Seamlessly Integrating Effective Links with Attributes for Networked Data Classification

Abstract

Networked data is emerging with great amount in various fields like social networks, biological networks, research publication networks, etc. Networked data classification is therefore of critical importance in real world, and it is noticed that link information can help improve learning performance. However, classification of such networked data can be challenging since: 1) the original links (also referred as relations) in such networks, are always sparse, incomplete and noisy; 2) it is not easy to characterize, select and leverage effective link information from the networks, involving multiple types of links with distinct semantics; 3) it is difficult to seamlessly integrate link information with attribute information in a network. To address these limitations, in this paper we develop a novel Seamlessly-integrated Link-Attribute Collective Matrix Factorization (SLA-CMF) framework, which mines highly effective link information given arbitrary information network and leverages it with attribute information in a unified perspective. Algorithmwise, SLA-CMF first mines highly effective link information via link path weighting and link strength learning. Then it learns a low-dimension link-attribute joint representation via graph Laplacian CMF. Finally the joint representation is put into a traditional classifier such as SVM for classification. Extensive experiments on benchmark datasets demonstrate the effectiveness of our method.