Design And Implementation Of Research Paper Recommender System Based On Multi-Features

Finding similar papers based on a target paper is a common requirement of researchers.The academic paper recommender system can help researchers filter and extract effective information from the rapidly increasing academic big data.Recommendation algorithm is one of the main research objects in recommender system.Different recommendation al-gorithms are suitable for different data.Academic papers contain multi-attribute features.We are able to use recommendation algorithms based on text features,we are also able to build a homogeneous network based on citations,co-citations or build a heterogeneous network based on information such as articles,authors,and institutions.Then graph-based features can be used for recommendation algorithms based on network.Existing paper recommendation methods have problems.Classic paper recommen-dation methods based on homogeneous network lose the structure and semantic informa-tion of the complex paper network composed of multiple vertex classes and edge classes,while most graph-based recommendation algorithms ignore the paper text information.Regardless of whether the graph embedding method is homogeneous or heterogeneous,it needs to be retrained after adding new nodes.There are also some methods for recon-structing the paper network based on the results of multi-attribute graph embedding.The text or structural features of their fusion are relatively simple,and they do not consider the text similarity of the abstract information and the structural relationship brought by the author and the institution of the paper.In response to the above problems,this the-sis proposes a multi-attribute feature recommendation algorithm for papers that integrates heterogeneous network representation learning and text representation learning methods on the basis of homogeneous networks:This recommendation algorithm first uses the citation network to convert the citation-based structural features of the paper.Then uses a heterogeneous network containing the article,author,and institution information to convert the paper to feature vector.After then converts the paper into a vector through title and abstract by doc2 vec.For each paper,re-construct the citation network using similar papers based on citation structure features,heterogeneous network features,and text features.Finally according to the reconstructed citation network with multi-attribute features,the graph embedding that retains the struc-tural and homogeneity features is obtained.The multi-attribute features of each paper will finally generate a recommendation list through the vector similarity of the target paper.We experiment our algorithms on the ar Xiv data subset,the result shows that our algorithm is partly better than the comparison algorithms in four metrics of precision,accuracy,recall,and F1.Combined with the above recommendation algorithm,this thesis designs and im-plements a paper recommendation prototype system,including basic query and search functions.It also supports user-defined paper recommendation methods based on multi-attribute features.For newly added papers,it avoids re-training of graph representation learning and directly performs recommend.This system uses Python for back-end devel-opment,Qt and flask frameworks for front-end development,and My SQL as the database implementation system.