MFRep: Joint user and employer alignment across heterogeneous social networks

Nowadays people join multiple social networks simultaneously to enjoy a variety of social services. Apart from common users, social networks also share other information entities such as organizations/companies. Aligning common information entities across heterogeneous social networks is challenging, but will facilitate many applications. Existing approaches for network alignment do not fully leverage indirect relations among users (i.e., friends of friends). This leads to relatively poor performance, especially when there are little overlapping social relations between different networks. Meanwhile, the mutual promotion between two kinds of information entity alignment is often neglected. In this paper, we propose a novel Matrix Factorization based Representation learning ( MFRep ) framework. MFRep investigates the joint learning for user and employer alignment across different networks. (1). We first compute the cross-network similarities in user attributes to extract seed potential anchor users. Its efficiency can be boosted by considering the similarities in employer properties. (2). Considering social relation differences across networks, we construct user relational matrix to preserve multi-step relational information. This embodies the information propagation from known and seed potential anchor users to other potential anchor users. (3). We extend the user relational matrix to preserve consistent associations between users and employers across networks. (4). We perform semi-supervised user representation learning and unsupervised employer representation learning concurrently via efficient matrix decomposition. The correspondence between users/employers across networks can be inferred based on vector relevance. Extensive experiments on real-world social network datasets justify the utility of MFRep , even with dissimilar social structures across networks. MFRep illustrates the mutual promoted alignments of users and employers with a middle degree of scalability.