Paper Recommendation via Correlation Pattern Mining and Attention Mechanism

In this paper, we improve the efficiency and effectiveness of the matrix factorization method in the paper recommendation system. We mainly address two problems. First, the vectors based on citation networks are undertrained because newly added papers are rarely cited. Second, current algorithms are mainly based on keyword search or global popularity and lack the organic combination of considering personalized interest and global popularity. To address the above two issues for the paper recommender, we propose a matrix factorization model that combines popularity analysis and attention mechanisms. The model effectively fuses the similarity of the citation network and topic using the multiplicative law, which can alleviate the data sparsity problem. Especially for cold-start papers, we add second-order neighbor nodes to makeup for the problem that newly joined papers don’t get enough training. We propose a keyword attention mechanism that combines user preferences and global popularity to personalize and balance the popularity of papers. Through comprehensive experiments on the CiteULike dataset, we show that our method can significantly improve the paper recommendation effectiveness.