Skip to content

XGBoost added features

  • When using gbtree::updater_colmaker::spit_evaluator==ElasticNet, it is possible to add feature specific additive penalties
    • Parameter format - a string "fid:valfid:val,..."
    • When spliting according to id fid, loss-change is decreased by val
    • if a given feature-id is not given in the string, the corresponding penalty is 0
    • In MedXGB intialization, give "feature-name:valfeature-name:val,..."  
  • It is actually XGB standard constraint, but we need some translation to reach the required format
  • Our parameter format: monotone_constraints=f1:d1#f2:d2#...
  • Where:
    • f is part of a unique feature name, and
    • d is a direction: 1 for up and -1 for down
  • Currently NO defense against raw values/format, and
  • Bad outcome (feature format not recognized by XGB) yield all predictions = 0.5 without warning