I have 62 data with 1 output and 7 input variables. i split the data into 40 data for training/validation and 22 data for testing the neural network. the R square is so small ( negative) and the train data fits well but the test data and the prediction of test data do not performs well so i think the net work is overfitted. I computed cv's for different number of hidden nodes, how can i decide which one is better? the more cv is for better model?
Hello,
Probably your input vector is not enough to be generelized otn new data (test). Thus, I think that you network only memorize the training data.
I suggest to try more tunning of learning parameters (learning rate, ...), if this doesnt work you should reconsider another input vector (more descriminant).
hi,
If you want to use cross validation, you can use 10- folds cross validation by splitting your data into 10 parts. Each time one of the ten is used for testing and the 9 parts for the training and you repeat this process.
In matlab, there is a direct function for Cross validation and NN. You can also use the predefined simulator for ANN. Its quite simple in matlab. Plz use it
The idea of cross validation is to split the training set into two: a set of examples to train with, and a validation set. The agent trains using the new training set. Prediction on the validation set is used to determine which model to use.
Please Refer following link for more detail:
http://artint.info/html/ArtInt_189.html
I advice to pre-process the data by normalizing the data with mean 0 and standard deviation 1 e.g. mapstd(set) is used in Matlab 2013a.
If you have only 62 records you could try the extreme form, known as leave-one-out resampling in psychology (62-fold cross-validation) you could use the weka package MLP classifier to see if it helps.Otherwise the solution is to collect more (and representative) data.
You can use sklearn kfold cross validation method to split your data and then feed them your model
- Badges
- Science topic