Heterogeneous co–transfer spectral clustering

Abstract

With the rapid growth of data collection techniques, it is very common that instances in different domains/views share the same set of categories, or one instance is represented in different domains which is called co-occurrence data. For example, the multilingual learning scenario contains documents in different languages, the images in the social media website simultaneously have text descriptions, and etc. In this paper, we address the problem of automatically clustering the instances by making use of the multi-domain information. Especially, the information comes from heterogeneous domains, i.e., the feature spaces in different domains are different. A heterogeneous co-transfer spectral clustering framework is proposed with three main steps. One is to build the relationships across different domains with the aid of co-occurrence data. The next is to construct a joint graph which contains the inter-relationship across different domains and intra-relationship within each domain. The last is to simultaneously group the instances in all domains by applying spectral clustering on the joint graph. A series of experiments on real-world data sets have shown the good performance of the proposed method by comparing with the state-of-the-art methods.