Coarse-grained privileged learning for classification

Abstract

Privileged information, a form of prior knowledge, can significantly enhance traditional machine learning performance through a novel paradigm known as learning using privileged information (LUPI). Although effective, current studies on LUPI require a distinct piece of privileged information per input, and these fine-grained priors are difficult to collect in practice. To this end, this paper proposes a brand new problem of learning with class-wise privileged information, where instances within the same class share identical privileged information. As far as we know, this problem has not yet been explored. We build a support vector machine with coarse-grained class-wise priors (CGSVM+) and put forward a novel and reliable augmenting strategy to solve it. In addition, two datasets are collected from nature reserves in Xinjiang, China, along with their class-wise privileged information annotated by professionals. Extensive experiments demonstrate the effectiveness of CGSVM+, with the best average accuracy of 80.16% (94.70%) and the best average F-score of 79.87% (94.57%) on the plant (animal) datasets.