0
$\begingroup$

What is the basic philosophy behind feature selection and modelling? How do you actually start? Could you please share your real (practical) inputs?

Bit of background: I am actually trying to analyse incoming service desk requests (emails) and predict/assign it to the concerned department (team).

I have identified 2 features, but they are giving ~70% prediction accuracy, because of few ambiguous entries. So, need to add more features to strengthen my model.

I am sure your inputs will help newbies (like me) to get the initial idea.

$\endgroup$
  • $\begingroup$ Any pointers please? I'm struggling to grasp this particular concept because most of the online blogs/tutorials will demonstrate on the test data sets (ex. Iris, Breast cancer, etc. packaged in Pandas). None focuses on the feature selection part. $\endgroup$ – ranit.b Feb 9 '19 at 20:38
  • $\begingroup$ To the moderators/admins, I strongly feel there should be a mandate for the name of user and a comment box when someone down votes a post. Now, I'm clueless who has down voted this question and for what reason. it's just so irritating! $\endgroup$ – ranit.b Feb 12 '19 at 11:55
1
$\begingroup$

In a lot of cases, you deal with too many features and you either try to reduce the dimensionality of your feature space with techniques like PCA or you try to select only the relevant ones. The latter is known as feature selection. However, what you have is a situation where you don't have a lot of features and the information you get from those does not seem to be enough for a solid classification.

Generally speaking, the best options you have in this case are cleaning your data as good as possible and gathering more features. This can either be done by using new sources of information or by deriving new information from existing sources. Usually, you derive the best feature candidates for your model from the literature of the respective subject area. Domain knowledge helps a lot!

In your case, I would guess that variables like time of day, weekday and day of the month could provide some valuable information, apart from the mail content of course. But these are simple guesses. In the end, you will have to dig into the literature and find the right features by trial and error.

Since you are trying to classify incoming emails, I assume you have access to the mail subject and the mail body. In this case, I would recommend starting with a simple Naive Bayes algorithm, since it is easy to use, effective and there are many implementations and guides how to use them. Also, Naive Bayes can easily be applied to multiclass problems like yours.

| improve this answer | |
$\endgroup$
  • $\begingroup$ what you guessed is exactly my problem and what you suggested is what exactly going inside my head. :) Yes, I do have email subject and body, and I'm about to start considering that using techniques like BagOfWords. $\endgroup$ – ranit.b Feb 11 '19 at 13:48

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.