Feature shift detection

During training and classification, instances are drawn from the instance space and mapped to the feature space. We focus on the problem of detecting hidden changes in the functions that map instances to feature vectors during classification. We call such changes feature shift and introduce an on-line method for detecting it. Our method is based on a robust similarity measure that uses one-class SVM to monitor distributional changes in the feature space. Unlike previous methods, ours can distinguish between changes in priors and feature shift. The method is empirically evaluated on visual categorization tasks and its advantage verified.