Robust Regression with Neural Networks Using Iterative Learning Trimmed Elemental Estimators

• Activity Number: 54
• Type: Contributed
• Date/Time: Sunday, August 8, 2004 : 4:00 PM to 5:50 PM
• Sponsor: General Methodology
• Author(s): Matthew Hall*+ and Matthew Mayo
• Companies: University of Kansas Medical Center and University of Kansas Medical Center
• Keywords: neural networks ; robust regression ; elementals

Abstract

It is well known that regression with neural networks (NNs) can be sensitive to the presence of outliers in the data. Since is it common to use the sum of squared errors as a metric when seeking to optimize the weights of the network, neural network linear regression shares many of the robustness problems that ordinary least squares (OLS) does. We will address the robustness problems for neural networks through the use of elemental subsets. Furthermore, we will reframe the work of Mayo and Gray (1997, 2001) in a neural network environment and compare the performance of an iterative learning version of Mayo and Gray's trimmed elemental estimators (TEEs) with that of OLS. We found that by using neural networks with iterative learning TEEs, the robustness of the parameter estimates is greatly improved over traditional methods.