Output of Neural Net operator

tsiklisloltsiklislol Member Posts: 2 Contributor I
edited November 2018 in Help
Hello everyone.

I am having difficulty in understanding the output of rapid miner using the neural network operator.

Let me explain the task at hand and i really hope that someone may be able to assist me.
Firstly i am trying to predict the dwell time of a container into a port terminal.
I have big datasets with attributes like, DT ( Days), day of container release from the port (day of the week), month of container release from the port (month of the year) , type of container ( reefer, general cargo), size of container (20ft,40ft) and type of cargo transfered .

Appart from the DT (which is the value i want to predict) i have transformed all other data into dummies (0 not true,1 true)

With running linear regression i have a R^2= 0.165
With neural nets i have R^2= 0.249 with just 500 training cycles and all values in the operator set as default
and if i add training cycles and hidden layers (4 hidden layers and 10.000 training cycles) i got R^2=  0.363

However, i don't know whats the best number of hidden layers i need to use or how to interpret the output of the system.
I really would appreciate any help since thats the subject of my dissertation for my masters degree.



Thanks in advance.
Tagged:

Answers

  • Immo001Immo001 Member Posts: 4 Contributor I
    You can find the right number of layers and neurons only by trial and error, as far as i know. I read it in many articles.

    I hope this information is useful for you.
  • mschmitzmschmitz Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 1,874  RM Data Scientist
    As Immo pointed out, the right number of hidden layers can just be found by trying it out.

    Keep in mind that you add more complexity to the problem the more layers you add. That might also introduce overfitting and the usual problem of ANN that they might do "something" in the region where they are not trained at.


    Cheers,
    Martin
    - Head of Data Science Services at RapidMiner -
    Dortmund, Germany
  • tsiklisloltsiklislol Member Posts: 2 Contributor I
    Thanks for your answers.

    The point is that if i use 5+ hidden layers the squared correlation is 0.000 and from 1-4 the 4 hidden layers produce the best R^2. So i think 4 hidden layers is the way to go.
Sign In or Register to comment.