A GENERALIZED KERNEL-BASED RANDOM K-SAMPLESETS METHOD FOR TRANSFER LEARNING

Abstract

Transfer learning allows the knowledge transference from the source (training dataset) to target (test dataset) domain. Feature selection for transfer learning (f-MMD) is a simple and effective transfer learning method, which tackles the domain shift problem. f-MMD has good performance on small-sized datasets, but it suffers from two major issues: i) computational efficiency and predictive performance of f-MMD is challenged by the application domains with large number of examples and features, and ii) f-MMD considers the domain shift problem in fully unsupervised manner. In this paper, we propose a new approach to break down the large initial set of samples into a number of small-sized random subsets, called samplesets. Moreover, we present a feature weighting and instance clustering approach, which categorizes the original feature samplesets into the variant and invariant features. In domain shift problem, invariant features have a vital role in transferring knowledge across domains. The proposed method is called RAkET (RAndom k samplesETs), where k is a parameter that determines the size of the samplesets. Empirical evidence indicates that RAkET manages to improve substantially over f-MMD, especially in domains with large number of features and examples. We evaluate RAkET against other well-known transfer learning methods on synthetic and real world datasets.