Applying a MRMR feature-selected ensemble model to new data

Hi,

I initially trained an ensemble model using 66 features, which were extracted from my data. I then applied the MRMR feature selection algorithm, limiting the number of features to 20.

I then went on to export the MRMR feature-selected model to my workspace and when inspecting the model parameters, noticed that the RequiredVariables is set to 1x66 cell

I was expecting this to be 1x20 cell i.e. the features that the MRMR algorithm identified as the best to use.

So do I need to retrain my ensemble model using the 20 features that the MRMR algorithm identified as the best to use or should the MRMR feature-selected ensemble trained model I exported into my workspace automatically identify the 20 features to use when presented with new data?

Any help would be greatly appreciated.

Thanks!

Answer 1

[Sai Teja G]

Hi Impala,

I understand that you used ‘MRMR feature selection’ algorithm to limit the number of features to 20.

The 'fscmrmr()' function, which utilizes the 'MRMR' algorithm, provides the feature importance of the predictors. It returns a 1x66 cell, representing the importance of all the features. From this cell, you can select the top 20 important features according to your specific needs and retrain the model.

You can follow the example about Rank Predictors by Importance for more details.

Hope it helps!

[Impala]

Hi Sai,

Thank you for the suggestion - I will select the features and retrain the model, as advised.

Thanks!

Gursharan