Joint feature importance

[grantirv]

Motivation
Sometimes features have a natural grouping, for example are sourced from a specific data provider. It is therefore usefull to know the joint importance of group(s) of features. Maybe that data provider is very expensive and you want to decide on a cost/benefit basis if you can drop those features from your model.

Additionally, for large datasets, the feature importance calculation is computationally expensive. It would therefore be useful to specify which features you want importance scores calculated on rather than have to calculate the full set.

Implementation
This PR adds an argument features to the FeatureImp$new call. It works as follows:

• If NULL (the default) behavior is unchanged. i.e. scores are calculated for all features individually.
• If a character vector of feature names, scores are calculated only for that subset of features.
• If a list of character vectors, then joint importance of each group is calculated.

[codecov-io]

Codecov Report

Merging #156 (b446afe) into master (9fdecc0) will increase coverage by 0.06%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #156      +/-   ##
==========================================
+ Coverage   90.66%   90.72%   +0.06%     
==========================================
  Files          17       17              
  Lines        1628     1639      +11     
==========================================
+ Hits         1476     1487      +11     
  Misses        152      152              

Impacted Files	Coverage Δ
R/FeatureImp.R	97.36% <100.00%> (+0.28%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 9fdecc0...b446afe. Read the comment docs.

[christophM]

That's awesome and a new feature many people might be looking for.