2
$\begingroup$

Suppose you need to give a talk about your final model for a machine learning project to the project partner. Take for example that you used a random forest model/neural network, where you cannot really interpret any coefficients. What can/should/should not/is great to present in such a talk? All i can think of is the features used and the test error?

$\endgroup$
  • 3
    $\begingroup$ Results from validation on a hold-out sample, together with crossvalidation results. Maybe also show results from some simpler, interpretable model and show with plots/visualizations how that model have problems with the fit, and how your complex model handles that. $\endgroup$ – kjetil b halvorsen Jun 22 '17 at 8:29
1
$\begingroup$

It depends on the context. Usually, a machine learning project does not stand by itself but is integrated in some kind of operational context. What are they trying to find out and why? One first part of the presentation should probably show that the model is really able to solve this exact question and can be used for this exact task that is needed for operationalization (best on a holdout sample or cross-validation set).

If you can try to give some indication of the business value your model provides. If your model is trying to get new customers (e.g. predict items to recommend), show how many more sales may be expected and why this model increases that number. It may be good to be conservative here, though, since overconfidence can lead to expectations that cannot be met.

Often, however, this also depends on the audience. For more technical oriented audience, test scores etc are just fine. For more business-oriented audiences the actual business value is much more convincing.

$\endgroup$

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.