Transfer Knowledge via Relational K-Means Method

Recent years have witnessed a large body of research works on mining knowledge from large volume of data to support decision making, where a primary assumption is that the training and test examples come from a same domain (i.e., the target domain). However, this assumption, in reality, is too rigorous to describe the training examples which may come from a different domain to the target domain (i.e., the source domain). Consequently, using knowledge of source domain to predict the target domain may achieve unsatisfactory results. Under this observation, in this paper, to unleash the full potential of the training examples to formulate genuine knowledge of the target domain, we propose a relational K-means (RKM) model to leverage both source and target domains by transferring knowledge from the source domain to the target domain. By doing so, we could restore the genuine knowledge of the target domain even if the source and target domain might vary from each other dramatically. Experimental results give some useful suggestions on setting the parameters of the RKM model.