An adaptive subsampling method for large-sample feature screening

A new feature-screening algorithm inspired by the multi-armed bandit problem promises to slash the computational cost of analyzing ultrahigh-dimensional datasets, according to a preprint posted to arXiv. The method, called BanditSIS, reduces the complexity of the standard sure independence screening approach from O(np) to O(√np) while preserving statistical guarantees [1][2]. The work, authored by Xiaxue Ouyang and submitted in September 2025, targets a bottleneck in modern statistics: the sure independence screening (SIS) method. SIS is widely used to filter out irrelevant variables in datasets where the number of predictors p is enormous, but its computational cost scales linearly with both the sample size n and p, making it expensive for large-scale applications [1][2]. Ouyang's proposed algorithm, BanditSIS, reframes the screening process as a multi-armed bandit problem. It progressively increases the subsample size used for evaluation and discards variables that show small empirical marginal Pearson correlations early in the process. This avoids spending computation on unpromising features [1][2]. The result is a theoretical cost reduction to O(√np) [1][2]. The preprint, now in its second version as of June 2026, includes a theoretical analysis that quantifies how the subsample size affects screening accuracy, revealing a direct trade-off between computational efficiency and statistical reliability [2]. The authors demonstrate that BanditSIS retains the "sure screening property" under mild regularity conditions, meaning it is guaranteed to include all truly important features with high probability [1][2]. Numerical experiments on both synthetic and real-world datasets showed that BanditSIS achieved screening and prediction performance comparable to standard SIS while substantially reducing computational time [1][2]. The method is positioned as particularly well-suited for large-sample, high-dimensional settings where computational resources are a constraint [2]. The paper was posted on arXiv, a preprint server that hosts scholarly articles prior to peer review. arXiv was founded in part through the efforts of astrophysicist Joanne Cohn and has become a central hub for rapid dissemination in physics, mathematics, and computer science [5]. The platform assigns a Digital Object Identifier to each submission, making preprints citable in other works [4]. The current version of the BanditSIS manuscript is 2,532 KB, slightly larger than the original 2,506 KB submission [1].