Dynamic Instance-wise Joint Feature Selection and Classification

Dynamic Instance-Wise Classification in Correlated Feature Spaces

“…In this article, the problem of selecting which features to use to classify each test instance as features sequentially arrive one at a time is explored. The current work extends our prior work [6], which formalized this problem with the simplifying assumption that features are conditionally independent given the class variable. Specifically, building upon [6], herein features ensure consistency: correlation vs dependency are efficiently and effectively modeled using a Bayesian network.…”

“…The current work extends our prior work [6], which formalized this problem with the simplifying assumption that features are conditionally independent given the class variable. Specifically, building upon [6], herein features ensure consistency: correlation vs dependency are efficiently and effectively modeled using a Bayesian network. In general, handling feature dependencies can lead to computationally intractable solutions [7].…”

“…In this subsection, IFC 2 F is compared with i) 2 dynamic feature selection methods: ETANA [6], F-ETANA [6], and ii) 6 streaming feature selection methods: OFS-Density [12], OFS-A3M [14], SAOLA [11], OSFS [10], Fast-OSFS [10], and Alpha-Investing [9]. In streaming feature selection meth- ods [9]- [12], [14], a feature is selected if it satisfies an appropriatelly defined criterion (e.g., belongs in the approximated Markov blanket of the class variable [10], [11], p-statistic is greater than a dynamically varying threshold [9]), or such that the boundary region of the decision is maintained as little as possible [12], [14].…”

“…The optimum solution is also derived for this more general formulation and important structural properties are analyzed that facilitate the design of a scalable method. The proposed method is evaluated and compared with prior work (including our own prior work [6]) on a variety of real-world datasets.…”