Publication classification prediction via citation attention fusion based on dynamic relations

摘要

Publication classification prediction aims to use existing publication information to infer the classification of uncategorized publications. With the popularity of online social networks, the researches based on citation networks are increasing. However, most of the existing models ignore the difference in the importance of citations. Therefore, this paper proposes an attention fusion method called Influencer-Affected Attention Fusion (IAAF). The IAAF method first considers the dynamic relationship between the publications and assigns different attention intensities to the publication’s neighbors according to the difference in importance of the front neighbors and the rear neighbors. Then the memory unit is constructed to enhance publication features according to the publication’s different attention to citations. Based on the Influence-Affected Attention Fusion method, a new deep learning model for publication classification prediction is proposed, called Deep Belief Network based on Influence-Affected Attention Fusion (IAAF-DBN). The model utilizes the multi-layer IAAF-Restricted Boltzmann machine (IAAF-RBM) to collect more hidden information. Meanwhile, the contrast divergence and back propagation method are used to optimize the model. The experimental results on two real citation network datasets show that the proposed model is superior to all comparative models in terms of prediction accuracy. The model can accurately predict the classification of scientific publications. Furthermore, it can replace manual classification and free people from complicated work. Finally, the paper considers a more accurate publication similarity method, which can provide new ideas for research on citation recommendation and data recommendation.