Discriminative machine learning for maximal representative subsampling.

Area of Science:

• Social Sciences
• Machine Learning
• Data Science

Background:

• Biased population samples are a significant challenge in social science research.
• Existing methods for bias mitigation may not fully address complex sampling issues.

Purpose of the Study:

• To introduce two novel positive-unlabeled learning methods, Maximum Representative Subsampling (MRS) and Soft-MRS, for mitigating bias in population samples.
• To evaluate the effectiveness of MRS and Soft-MRS in correcting biased datasets and improving downstream analytical tasks.

Main Methods:

• Developed two machine learning methods, MRS and Soft-MRS, utilizing auxiliary information from representative datasets.
• Trained classifiers to determine sample weights, with MRS iteratively removing instances and Soft-MRS adapting sample weights.
• Validated methods on a biased public census dataset and compared performance against existing techniques.

Main Results:

• Both MRS and Soft-MRS demonstrated effectiveness in reducing bias in artificially created biased datasets.
• Sample weights generated by MRS and Soft-MRS minimized differences and enhanced performance in downstream classification tasks.
• MRS is recommended for classification tasks, while Soft-MRS is suitable for tasks where dependent variable bias is critical.

Conclusions:

• The proposed MRS and Soft-MRS methods offer a versatile machine learning-based approach to bias reduction in social science research.
• These methods provide practical solutions for improving the reliability and generalizability of findings from social science studies.
• The study highlights the applicability of these techniques in real-world scenarios, such as analyzing the influence of resilience on voting behavior.