Learning under Class-Balance Change: Distribution Matching via Direct Divergence Estimation
C.Duples Marthinus · Institutional Repositories DataBase (IRDB) · 2014
In most machine learning algorithms, it is assumed that the training and target environment are the same and that the supervisor (teacher) assigns labels to all training samples. In many real-world datasets, these assumptions are however violated due to a changing environment or imperfect supervision. Many of these situations in the classification scenario can be characterized as a change in class balance. In this thesis, three such problems are considered: classification under class-balance change, labeling of unsupervised datasets differing by class balance and classification from partially labeled data. In this thesis we reformulate these problems in terms of divergences. Analysis of existing methods according to this framework reveals that these methods may be interpreted as indirectly estimating divergences. We propose to directly estimate the divergences leading to efficient solutions to the problems. This approach is empirically validated by experiments on several real-world datasets.