Building models with SAS Enterprise Miner, SAS Factory Miner, SAS Visual Data Mining and Machine Learning or just with programming

NN model improvement

Accepted Solution Solved
Reply
Occasional Contributor M23
Occasional Contributor
Posts: 5
Accepted Solution

NN model improvement

Hi ,

 

 

How to improve the Neural netwrok model in SAS Eminer(12.1),any options to improve the misclassification and reduce the ASE.

 

Thank you

 

 


Accepted Solutions
Solution
‎05-24-2017 03:29 AM
SAS Employee
Posts: 2

Re: NN model improvement

There are many answers to this question. 

 

The obvious one is to change (probably increase) the number of hidden units in the hidden layer. If divergence is not a problem, then increasing the number of training iterations might also help. And changing the optimization algorithm,or the architecture (say from MLP to NRBF), or even changing the number of preliminary training starts can sometimes improve model fit.

View solution in original post


All Replies
Solution
‎05-24-2017 03:29 AM
SAS Employee
Posts: 2

Re: NN model improvement

There are many answers to this question. 

 

The obvious one is to change (probably increase) the number of hidden units in the hidden layer. If divergence is not a problem, then increasing the number of training iterations might also help. And changing the optimization algorithm,or the architecture (say from MLP to NRBF), or even changing the number of preliminary training starts can sometimes improve model fit.

SAS Employee
Posts: 8

Re: NN model improvement

[ Edited ]

Hello,

 

I would add the following:

 

Include a weight decay value (L2) and tune the value on your validation data.

You mentioned ASE... if your target is interval, consider changing the error function and output activation function to match your target's understood distribution.

 

I can typically achieve performance improvements in my neural networks when I divide and concur the input space.  You can do this in several ways.  One that is easiest for me is to build several networks, giving each network a different subset of inputs and average the predictions of the networks.

I hope this helps.

 

Best,

  Robert

☑ This topic is solved.

Need further help from the community? Please ask a new question.

Discussion stats
  • 2 replies
  • 247 views
  • 0 likes
  • 3 in conversation