.png){kind=icon}
.png){kind=icon}
Return
Robust subsampling framework for data with influential points
.png){kind=icon}
PRE
AI
DOI：10.1080/03610918.2026.2658743.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Subsampling is a widely used technique, with its core principle being the extraction of a small, representative subset from a large-scale dataset. Among various subsampling approaches, the D-optimal method has attracted considerable attention due to its ability to maximize statistical inference accuracy. However, a critical limitation arises when the data contain outliers: conventional subsampling methods, particularly D-optimal subsampling, tend to select anomalous points, which can introduce substantial bias and severely compromise analytical performance. Especially, influential points cannot be simply removed prior to modeling, as their identification inherently depends on the underlying statistical framework. To overcome this challenge, we propose a novel robust D-optimal subsampling method. Our approach begins by classifying the original data into three distinct categories-"clean" data, "ambiguous" data, and "outliers"-based on their likelihood of contamination. Leveraging this classification, the method first selects an initial subsample from the "clean" data and then iteratively refines it by conditionally incorporating points from the "ambiguous" subset. This strategy ensures the construction of a robust subsample that closely adheres to D-optimality criteria while mitigating the influence of outliers. Experimental results show our method consistently outperforms existing techniques across datasets with different outlier levels, as well as three real-world applications.
Keywords:
D-optimality
Experimental design
Robust regression
Subsampling
Journal
C
IF:
0.8
Papers:
213
Citations:
4.7K