I used nnet package in R to train the neural network and make prediction. At first, because the output values were large, i used the formula (x-xmin)/(xmax-xmin) to standardize them in range of 0 to 1. After training the network, i predicted the output values. The result is a range of data in range of 0 and 1. How can i un standardize the predicted values to have predicted values of the first unit?
thanks
nnet R package can be set up both for regression problems and for classification problems and it supports several regression and classification metrics. Usually, you don't need to do the transformation you mentioned to output variables.
Actually, the set up can be a bit difficult, so I suggest to use caret R package that makes the set up to nnet (and many other R models) very easy. The method you need to specify in order to use nnet in caret is "nnet".
here you can an example how to tran a NN using nnet through caret (for regression).
http://stats.stackexchange.com/questions/21717/how-to-train-and-validate-a-neural-network-model-in-r
For more
- Max Kuhn's caret Manual - Model Building is a great starting point: http://cran.r-project.org/web/packages/caret/vignettes/caret.pdf
- Max Kuhn's Applied Predictive Modeling dedicates 2 chapters to how to set nnet through caret (one for regression and one for classification: http://appliedpredictivemodeling.com/
Dear Gino,
Thanks for information gave me. I tried caret package, but because my data are large, i could not get the right answer. the error criteria are so large and it converged after 2 steps!
I have to standardize the data to make them small, but i do not know how to transform them to the first unit after doing the prediction with neural network?!
Dear Ghazaleh,
the case you describes is pretty interesting. In order to understand better the case I need some more information:
- you're working on a regression problem or a classification problem? In case you are working on a classification problem is it a binary classification problem or a multi-class classification problem? in the latter case, how many classes are there?
- size of feature matrix, i.e. number of predictors (or input variables) and number of observations you are using; are they all numeric or is there some categorical variables?
- which kind of metric are you using (e.g. RMSE, AUC, LOGLOSS)?
Dear Gino,
sorry to be late
- The case i am working is a regression problem.
- I have 1 output variable ( 1 neuron in output layer) , 14 neurons in hidden layer and 6 neurons in input layer. the output variable is in currency ( quantitative). the values of input variables are the mixture of binary and quantitative currency.
- the error criteria that i used is MAD and RMSE
- the number of training data is 1980 and the number of test data is 540
as i mentioned, when i standardized data and use nnet function in R, the result is good . but the problem is that how can i change the predicted values ( of course the standardized ones) into the first unit of currency ( in Dollar or IRR rials) ?
thanks in advance
Dear Ghazaleh,
here are some suggestions based on the information you provided.
As the values of input variables are a mixture of binary (hence, they should belong to range [0,1]) and quantitative currency (hence, they should belong to to range like [1000$, 1000000$]) I suggest to scale the input variables so that all ones belong to the same range [0,1]; so it means applying the transformation (X-mean(X))/sd(X) to all input variables
If you use caret you can just specify this transformation inside the the pre-process object. For example, if XX is your train set (not scaled) of 1980 observations, yy the related response variables (not scaled), and ZZ the test set (not scaled) of 540 observations,
fit
Hi, you need to denormalize your data with the opposite way as you have normalized them. see the following link -> https://stackoverflow.com/questions/39355942/denormalize-data
Hope that it helps!!!
Best regards,
Katerina
- Badges
- Science topic