1
$\begingroup$

I'm researching a regression model to predict a target value that has four features, all of which are categorical.

The categories are not fixed, e.g. one is a customer identifier. How could my model handle making a prediction for a customer identifer it hasn't seen before, based on the remaining features it has been trained with?

I have considered having a model for each of the features that could predict which category label is most similar based on the other three remaining features (or multiple similar category labels could be used and take an average of the target values for those).

My only concern with with this method is that it's not that scalable, I'm going to want to extend the model with more and more categorical features.

Is there some technique that could create an 'unknown' label for each of the features so the model can handle this case or would the prediction likely be completely inaccurate?

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

Technically, you can't. That is one of the limitations of regression models; they are really only effective for values/ranges that they have seen before. Your use of categorical values makes it even more complex. But even with continuous variables, it is not recommended to use regression models for these "unknown" values.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks for the answer. Do you know of any models that are best suited to handling unknown categorical values? Instead of regression, a binary classifier with a target value that represented the confidence or probability of being in either the positive or negative set could work for this application. $\endgroup$
    – Guy Thouret
    Mar 17, 2018 at 0:00
  • $\begingroup$ @GuyThouret if that's the case, then most neural networks would fit your need. Something like an LSTM neural network, with binary settings, would give you a 0 - 1 probability value of being one of your binary choices. $\endgroup$
    – I_Play_With_Data
    Mar 21, 2018 at 18:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.