
NEURAL NETWORK MODELS
Overview
A graduate level introduction to and illustrated tutorial on neural network analysis.
Why we think it is important: Neural network analysis is a valuable tool for prediction of continuous target variables or classification of categorical target variables. It is robust for noisy and missing data, and is particularly useful when nonlinear relationships which cannot be addressed through data transformations or generalized link functions exist in the data.
New title in 2014:
The full content is now available from Statistical Associates Publishers. Click here.
Below is the unformatted table of contents.
NEURAL NETWORK MODELS Overview 6 Data examples 8 Artificial neural network software 9 Key concepts and terms 10 Abbreviations 10 Types of artificial neural network models 10 Multilayer perceptron (MLP) models 10 Radial basis function (RBF) models 11 Kohonen selforganizing models 11 Networks, nodes, and weights 13 Models 16 Datasets 16 Training, recall, and learning 17 Training dataset considerations 18 Setting learning parameters 20 Convergence 22 Activation functions 23 Normalization 24 Multilayer perceptron (backpropagation) models 25 Overview 25 MLP models in SPSS 26 SPSS input for ANNMLP 26 SPSS output for ANNMLP 40 MLP models in SAS Enterprise Miner 49 Overview 49 Overview of SAS Enterprise Miner steps 50 MLP flow chart 60 Data Partition 60 Modeling 61 Architecture 62 Optimization 63 Model selection criterion 65 Output 66 Model Comparison 73 Scoring 75 MLP Models in SAS PROC NEURAL 77 Overview 77 SAS syntax 77 SAS output 78 Autoneural models in SAS 84 Overview 84 Example 85 Radial basis function models 86 Overview 86 RBF models, data order, and randomization 87 ANNRBF models in SPSS 88 SPSS input for ANNRBF 88 SPSS output for ANNRBF 97 ANNRBF models in SAS 109 Overview 109 Example using SAS Enterprise Miner 110 Neural network modeling in Stata 112 Assumptions 112 Data level 112 Adequate variance 112 Representative training cases 113 Randomization 113 Few outliers 113 Frequently asked questions 113 What are the “NIST Studies” in relation to ANN? 113 What is a backpropagation model? 114 How can I tell if my results are significant? 116 How can I improve the generalization of my model? 117 Explain neural weights 118 Explain activation (transfer) functions 119 Explain settings for learning rate parameters 121 What are strategies for model complexity vs. model parsimony? 123 Explain quartile analysis 124 Is generalized ANN available? 125 Do I need to transform my input variables? 125 Do I need to standardize my input variables? 125 How should I code binary variables? 127 How do I handle “DK= Don’t Know” and similar codes for my dependent variable? 127 What are pretrained networks? 128 What is a PNN model? 128 What is a GRNN model? 128 What are “constructive algorithms” in ANNRBF? 129 What software is available to implement ANN models? 129 What are some drawbacks to use of ANN? 129 Bibliography 132 Appendix A: SAS Optimized Data Step Code 136 Appendix B: SAS Results for the “Score” node 141 Pagecount: 144