Uploaded image for project: 'Spark'
  1. Spark
  2. SPARK-5133

Feature Importance for Random Forests

    XMLWordPrintableJSON

    Details

    • Type: New Feature
    • Status: Resolved
    • Priority: Major
    • Resolution: Fixed
    • Affects Version/s: None
    • Fix Version/s: 1.5.0
    • Component/s: ML, MLlib
    • Labels:
      None
    • Target Version/s:

      Description

      Add feature importance to random forest models.
      If people are interested in this feature I could implement it given a mentor (API decisions, etc). Please find a description of the feature below:

      Decision trees intrinsically perform feature selection by selecting appropriate split points. This information can be used to assess the relative importance of a feature.
      Relative feature importance gives valuable insight into a decision tree or tree ensemble and can even be used for feature selection.

      More information on feature importance (via decrease in impurity) can be found in ESLII (10.13.1) or here [1].
      R's randomForest package uses a different technique for assessing variable importance that is based on permutation tests.

      All necessary information to create relative importance scores should be available in the tree representation (class Node; split, impurity gain, (weighted) nr of samples?).

      [1] http://scikit-learn.org/stable/modules/ensemble.html#feature-importance-evaluation

        Attachments

          Issue Links

            Activity

              People

              • Assignee:
                josephkb Joseph K. Bradley
                Reporter:
                pprett Peter Prettenhofer
                Shepherd:
                Yanbo Liang
              • Votes:
                3 Vote for this issue
                Watchers:
                7 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved:

                  Time Tracking

                  Estimated:
                  Original Estimate - 168h
                  168h
                  Remaining:
                  Remaining Estimate - 168h
                  168h
                  Logged:
                  Time Spent - Not Specified
                  Not Specified