4
$\begingroup$

Our response variable is something like Sales which should be very large(Mean at the million level), and one predictor is also numeric but with mean at ten thousand level. The other predictors are all dummy variables.

From my perspective, there are too many dummy variables which mean linear model will not perform well lm(Y~ x1 + Dummies). The reason why I thought about neural network is that I really need make a good fit and make a decent prediction about the Sales.

Should I standardize the whole dataset and then build the neural network model?

$\endgroup$
7
  • $\begingroup$ For a neural network i would aim to capture as much of the behaviour already in a linear model. I would then guess it would make more sense to model the log of sales, on the assumption that the predictors tend to have a multiplier effect on sales rather than additive. please edit your question to explain where this question is coming from . are you a student? why have you chosen neural network ? $\endgroup$
    – seanv507
    Sep 18 at 10:00
  • $\begingroup$ @seanv507 I just add my thoughts $\endgroup$ Sep 19 at 7:11
  • $\begingroup$ a regularised linear model will perform well with lots of dummy variables. just make sure you use sparse matrices so you don't run into memory issues. What I am saying is that I suspect your dummies don't have an additive effect but a proportional effect $\endgroup$
    – seanv507
    Sep 19 at 8:05
  • $\begingroup$ therefore it might make sense to model lm(ln(Y) ~ln(X1) +dummies). See eg dummies.com/article/business-careers-money/business/economics/… $\endgroup$
    – seanv507
    Sep 19 at 8:13
  • $\begingroup$ a linear model is a good baseline. what I am saying is that a neural network starts with a linear combination of inputs. so whatever makes your relationships more linear ( eg taking logs - for this case) will aslo help the neural network. $\endgroup$
    – seanv507
    Sep 19 at 8:15

1 Answer 1

3
$\begingroup$

Standardizing would be advisable. Neural networks usually do not behave well with large numbers due to disproportionate gradient steps.

$\endgroup$
1
  • $\begingroup$ Totally agree, if I don't standardize the whole dataset, there will raise a error algorithm did not converge in 1 of 1 repetition(s) within the stepmax. $\endgroup$ Sep 18 at 7:39

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.