Abstract

Two linked ensembles are used for a supervised learning problem with rare-event counts. With many target instances of zero, more traditional loss functions (such as squared error and class error) are often not relevant and a statistical model leads to a likelihood with two related parameters from a zero-inflated Poisson (ZIP) distribution. In a new approach, a linked pair of gradient boosted tree ensembles are developed to handle the multiple parameters in a manner that can be generalized to other problems. The result is a unique learner that extends machine learning methods to data with nontraditional structures. We empirically compare to two real data sets and two artificial data sets versus a single-tree approach (ZIP-tree) and a statistical generalized linear model.

Links and resources

Tags

community

  • @yourwelcome
  • @dblp
@yourwelcome's tags highlighted