A1641
Title: Scalable Bayesian classifier with many relevant features
Authors: Myungjin Kim - Kyungpook National University (Korea, South)
Gyuhyeong Goh - Kyungpook National University (Korea, South)
Jieun Lee - Kyungpook National University (Korea, South) [presenting]
Abstract: Probabilistic classification based on the Bayes rule is often infeasible in high-dimensional settings due to the need to estimate a large covariance structure among features. While Naive Bayes classifiers circumvent this issue by imposing conditional independence, their performance can deteriorate substantially when features exhibit strong dependence, which is common in modern applications. A scalable Bayesian classification framework is developed that accommodates complex dependence among a large number of relevant features without requiring full covariance estimation. The proposed approach employs a Vecchia-type approximation to construct a sparse representation of the joint distribution. A key challenge in this setting is the absence of a natural notion of neighborhood outside spatial contexts. To address this, a data-driven neighborhood construction based on the link between variable selection and conditional independence is introduced. The resulting classifier achieves a balance between statistical accuracy and computational efficiency in high-dimensional regimes. The performance of the method is demonstrated through simulation studies and real data analysis.
