First, I should say I'm beginner in data mining and use Matlab. I have a huge dataset: approximately 40 millions observation. It consists of 24 features including numeric and nominal values like below:

id, click, hour, C1, banner_pos, site_id, site_domain, site_category, app_id, app_domain, app_category, device_id, device_ip, device_model, device_typ,e device_conn_type, C14

one observation is:

1.00E+18,   0,  14102100,   1005,   0,  1fbe01fe,   f3845767,   28905ebd,   ecad2386,   7801e8d9,   07d7df22,   a99f214a,   ddd2926e,   44956a24,   1,  2,  15706

All the features and some headers are annonymized.

The problem is to estimate the probability of clicking an advertisement id (click is target feature). What I need is a probabilistic classifier to estimate this that also uses both nominal and numeric features. Lately I used Naive Bayes but it just use numeric and does not give accurate predictions. What is your suggestion?

  • $\begingroup$ Please consider editing your question because we could only guess what does "device_conn_type" or "C4" etc. mean. $\endgroup$ – Tim Dec 12 '14 at 18:29
  • $\begingroup$ some headers and feature values are anonymized. i add an observation corresponding the headers $\endgroup$ – Milad Dec 12 '14 at 18:39
  • 1
    $\begingroup$ I edited your question so that it will be more readable. Next time consider paying more attention to text formatting. Btw, welcome to CV! Consider taking a tour that will help you to know the site better and get the best of it. $\endgroup$ – Tim Dec 12 '14 at 18:47
  • $\begingroup$ I appreciate you Tim to helping me not to holding on this question. $\endgroup$ – Milad Dec 13 '14 at 7:15

If you are a beginner I would recommend to catch up with theory so to have a better understanding of different methods. For example, there is great series of lectures on data mining by Trevor Hastie and Rob Tibshirani available here online. There also two books by the lecturers that are also available freely online: an introductory handbook and more advanced one.

As about a specific method that you could try, consider using logistic regression. This is a pretty typical method for this kind of analysis. Also you have to remember that Naive Bayes classifier does not estimate probabilities precisely because it is based on some simplified assumptions on probability (so it is called "naive"). Because of that you should consider the estimates of Naive Bayes rather as some subjective weights rather than as reliable estimates of probability.

  • $\begingroup$ Logistic Regression classifier uses numeric or categorical nominal values as predictor but my nominal features are not categorical. i can not find any solution how to use these features all together. $\endgroup$ – Milad Dec 18 '14 at 21:35
  • $\begingroup$ The nominal features which are not categorical like site_id, site_domain... $\endgroup$ – Milad Dec 18 '14 at 23:08
  • $\begingroup$ Dear Tim. the header of data and one observation is mentioned in above question. $\endgroup$ – Milad Dec 19 '14 at 8:01
  • $\begingroup$ Yes. because all features and data are anonymized. this is what i have!! dose not matter what is the content of a feature. $\endgroup$ – Milad Dec 19 '14 at 9:52
  • $\begingroup$ it`s my fault. I just want to classify these non categorical nominal data. but in help documents i can just find instruction for categorical nominal values. like this:link $\endgroup$ – Milad Dec 19 '14 at 10:34

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.