0
$\begingroup$

i want to build a neural network which gives a forecast for a date in the future.

I got a lot of Training Datas. At first i tried to get some characteristic.

Some lines of the Blank datas:

2005-11-10 05:00:00.000 0
2005-11-10 05:30:00.000 0
2005-11-10 06:00:00.000 1
2005-11-10 06:30:00.000 3
2005-11-10 07:00:00.000 12
2005-11-10 07:30:00.000 36
2005-11-10 08:00:00.000 89
2005-11-10 08:30:00.000 120
2005-11-10 09:00:00.000 209
2005-11-10 09:30:00.000 233
2005-11-10 10:00:00.000 217
2005-11-10 10:30:00.000 199
2005-11-10 11:00:00.000 244

The first variable is the date, the second is the interval. the fourth are the incomming calls.

Now what i think might make sense as inputs for the NN are:

  • Week of the Year
  • Day of the week
  • interval( a interval is a number between 1-48, 30min intervals per day (24*2)

Output:

  • Incoming Calls

I reparsed the Datas. They're looking now like this:

15 5 33 165
15 5 34 122
15 5 35 123
15 5 36 95
15 5 37 51
15 5 38 50
15 5 39 40
15 5 40 37
15 5 41 20
15 5 42 22
15 5 43 16
15 5 44 9
15 5 45 7
15 5 46 2
15 5 47 2
15 5 48 0
15 6 1 1
15 6 2 0
15 6 3 1
15 6 4 0
15 6 5 0
15 6 6 0
15 6 7 1
15 6 8 0
15 6 9 0
15 6 10 0

Are the inputs well selected?

Improve this question
$\endgroup$
8
  • $\begingroup$ "Are the inputs well selected?" - What do you mean by well selected? And what are your thoughts? Are you asking, "how do I choose features for machine learning?" If so, that's also discussed in many standard resources. Basically, the answer depends on (a) what data is available, and (b) what data you suspect will be predictive, based on your domain knowledge of the thing you're trying to predict. But you should do some reading of standard resources on machine learning if that's what you're really asking. $\endgroup$
    – D.W.
    Commented Nov 16, 2016 at 17:01
  • $\begingroup$ Please don't cross-post your question on more than one Stack Exchange site. If you'd prefer for your question to appear on Stats.SE, please click "flag" under your question to flag it for moderator attention and ask the moderators to migrate it to Stats.SE. You can choose which site you want the question to appear on -- either choice is OK. You might get better answers on Stats.SE. But I recommend you first address the feedback you've already received, to make sure you don't just get the same feedback over on Stats.SE. $\endgroup$
    – D.W.
    Commented Nov 16, 2016 at 17:02
  • $\begingroup$ I recommend encoding cyclic elements onto unit circle. For example day of week will be encoded as two parameters instead of one: cos(day_of_week*2*pi/7) and sin(day_of_week*2*pi/7). The reason is that there is 1 day difference between day 2 and day 3, but also between day 7 and day 1. When you encode as 1,2,3,4,5,6,7 the distance between day 1 and 7 creates an asymmetry. When you encode it over unit circle day neighborship is maintained. Same for week of year or hour of day (or half hours). $\endgroup$
    – Cagdas Ozgenc
    Commented Nov 16, 2016 at 17:48
  • $\begingroup$ Thanks for your comment, thats makes sense. Many Values are now negative. Thats no problem or? @cagdas $\endgroup$
    – T.Setso
    Commented Nov 16, 2016 at 18:15
  • $\begingroup$ No not at all. For neural networks it is good practice to normalize all inputs to [-1,1] range anyhow, otherwise they saturate the activation functions in hidden layers. $\endgroup$
    – Cagdas Ozgenc
    Commented Nov 16, 2016 at 18:21

1 Answer 1

Reset to default
0
$\begingroup$

I wanted to tell you that I solved it differently: I have not solved it as a time series, but as follows: I surrender my network The features as input, and get as output a real value (call volume) This works very well, and Even without exogenous factors. Also I use the coding you suggested to me to ensure the neighborhood (unit circle, with cos and sin formula). The question I ask myself is why does it have advantages, for example, to pass the day of the week as two parameters instead of coding it as 1,2,3,4,5,6,7? My mathematics knowledge is unfortunately not so good, to understand immediately. Do you have a good explanation for me? Do you think I would get the same results if I encode my data differently? So if I give it only as a parameter with, but normal normalize them between 0-1 @CagdasOzgenc

Improve this answer
$\endgroup$
3
  • $\begingroup$ @CagdasOzgenc That is to you $\endgroup$
    – T.Setso
    Commented Jan 16, 2017 at 1:33
  • $\begingroup$ I do not understand this answer. $\endgroup$
    – Michael R. Chernick
    Commented Jan 16, 2017 at 2:10
  • $\begingroup$ This answer is dedicated to @CagdasOzgenc, i had not enough space to write this as a comment. $\endgroup$
    – T.Setso
    Commented Jan 16, 2017 at 14:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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