Tree-structured multiclass probability estimators

Abstract

Nested dichotomies are used as a method of transforming a multiclass classification problem into a series of binary problems. A binary tree structure is constructed over the label space that recursively splits the set of classes into subsets, and a binary classification model learns to discriminate between the two subsets of classes at each node. Several distinct nested dichotomy structures can be built in an ensemble for superior performance. In this thesis, we introduce two new methods for constructing more accurate nested dichotomies. Random-pair selection is a subset selection method that aims to group similar classes together in a non-deterministic fashion to easily enable the construction of accurate ensembles. Multiple subset evaluation takes this, and other subset selection methods, further by evaluating several different splits and choosing the best performing one. Finally, we also discuss the calibration of the probability estimates produced by nested dichotomies. We observe that nested dichotomies systematically produce under-confident predictions, even if the binary classifiers are well calibrated, and especially when the number of classes is high. Furthermore, substantial performance gains can be made when probability calibration methods are also applied to the internal models.