3
$\begingroup$

For a paper, I am training different models and using LIME to simplify the blackbox models into a transparent decision tree model that I can visualize with view(tree, "mode", "graph").

I don't want to show the original plot, so here is a similar representation of the resulting simple tree given a classification between classes A, B, C, and D:

Simple tree

The questions that I have are the following:

  • given that the tree idenitfies A and B early (first two nodes), does this mean that the model finds it easier to identify these groups when compared to C and D?
  • Similar but a bit different: Does this hierarchy also mean that the classification accuracy of these groups are also affected by this pattern? Currently I see that indeed A has the highest classification accuracy, with B, C, and D following the same pattern.

I have looked through some early posts about interpreting decision trees and I did not see a similar question being asked. CART books are also not including this information. Eventually I would need to find a source that I can cite for the explanation.

Thanks in advance for any input!

$\endgroup$

1 Answer 1

6
$\begingroup$

The answer to your first question depends on what, exactly, you mean by "easier".

The answer to your second question is "no". The fact that A was found on the first branch simply means that a single variable (whichever created the branch) did a good job of distinguishing A from the others, and that this choice did the "best" job of improving the fit, for whatever measure of fit you chose.

I have often seen the nodes that come after many branches have the highest accuracy.

$\endgroup$
6
  • $\begingroup$ Thanks for the quick reply, and in some way you already answered both of my questions. It is easier because it only takes the one variable to say if this is class A or not A. For the other question I learnt that it is not necessarly a universal pattern. Out of curiosity and necessaty I would like to ask you if you can recommend a CART book other than Breiman (1984)? Thanks in advance! $\endgroup$
    – Tino D
    Commented May 22 at 10:36
  • 3
    $\begingroup$ I haven't looked at the literature in a long time, and I downsized my statistics library when I retired. You might want to ask a separate question about good books, with details of what you are looking for, your math background, and so on. $\endgroup$
    – Peter Flom
    Commented May 22 at 10:55
  • 2
    $\begingroup$ Here are threads tagged both "cart" and "reference". $\endgroup$ Commented May 22 at 19:19
  • 1
    $\begingroup$ @TinoD if you want to see how general this easy recongition of A is, you could train four separate models on A-not A, B-not B, C-not C, and D-not D, and see how well they perform. You can also train a random forest on your data (each tree on a random subset) and see if the pattern is consistent. $\endgroup$
    – Davidmh
    Commented May 22 at 20:19
  • $\begingroup$ @Davidmh thanks for the comment David! I already did as part of the analysis. I trained ECOC, SVM, Bagged trees, discriminant and decision tree models for the task. They all performed well (val. Acc >90%). The trend of A, B, C, and D having a decreasing val. Acc was present for all models. I already included that information in the manuscript but I wanted to add more about that given the transparent tree plot of the explained model. $\endgroup$
    – Tino D
    Commented May 23 at 6:30

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.