It seems to me as if the go-to technique that is used to make any prediction model more "interpretable" is to reduce the number of input variables that are used in the model.

I'm wondering if there are any other big picture approaches that one can take to make these models more interpretable?

By a model that is interpretable, I mean a model that is easier to use and understand.

  • $\begingroup$ If by "reduce the number of input variables" you mean some sort of data reduction, that's the only way I've seen as well. But data reduction could mean things like PCA and factor analysis which really don't reduce the number of input variables. $\endgroup$
    – user765195
    Commented Oct 2, 2012 at 3:04
  • $\begingroup$ What do you mean by more "interpretable"? In my view, one way to do this is to use graphs. But since your question is so broad, it's hard to say exactly what kind of graph. $\endgroup$
    – Peter Flom
    Commented Oct 2, 2012 at 10:35
  • $\begingroup$ Sorry! The question is meant to be broad. By "interpretable", I effectively mean a model that can be easy to use and understand but may take longer to compute. Ideally, an "interpretable" model would allow a user to compute predictions by hand, or strive to at the very least. Models which require a computer to carry out predictions would not be interpretable. As an example, I would say that a sparse linear regression is "interpretable" (i.e. 4-6 variables) but an SVM is not. $\endgroup$
    – Berk U.
    Commented Oct 2, 2012 at 12:10
  • 2
    $\begingroup$ You ask about "prediction model[s]", but I'm wonder if you are really wondering about explanatory models. Whether or not a prediction model is useful is based on how easy it is to get the required data to make the prediction, & whether the prediction bears out within a tolerance that is sufficiently narrow for your purposes. Who cares if you can interpret it? Many machine learning prediction models are out & out black boxes, for instance. Interpretability is important when you want an explanation for a given phenomenon. $\endgroup$ Commented Oct 2, 2012 at 17:11
  • $\begingroup$ I don't mean to make too big of a deal out of this, but these terms are often used loosely, & it's worthwhile to at least be clear on the underlying ideas, even if we continue to use 'predictive model' to mean 'explanatory model' (as I often do myself, TBH). Here is a good discussion on CV on the distinction: practical-thoughts-on-explanatory-vs-predictive-modeling. $\endgroup$ Commented Oct 2, 2012 at 17:15

2 Answers 2


We deliver a "verbal.txt" file which explains in a narrative form precisely what the prediction equation is all about. We also deliver a breakout table to detail how much of the forecast is due to each and every variable in the model. Additionally we present the equation in such a way that one can then use arithmetic (multiplication/addition) to compute the forecast. Finally we allow a "what-if" to enable the user to set different values of the cause variables in the model to determine sensitivity and to actually deliver elasticities.

  • $\begingroup$ I think there is some good, useful information here, @IrishStat, but I think some caution is in order re suggesting your software to the OP (even if you state that there is "no obligation" & I fully trust that your intentions are noble), as it violates our site's policies. $\endgroup$ Commented Oct 2, 2012 at 17:22
  • $\begingroup$ Ok . i was simply saying to the OP that if he wanted to see some examples of the things I was pointing out, he could. In this way it is not "imaginary" but something he cold get his hands around as an example of improvements in model presentation style. There was no intention to "sell" but rather to give him an "example". Perhaps I should have actually shown these kinds of output but I thought that was uneeded. $\endgroup$
    – IrishStat
    Commented Oct 2, 2012 at 17:54

Interpretation (or explanation) and prediction are different missions. Asking for a model that does both well, might sometimes be impossible. This will depend on the data generating process. 1. Convert continuous variables to intuitive categories (height-> {tall, short}). While this might bear a cost on goodness of fit, it is much more interpretable. In particular, when allowing for interactions. 2. Stick to interpretable classes of models. A linear predictors, or a CART is much more interpretable than a neural network of random forest. As I previously mentioned, this might (but not necessarily) bear a cost on the goodness of fit and predictions. 3. As you mentioned in the question- a dimension reduction stage before the model fitting could improve interpretability, but not necessarily. Say, PCA could suggest rather weird factors, which are both impossible to interpret and might not improve the prediction (since the dimension reduction was done while disregarding the dependent variable).


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.