Dual-AutoEncoder & Bipartite Graph Embedding for article recommendation

Existing methods in article recommendation fail to fully use the article information, or pay less attention to the correlations among articles and "User-Article"s, resulting in inaccurate recommendation performance and weak personalized recommendation ability. Therefore, we design an article recommendation model based on Dual-AutoEncoder & Bipartite Graph (DAEBG): we build a Dual-AutoEncoder, consisting of an Attention-based AutoEncoder and a Graph AutoEncoder, for article basic features of "Title and Abstract" and the correlation features among articles by the "Tag and Citation" information, which are the input of DAEBG. Second, we design a correlation extraction network with Resource Allocation-based Bipartite Graph to extract the correlation feature embeddings of "User-Article"s. Finally, we employ the Probabilistic Matrix Factorization to fuse and update these feature embeddings, consider all kinds of features comprehensively, and avoid the sparse matrix and cold start problems. The output of DAEBG is a rating matrix, which is used for article recommendation. Experimental results show that DAEBG outperforms the other recommended methods in all experiments with different data characteristics (sparse or dense), which could help people choose the targets that better match their interests from a huge amount of articles.