[Objective]This paper proposed a method of scientific collaboration prediction combing the fusion network representation learning and author topic model.

[Methods]Based on classical network representation learning method, the embedding vector representation of authors was calculated, and the structural similarity of authors was calculated by cosine similarity. Based on the author-topic model, the topic representation of authors was obtained, and the topic similarity of authors was calculated by Hellinger distance. Then the two similarity measures were fused linearly, and the Bayesian optimization method was used to fuse the hyperparameter selection.

[Results] Empirical research based on the NIPS datasets shows that after Bayesian parameter selection, the node2vec+ATM model achieves the AUC value of 0.9271, which improved 0.1856 than the benchmark model.

[Limitations] This article only considers the content of the author’s publications, but does not incorporate more attribute such as the author’s institution and geographic location into the model.

[Conclusions] The proposed fusion model take structure and content features into consideration and can improve the prediction effect of network representation learning.