Unsupervised domain adaptation without source domain training samples: a maximum margin clustering based approach

Abstract

Unsupervised domain adaptation (DA) techniques inherently assume the presence of ample amount of source domain training samples in addition to the target domain test data. The domains are characterized by domain-specific probability distributions governing the data which are substantially different from each other. The goal is to build a task oriented classifier model that performs proportionately in both the domains. In contrary to the standard unsupervised DA setup, we propose a maximum-margin clustering (MMC) based framework for the same which does not consider source domain labeled samples. Instead we formulate it as a joint clustering problem of all the samples from both the domains in a common feature subspace. The Geodesic Flow Kernel (GFK) based subspace projection technique in the Grassmannian manifold is adopted to cast the samples in a domain invariant space. Further, the MMC stage is followed to simultaneously group the data based on the maximization of margins and a classifier is learned for each group. The data overlapping problem is taken care of by specifically learning a SVM-KNN classifier for the potentially unreliable samples per group. We validate the framework on a pair of remote sensing images of different modalities for the purpose of land-cover classification and a generic object dataset for recognition. We observe that the proposed method exhibits performances at par with the fully supervised case for both the tasks but without the requirement of costly annotations.