MELP: Multi-Embedding-Based Link Prediction in Attributed Networks

Abstract

Attributed networks are complex networks that incorporate the basic network structure of simple networks with node and edge attribute information associated with actors and relationships in a social network. Link prediction on such networks is the task of predicting possible relationships between nodes in such a social network while taking into account both structural and attribute-based information. Node embedding is the task of generating low-dimensional vector features for node representation on a social network graph. Embedding such network data has been of great help in multiple real-world applications, such as node classification and entity retrieval. This paper proposes a unified embedding framework for link prediction called multi-embedding-based link prediction (MELP) in attributed networks. The proposed framework is built on a multi-embedding-based method to link prediction that considers network topology and rich node attributes. For the link prediction problem, the proposed framework takes into account a wide range of node attributes, including structural and attribute proximity, vertex features at several scales, text features of vertices, network structure and content of the node, node context, and node attribute. We have concatenated the properties of seven popular attributed node embedding-based methods and provide rich information for subsequent node embedding-based tasks. Also, we propose an ensemble model to integrate various graph embedding into a new representation of each node because each method has its own strengths and drawbacks. For the task of link prediction in attributed networks, our proposed approach shows enhanced performance on four real-world datasets and four performance evaluation metrics. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.