LOCAL CASE-CONTROL SAMPLING: EFFICIENT SUBSAMPLING IN IMBALANCED DATA SETS.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Class imbalance in classification presents challenges for model parameter estimation.
• Subsampling methods reduce computational costs but can inflate variance.
• Standard case-control sampling may lack consistency under model misspecification.

Purpose of the Study:

• To develop an efficient subsampling method for logistic regression in imbalanced datasets.
• To improve parameter estimation accuracy and consistency compared to existing methods.
• To address the trade-off between computational efficiency and statistical variance.

Main Methods:

• Proposes an accept-reject scheme to adjust class balance locally in feature space.
• Utilizes a pilot estimate to preferentially select rare examples.
• Employs a post-hoc analytic adjustment to correct for biased subsampling.

Main Results:

• The proposed method generalizes standard case-control sampling.
• Achieves consistency for population risk-minimizing coefficients under pilot estimate consistency.
• Demonstrates substantial performance improvements over standard case-control subsampling in simulations and real-world data.

Conclusions:

• The novel subsampling technique offers an efficient and statistically robust approach for imbalanced classification.
• Provides a consistent estimator that outperforms traditional methods, especially in severely imbalanced scenarios.
• The method is simple, parallelizable, and adaptable for improved variance reduction.