So i have some dataset, which is basically a count dataset. I have my own code for the classification using neural networks. Turns out that the data does not have a lot of correlation so accuracies as high as 55% is acceptable.For some reason the output of my neural network turns out to be the same prediction no matter what the input is on the testing set.I think this has something to do with the data, as the sigmoid function saturates outputs of layers to 1. The code i have written was tried on 5 different UCI classification datasets and it worked perfectly. Any kind of help is appreciated ,i would not mind sharing the code as well. please do help,i need to know how i can transform the data(normalise or something similar) so that i can get the neural network to learn it.

  • $\begingroup$ Pls describe your data. So your response is a count of some events(?) and you have a bunch of features from which your NN predict. And the response is mostly zero. Correct? $\endgroup$ – horaceT Jun 15 '16 at 21:47
  • $\begingroup$ So the data contains lots of 0's.The input data is the count of some events, which belong to certain classes. The network prediction is putting all the test inputs into the same class(constant prediction).The output of the first sigmoidal layer is saturating to 1,which i guess is because of the data not being of the right form. $\endgroup$ – aditya ramesh Jun 15 '16 at 22:09
  • $\begingroup$ i think you meant to say the features, ie. your input data, have lots of zero. right? $\endgroup$ – horaceT Jun 15 '16 at 22:32
  • $\begingroup$ yes lots of zeros, and yes the features. $\endgroup$ – aditya ramesh Jun 16 '16 at 3:17

Could you be solving the wrong problem? The underlying neural net may be oscillating or any number of things based on the parameters. If you have many zeros you may be up against against the zero inflated model. In this case you would predict whether you will have zero or non-zero counts then on some the actual number. So your net would first be a binary classification problem. Then you would have another model where you only train on non-zero cases.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.