0
$\begingroup$

In a classifier model, we can predict the outcome class, but here I need to find out the features that drive towards a particular result in a classification problem, that are a strong indicator of a particular result eg: driver features for loan default or features indicating for a successful sale and so on. Is it the same as finding the important features for the classifier model, if not, how do I proceed towards this. Also which model could be the most useful in such a case. Thanks in advance.

$\endgroup$
0
$\begingroup$

If you want to know what features are important to predict some outcome, you speak about feature importance (FI). FI can differ between models, e.g. between a random forest and logistic regression. This is because the models take a different perspective on how to predict some outcome.

A mathematically sound way to assess FI in binary classification would be to run a logistic regression with a lasso or ridge penalty. This is easy and the results (the log odds) can be directly interpreted. Larger positive or negative values indicate a larger positive or negative FI et vice versa.

Here is a good R example: https://web.stanford.edu/~hastie/glmnet/glmnet_alpha.html

See „Introduction to Statistical Learning“ Ch. 6.2 for some background and demo: https://www-bcf.usc.edu/~gareth/ISL/

| improve this answer | |
$\endgroup$
0
$\begingroup$

I think that some causal feature selection model would work well for you, like the FCI or GFCI algorithm. There's a good library in R called rcausal that implements these algorithms. There's also a corresponding GUI that has some more information about the algorithms used to determine causality.

| improve this answer | |
$\endgroup$
0
$\begingroup$

I found the answer I was looking for at "https://www.kaggle.com/learn/machine-learning-explainability". Hope others with the same doubt might help from it.

| improve this answer | |
$\endgroup$

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.