2
$\begingroup$

I am currently working on a binary classification problem with 1000 rows and 10 features (after feature reduction). This model is for the purpose of business

Despite multiple attempts/approaches to improve the performance of the ML model, I get only around 70% accuracy with traditional models like logistic regression.

When I tried some AutoML solutions, they were able to do an exhaustive search and suggest me Xgboost, LGM models, neural network models and better feature engineering approach that gave me 80% accuracy.

But since my model is for business, I felt it might be hard to explain to the business.

So, my questions are as follows

a) Should I make use of Lime, SHAP etc to explain the output of autoML models (whatever model they use)?

b) I was confused as to choose between interpretability and predictive power. Felt Xgboost for 1000 rows is bit overkill? I thought they are to be used only for large datasets? Can shed some light on this? Is it okay to use boosting models for small datasets?

c) How should the approach be different when ML is used for business and competition (where only accuracy matters)

d) Any tips/suggestions on what is the approach that users here adopt when building a ML model for business

$\endgroup$

1 Answer 1

1
$\begingroup$

I don’t think much of this should be up to you. Your customer wants a product. You’ve given them several options and can explain their relative advantages. Now it’s time for the decision-makers to make decisions.

By analogy, Chevrolet is being reasonable in making a Corvette sports car and Silverado pickup truck, leaving it to me to decide which car suits my needs.

$\endgroup$
3
  • $\begingroup$ thanks, upvoted, So, you suggest whatever model automl uses, I just make use XAI solutions like LIME, SHAP etc and close the chapter with that. $\endgroup$
    – The Great
    Commented Jan 24, 2022 at 10:43
  • $\begingroup$ My point is that you’re not the one who should decide what the customer values. They might value predictive performance over ease interpretation. They might value interpretability at the expense of some performance. Who are you to decide, however, if they want a sports car or a pickup truck? $\endgroup$
    – Dave
    Commented Jan 24, 2022 at 10:48
  • $\begingroup$ Okay, understand that. But as data scientists, shouldn't my job end at explain our predictions using LIME, SHAP etc? After that, I should let them to decide on whatever they wanted to do.. $\endgroup$
    – The Great
    Commented Jan 24, 2022 at 10:51

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.