Abstract
Transductive learning, which is a special case of semi-supervised learning, is a task to estimate labels of whole data using known labels of a piece of data. Differently from supervised learning, we can use distribution of whole data to estimate the labels in the transductive learning. In this research, we assume that the members of each cluster in data have the same label, and use the orthogonal representation of the Lovasz number for the transductive learning, which has an effect of cohesiveness for clusters. We construct a classifier in the space of the orthogonal representation to achieve the trainsductive learning.
