I am using RNN package in R to do regression. First created the model with training dataset and checked the model output using the same training dataset. It was fairly accurate with the predicted output always following the pattern of targeted variable, even if in absolute terms there was error which is acceptable. But when checked with test dataset, the output was nowhere close to pattern I expected. I have tried changing the parameters in ‘predictr’ function, but that hardly helped. Is this because I over-fitted the rnn model? or what else could be reason.
$\begingroup$
$\endgroup$
2
-
$\begingroup$ Yes, you are either overfitting or the data distribution in your test is different from your training. If you adopt a cross validation scheme instead of a single validation data set, and keep getting the same result, then you are most likely overfitting. I would advise, instead of looking at the prediction vs truth plot, to look at the evolution of the training and validation error across the epochs of training. $\endgroup$– TomJul 9, 2018 at 13:40
-
$\begingroup$ Dear Tom,Thanks for reply .could you elobrate more on the last 2 sentences. I am relatively new to neural networks $\endgroup$– Mahesh SubramaniamJul 9, 2018 at 13:51
Add a comment
|