Link prediction based on feature representation and fusion

Link prediction is one of the core problems in social network analysis. Considering the complexity of features in social networks, we propose a link prediction method based on feature representation and fusion. Firstly, based on the sparseness and high-dimensionality of network structure, network embedding is applied to represent the network structure as low-dimensional vectors, which identifies the spatial relationships and discovers the relevance among users. Second, owing to the diversity and complexity of text semantics, the user text is converted into vectors by word embedding models. As user behaviors can reflect the dynamic change of links, a time decay function is introduced to process the text vector to quantify the impact of user text on link establishment. Meanwhile, to simplify the complexity, we choose the top-k relevant users for each user. Finally, due to the attention mechanism can improve the expression of user's interests in text information, a link prediction method with attention-based convolutional neural network is proposed. By fusing and mining structural and text features, the purpose of synthetically predict link is finally achieved. Experimental results show that the proposed model can effectively improve the performance of link prediction. © 2020 Elsevier Inc.