I am training a model using the DNN Regressor estimator from the Tensorflow API to predict prices based on 1035 features.

My dataset contains a little over 500 millions inputs and none of the target output is negative. All prices are positive.

After training for 3 epochs, my NN is still making negative predictions.

Is this normal ? Shouldn't the network learn that prices are always positive

  • 1
    $\begingroup$ Hi, welcome to AI.se! This is not normal, but it would hard to diagnose without code. Also, your question might be better suited to DataScience if you need someone to help you find a bug in your implementation. $\endgroup$ – Philip Raeisghasem Mar 20 at 6:39
  • $\begingroup$ @PhilipRaeisghasem I'm wondering if this question is general enough, dealing with theory, that it might be useful for SE:AI, in that there is always going to be some overlap. (We definitely act as a feeder site for DS, so I pro basic DS questions on AI, b/c the majority of the field is statistical nowadays.) $\endgroup$ – DukeZhou Mar 20 at 20:05
  • $\begingroup$ @PhilipRaeisghasem Hey ! Thanks for your comment. The value i am predicting is a percentage change. It usually hovers around 1-3 %. The features I have engineered are my best guess at what determines this price change so they are not perfect. Can it be that percentage change is so close to 0 that for some combinations of feature values, the formula that predicts y (% delta) goes under 0 ? From my understanding, the concept of 0 is based on our framework of math and "0" can be 1 or 2 or whatever. Is it just possible that instead of predicting 1% it is predicting -2%, for a loss of 3 ? $\endgroup$ – user1776576 Mar 21 at 20:12

Your Answer

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

Browse other questions tagged or ask your own question.