2
$\begingroup$

I am not sure if this is the right forum to ask this.

I have some data of the houses, like their size(in square meters), if they use aircondition, how many residents live in, I have their electricity consumption as well. I want to train any Machine Learning Algorithm to the dataset above, in order to create a model that estimates the houses consumption.

I tried many different algorithms (using weka tool), but I did not have good results. I was said that SVMs could solve this problem, with the right preprocessing. However, i did not have good results either.

Can anyone help me, in the way i should approach this problem?

Thanks in advance

$\endgroup$
3
$\begingroup$

This is a regression problem, meaning that you are trying to approximate a function, as opposed to a classification problem in which you would be trying to reproduce a discrete category.

I think the first step should be to use something simple like linear regression. Did you try that, and if so, what was unsatisfactory with the results?

$\endgroup$
  • $\begingroup$ Thank you very much for your answer. Yes, I have tried that. However I have an error of 90% ... It can be something in my data. Just to mention, I have in total 265 records. I know that is a small set, but I bilieve that is sufficient for my algorithm. My data are stored as integers. I have three variables which are about the size of the house, the children, and the adults residents, as well as zeros and ones (flags) if they have something else for example I have a variable which takes the value 1 if the house has aircondition and 0 otherwise. I have a total of 15 zero-ones varuables. $\endgroup$ – Andreas Mar 11 '13 at 20:13
  • $\begingroup$ I do not know If some preprocessing of the data I mentioned above is required, and if yes, what kind of preprocessing. Thanks again for your answer. I appreciate it $\endgroup$ – Andreas Mar 11 '13 at 20:14
  • 1
    $\begingroup$ How is your error $90\%$? Are you counting it as a total miss if you don't get the exact value? Usually, when trying to predict a continuous value, people use a cost function like the squared error. It might make sense to rescale the output, but be careful that when you don't have a lot of data, you want to use a simple model. Trying a lot of ways to rescale the output is equivalent to using a more complicated model which may tend to overfit on a small sample. $\endgroup$ – Douglas Zare Mar 11 '13 at 21:31
  • $\begingroup$ Well the error I mentioned is the relative absolute error produced by the tool I am using, Weka. Basically while the std dev is 492, my algorithm has about 350 mean absolute error.... something is not going well... $\endgroup$ – Andreas Mar 12 '13 at 8:37
  • $\begingroup$ I wouldn't call that a $90\%$ error. That sounds like linear regression explains up to $50\%$ of the variance. Depending on the information you have, you might not be able to do much better. $\endgroup$ – Douglas Zare Mar 12 '13 at 14:47

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.