Supervised Learning, how to examine the model produced

For a type of supervised learning like ,ClassificationDiscriminant.fit, ClassificationKNN.fit, or NaiveBayes.fit Is there a way to examine the model that you produce so that you identify the columns that where important to determining the label? Like a high integer in column 13 gives a large amount of evidence that this is lable 'passing'

Answer 1

[Ilya]

The answer varies per model.

For linear types of ClassificationDiscriminant, you can use the DeltaPredictor property. This property defines thresholds on the predictor coefficients (predictor means column). If you say "I only want predictors with coefficients above T", you effectively set coefficients for all predictors with DeltaPredictor below T to zero. A large value of DeltaPredictor indicates an important predictor.

ClassificationKNN does not have a built-in method, but you could use function relieff since this function provides estimates of variable importance using nearest neighbor rules.

I have no suggestion for NaiveBayes.

For any data, you can find predictors most relevant for classification using the chosen classifier with the sequentialfs function.

[Jordan]

Thankyou for your help