative subsampling

Area of Science:

• Social Sciences
• Data Science
• Machine Learning

Background:

• Debiasing algorithms are crucial for valid conclusions in social science studies.
• Existing methods using sample weights can introduce bias into representative features when only a subset is biased.

Purpose of the Study:

• To develop a novel debiasing method that minimizes the impact of highly biased features on sample weight computation.
• To address the issue of introducing new biases into representative variables during the debiasing process.

Main Methods:

• Developed feature-weighted Maximum Representative Subsampling (FW-MRS).
• FW-MRS uses feature weights derived from domain classifier importance to down-weight biased features.
• Algorithm iteratively removes elements from non-representative samples to align with representative ones.

Main Results:

• FW-MRS was validated on eight artificially biased tabular datasets.
• The method showed no statistically significant difference in generalization performance on downstream tasks compared to other methods.
• Applied FW-MRS to a real-world social science dataset.

Conclusions:

• FW-MRS effectively debiases datasets by reducing the influence of highly biased features.
• The approach retains more data instances for downstream tasks without compromising generalization.
• Offers a promising solution for debiasing complex datasets in social sciences.