| Previous: Calibration |
Reductions
A machine learning reduction solves a problem by reducing it to a problem that has already been solved. This approachs leverages existing theory from a well-understood problem to solve a problem of interest. The following algorithms compose two machine learning reductions that have been shown to work well in practice.
Beygelzimer, Alina, Varsha Dani, Tom Hayes, John Langford, and Bianca Zadrozny. "Error Limiting Reductions between Classification Tasks." Paper presented at the 22nd International Conference on Machine learning, Bonn, Germany 2005.
The costing algorithm employs rejection sampling proportional to the cost of an example to perform importance weighted classification. The algorithm performs multiple rounds of rejection sampling, averaging all over all classifiers.
Zadrozny, Bianca, John Langford, and Naoki Abe. "Cost-Sensitive Learning by Cost-Proportionate Example Weighting." Paper presented at the Third IEEE International Conference on Data Mining, Melbourne, Florida 2003.
The probing algorithm estimates the probability membership of a classifier even when the classifier produces a discreet output. Moreover, a small error rate for classification theoretically translates to small error in probability estimates. An accurate probability estimate is important in several ways. First, when a confidence in a prediction is just as important as the prediction itself. Second, when a ranking of examples is important. Third, when the classifier is used in a larger system.
The Probing Algorithm
Langford, John, and Bianca Zadrozny. "Estimating Class Membership Probabilities Using Classifier Learners" Paper presented at the Tenth International Workshop on Artificial Intelligence and Statistics, Barbados 2005.
| Previous: Calibration |