Identity alignment algorithm across social networks based on attention mechanism

Abstract

In recent years, more and more business scenarios need accurate cross social network user identity alignment. Existing methods directly splice user attributes and network structure features in the case of sparse network structure, which affects the accuracy of alignment. At the same time, the semantic expression deviation of the same user attribute in different networks further increases the challenge of user identity alignment. This paper proposes a cross social network user identity alignment algorithm based on attention mechanism. Firstly, to learn distinguishing semantic features, a feature extraction method of user attribute text based on attention mechanism is designed. It uses Highway network to dynamically balance the word and character embedding of attribute text, extracts different granular fusion semantic information through convolution neural network with three convolution kernels and uses attention mechanism to give higher weight to key semantics to learn distinguishing semantic features. Secondly, to strengthen the attribute information and weaken the influence of network structure sparsity, a new fusion loss method is proposed. It uses Cosine and cross-entropy loss for attribute feature and fusion feature of attribute and network structure respectively and carries out weighted fusion calculation. The experimental results show that the accuracy of this method can reach 94.95% and the F1 score can reach 92.52% on the Aminer-LinkedIn social network matched data set, which is better than other algorithms.