|
Activity Number:
|
131
- Predictive Modeling in Data Science
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, July 31, 2017 : 8:30 AM to 10:20 AM
|
|
Sponsor:
|
Section on Statistical Learning and Data Science
|
|
Abstract #324658
|
View Presentation
|
|
Title:
|
Bin-Weighted Ensemble Classifiers
|
|
Author(s):
|
Karsten Maurer* and Walter Bennette
|
|
Companies:
|
Miami University and Air Force Research Lab
|
|
Keywords:
|
ensemble classifiers ;
binned partitions ;
weighted ensembles ;
classification
|
|
Abstract:
|
We present a methodology for adapting weighted ensemble classifiers to incorporate localized accuracy estimates for member classifiers using binned partitions of the feature space. We propose the concept of bin-weighted ensemble classifiers, using feature binning to partition the feature space in order to group observations to create locally weighted ensembles. The goal is to fit the member classifiers on the full training set but then adjust the ensemble weights based on empirical accuracies of the members within the specific partition of the feature space where the observation resides. A computational experiment is conducted to evaluate the benefits of bin-weighted ensembles versus those using traditional weighting schema.
|
Authors who are presenting talks have a * after their name.
