Activity Number:
|
175
- Contributed Poster Presentations: Section on Statistical Learning and Data Science
|
Type:
|
Contributed
|
Date/Time:
|
Monday, July 31, 2017 : 10:30 AM to 12:20 PM
|
Sponsor:
|
Section on Statistical Learning and Data Science
|
Abstract #322529
|
|
Title:
|
Mixture DOE Approach for Weights Optimization in Regression Ensembles
|
Author(s):
|
Stanislav Zakharkin*
|
Companies:
|
PepsiCo
|
Keywords:
|
regression ensemble ;
predictive modeling
|
Abstract:
|
Ensemble methods for classification and regression can improve predictions compared to individual algorithms and are an active area of research. In this study, we demonstrate algorithm selection and their weights optimization using Mixture DOE (Design of Experiments) approach. A mixture design was used to generate combinations of weights for various regression algorithms. The R package caret was used to tune parameters and retain the best performing models for each algorithm. Performance was evaluated using multiple rounds of double cross-validation to avoid over fitting and was compared to performance on randomized sets. A quadratic model fitted to performance results was used to select and optimize algorithms weights. The approach was demonstrated on the publicly available data.
|
Authors who are presenting talks have a * after their name.