Transfer affinity propagation-based clustering

Abstract

Designing a clustering algorithm in the absence of data is becoming a common challenge because the acquisition of annotated information is often difficult or expensive, particularly in the new fields. Because transferring knowledge from the auxiliary domain has been demonstrated to be useful, it is possible to develop an appropriate clustering algorithm for these scenarios in view of transfer learning, where useful information from relevant source domains can be used to complement the decision process and to identify the appropriate number of clusters and a high quality clustering result. In this paper, a novel transfer affinity propagation-based clustering algorithm known as TAP is presented for the scenarios above. Its distinctive characteristics can modify the update rules for the two message propagations used in affinity propagation (AP). Specifically, the most representative points called “exemplars” and the preferences in the source domain are considered for helping in the construction of the high-quality clustering model for insufficient target data. With the corresponding factor graph, the addition of a new term in the objective function for AP allows TAP to cluster in a AP-like message-passing manner for transfer learning, i.e., TAP can identify the appropriate number of clusters and can extract the knowledge of the source domain to enhance the clustering performance for target data, even when the new data are not sufficient to train a model alone. Extensive experiments verify that the proposed algorithm outperforms the state-of-the-art algorithms on insufficient datasets.