6
$\begingroup$

I am new to ML, so please interpret this question accordingly... I am not sure if this is a common issue or not, or if I am thinking about this the right way.

Here is what I am trying to do:

I have a bunch of text fragments which I want to classify into certain topics. The text fragments are the titles of support tickets, so for example the title "My laptop is broken, please help" might get classified into the Hardware category, and the title "I would like a refund for my July bill" might get classified into the Finance category.

So far this is straightforward. However, I have a lot of metadata that would probably be useful to include in my model. For example I know how long somebody has been a customer w/ the company, which could be one feature. I know the age of each customer, which could be another feature. Etc.

What I'm not sure is, what is the best way to combine these metadata features with the text features? For the text features I am using something like tf-idf, so I'll have one feature for each word in the vocabulary, and the feature list will be very long since the vocabulary is large. I suppose I could manually append these metadata features to the end of the vocabulary, but it seems a little ridiculous to append 10 features to a feature vector 100k features long. And I'm not sure if it would work right.

FWIW I am using scikit-learn, but I'm not sure if it has any functionality that would help here.

Improve this question
$\endgroup$

1 Answer 1

Reset to default
3
$\begingroup$

I am not aware of a standard way as such but here's a one thing I'll try. This will contain two models in the pipeline.

  1. Train on the textual data to predict a class (like Fiance, Hardware) and get the model's prediction as a one categorical variable.
  2. Append that categorical variable to the existing metadata featues, and train a new model

I could also replace the step 1 with: rather than outputting a one class (one that has the highest probability), using the whole set of probabilities that the first model predicted for each class and append those numeric features to the metadata featues to use in the step 2.

Improve this answer
$\endgroup$
5
  • $\begingroup$ These are both great ideas! Thank you for your answer. $\endgroup$
    – Stephen
    Commented Aug 12, 2017 at 18:04
  • $\begingroup$ Glad that I could be of help. Note that there should be many more alternative approaches as well :) $\endgroup$
    – akilat90
    Commented Aug 13, 2017 at 15:58
  • $\begingroup$ Hi akilat90, can you recommend some papers which uses this idea of combining text and non-text features? $\endgroup$
    – Kaushik Acharya
    Commented Dec 3, 2018 at 4:46
  • $\begingroup$ @KaushikAcharya Sorry I haven't done any reading on this topic. The method mentioned in this answer is straightforward to implement though. $\endgroup$
    – akilat90
    Commented Dec 3, 2018 at 5:57
  • $\begingroup$ datascience.stackexchange.com/a/9642/15808 David's answer(approach #3) is in similar lines. He mentions ensembling via stacking to combine sparse text features with dense features. $\endgroup$
    – Kaushik Acharya
    Commented Dec 25, 2018 at 2:34

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.