Collaborators Recommendation Based On Heterogeneous Network Embedding

In academia,researchers collaborate with their peers to improve the quality of their research and expand their research fields.However,the information overload resulting from the production of large-scale academic data poses challenges for identifying potential collaborators.To address this problem,several collaborator recommendation models have been proposed by researchers.However,these models primarily rely on collaboration relationships between scholars and ignore some important content information factors,resulting in dependency on the recommendation results of collaboration.Furthermore,many models typically choose first-order proximity as information supplementation when utilizing structural information,but the number of first-order proximity relationships in networks is limited,which cannot effectively express relationships between nodes.To address these issues,this thesis introduces content information,including research topics,affiliations,academic levels,and publication venues,as the basis for model training,and utilizes second-order proximity,which can express local structural information,combined with biased random walk algorithm,to propose two collaborator recommendation models based on heterogeneous network embedding.The main contributions of this thesis are as follows:(1)We propose a collaborator recommendation model based on a heterogeneous information network.Firstly,nodes,edges and weights are defined,and the corresponding calculation methods for weight of different information graphs are designed to construct a logically heterogeneous information network.Secondly,second-order proximity,which can more comprehensively express structural information,is considered when learning network node embeddings.Finally,the content and structural information of multiple information graphs are encoded into a latent space to jointly learn the embeddings of scholars,and various optimization algorithms,including negative sampling optimization and margin sampling method,are used to improve the model performance.(2)We propose a scientific collaborator recommendation model based on multiinformation graph random walk.Firstly,the nodes that store embedding information are reorganized into multiple author Homogeneous graphs based on different information.Then,the similarity between nodes is calculated using embedding information as transition probabilities,and biased random walk is performed to calculate candidate scores based on individual information.Finally,the model assigns different weights to candidate scores calculated based on different information to calculate the final candidate ranking.(3)We verify the proposed two models on the Aminer,DBLP,and ACM datasets and compare them with multiple baseline methods.The experimental results show that the collaborator recommendation model based on the heterogeneous information network achieved 4.8%,2.8%,and 4.3% improvements in precision,recall,and F1 index,respectively,compared with the baseline methods.The scientific collaborator recommendation model based on multi-information graph random walk achieved 2.9%,3.1%,and 2.8% improvements compared with the baseline methods,and achieved better recommendation results than the collaborator recommendation model based on the heterogeneous information network.In addition,to analyze the influence of model parameters on experimental results,this thesis conducts experiments and analyses on the values of multiple parameters.