Research On Modeling And Representation Of Heterogeneous Hypergraphs For Academic Recommendations

With the rapid development of information technology,more and more academic papers are being published on the Internet and researchers can find papers in various fields.However,as the number of academic papers increases,it has become a challenge for researchers to quickly find papers that match their research interests.In recent years,academic recommendation methods based on heterogeneous graphs have been proven to improve recommendation performance.Heterogeneous graphs are not only able to fuse more information about papers and contain richer semantics than homogeneous graphs,and they are also able to satisfy users’ complex preferences.However,the method only focuses on the impact of other types of nodes on paper,ignoring the correlation of other nodes.Furthermore,due to the intricate and complex correlations between different types of nodes,heterogeneous graphs have limited effectiveness in capturing high-order complex relationships between nodes and integrating features of multiple types of nodes.Also,existing studies ignore the presence of nodes and edges with no or negative influence in the construction of academic graphs,which can affect the efficiency of information dissemination and introduce noise.Based on these issues,the main work of this paper is as follows:(1)This paper adopts a hypergraph approach to construct academic heterogeneous networks.Unlike traditional homogeneous graphs,hypergraphs can synthesize the metadata information of academic papers.This paper defines various types of hypergraphs based on the interplay between metadata,such as hypergraphs that synthesize papers and their metadata,hypergraphs based on authors writing multiple papers,and so on.In this way,an academic heterogeneous hypergraph containing multiple types of nodes is constructed in this paper to capture the higherorder complex relationships between nodes and to fuse multiple types of node features.(2)We propose a hypergraph neural network with an attention mechanism for different hyperedge types.Traditional hypergraph neural networks usually use homogeneous hypergraphs and do not distinguish the importance of different hyperedge types,but different types of hyperedges contain different semantic information.Therefore,this paper introduces an attention mechanism for specific types of hyperedges,which adaptively aggregates the embedding representations of the same hyperedges by different nodes in heterogeneous hypergraphs.In this way,to better capture users’ personalized preferences and improve the performance of heterogeneous hypergraphs for the academic recommendation.(3)This paper introduces a self-supervised learning approach to enhance the heterogeneous hypergraph.To address the situation that the heterogeneous hypergraphs constructed based on reference relations are noisy,this paper adopts a Dropout strategy to generate two new heterogeneous hypergraphs by randomly dropping nodes and hyperedges,to effectively alleviate the problems of data redundancy and model overfitting and improve the generalization performance of the model.(4)Extensive experiments are conducted on three commonly used datasets for academic paper recommendation and academic journal recommendation tasks,respectively.The effectiveness and scalability of the proposed approach for the academic recommendation task are verified by comparing it with a variety of classical models.