Conference Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 115 - Advances in Clustering and Classification
Type: Contributed
Date/Time: Monday, August 8, 2022 : 8:30 AM to 10:20 AM
Sponsor: Section on Statistical Learning and Data Science
Abstract #323628
Title: Domain-Dependent Classification with Geometric Digraphs
Author(s): Antony Pearson* and Elvan Ceyhan
Companies: Auburn University and Auburn University
Keywords: hybrid classification; nonparametric; prototype selection; graph-based learning

Class cover catch digraph (CCCD) classifiers are a family of non-parametric prototype selection learners. Previous work has demonstrated that CCCD classifiers perform well in the presence of class imbalance, whereas state-of-the-art classifiers require resampling or ensemble schemes to achieve similar performance. Furthermore, one of the two varieties of CCCD classifier, the random walk (RW-) CCCD classifier, performs better than the pure (P-) CCCD classifier when two classes have some level of support overlap. RW-classifiers suffer from computational complexity and are less accurate when classes are separable, but performs well when overlap occurs. In this work we describe a decision framework that combines P- and RW-CCCD classifiers, achieving superior classification accuracy and sub-cubic computational complexity. We also describe a domain-dependent hybrid classification scheme, which first partitions the domain into class-overlap and well-separated regions, then classifies new examples within each region using RW- and P-CCCD classifiers, respectively.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2022 program