Multi-relational Neural Topic Model And Its Application In Document Networks Link Prediction

With the rapid development of Internet technology,a large number of documents are generated every day,such as various types of web pages,academic papers,news reports,and government documents.These documents are numerous and can’t be effectively integrated,resulting in information overload.Therefore,extracting useful information from large-scale documents quickly,accurately,and comprehensively has become a hot research direction.Topic models can quickly extract information,and play an important role in text classification,information retrieval,and other aspects.However,in the current modeling process of topic models,most of the focus is on the content of the document itself.There are also models that consider factors such as author,time,and citation links contained in the document,such as author topic models(ATM),dynamic topic models(DTM),and link topic models(RTM).These models have verified that adding the remaining information can improve the performance of the topic model based on the document content itself.However,there is currently no topic model that comprehensively considers multiple network relationships between documents.Therefore,in order to fully utilize all aspects of information contained in documents and capture the relationships between documents.This paper proposes a Multi-Relational Neural Topic Model(MRNTM).This model combines Variational Auto-Encoders(VAE)and multiple neural network structures,making full use of the advantages of both.VAE can use inference networks to learn more representative topics,and neural networks can use topics to learn citation relationships,citation coupling relationships,and author relationships between documents.Through the network relationships between documents,VAE can further promote topic reasoning,ultimately achieving model optimization,and can be used for citation prediction between documents and author collaboration prediction.Finally,the existing Citeulike-a dataset and the self built dataset of 11914 scientific research articles published in different fields in the past seven years collected in the Web of Science database were used to verify the performance of the model in the three tasks of topic learning,citation prediction,and author collaboration prediction.When the number of topics is selected as 50,experimental results on the Citeulike-a dataset show that our model MRNTM has a 0.021 improvement in the score of topic learning tasks compared to NRTM.In the citation prediction task,HR@K and NDCG@K increased by 0.02 and 0.018,respectively.In the author collaboration prediction task,HR@K and NDCG@K increased by 0.055 and 0.039,respectively.On our self built dataset,the experimental results of the three tasks are also higher than the existing models.This indicates that MRNTM performs well in three tasks: topic learning,citation prediction,and author collaboration prediction.