A Novel Domain Adaptation Solution to the Transductive Transfer Learning Problem

Abstract:

Most classification algorithms rely on the assumption that training and testing data
come from the same distribution. When this assumption is violated, classification
performance can be seriously affected. Transfer learning, a new and increasingly
popular branch of machine learning, seeks to remedy this potential problem. In this
thesis, we introduce a new transductive transfer learning technique that functions
by leveraging two separate classification hypotheses to geometrically align the source
and target datasets before classification. We show several examples of this technique,
and compare it to other methods commonly used in transfer learning.