【作者】 赵志；

【导师】 王玉明；

【作者基本信息】 华中科技大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 随着国内科研领域信息化建设的飞速发展,科研管理平台积累了海量数据。基于海量数据构建的科研领域知识图谱,也为科研管理数据的深度应用奠定了基础。一方面,随着科研数据的日益增长,科研人员面临日益严重的信息过载问题,需要更高效的工具帮助科研人员过滤信息;另一方面,科研人员希望能获取个性化科研信息,并与具有相同研究兴趣的人员交流合作。科研项目作为科研人员在所属科研领域研究内容的集中体现,具有很高的参考价值。本文构建了一个科研项目推荐系统,以此帮助科研人员高效地发现符合其兴趣的科研项目、了解领域内相关人员的研究方向,从而促进各项目成员间交流合作。本文广泛调研了基于知识图谱的推荐模型,选取了基于知识图谱和图神经网络的KGIN(Knowledge Graph-based Intent Network)模型作为基础模型,该模型通过用户意图建模与关系路径聚合获得了优秀的推荐性能和可解释性。基于KGIN模型,本文提出了改进模型WKIU(Weighted KGIN with user-interest-boundary loss)。WKIU模型通过注意力机制组合不同深度聚合层的嵌入生成最终表示,有效地区分了高低阶信息的重要性。此外,模型在损失函数中引入个性化兴趣边界,缓解了训练中样本梯度消失的问题,帮助模型更高效地区分正负样本。经过在科研数据集上对比实验证明,WKIU的推荐性能和训练效率均优于KGIN,具有可观的研究及应用价值。另外本文通过消融实验确定了模型最优的独立性意图建模方案、意图划分粒度以及关系路径聚合层数。基于WKIU模型,本文从功能和技术层面出发,设计并实现了科研项目推荐系统。系统构建了数据预处理、推荐评分以及推荐展示三大模块。特别的,推荐展示模块基于推荐评分结果、意图关系分布以及知识图谱关联关系,在为科研人员提供个性化项目推荐的同时也提供了可视化推荐理由。更多还原

【Abstract】 With the rapid development of information construction in the domestic research field,massive data has been accumulated on the research management platform.The scientific research knowledge graph built on the massive data has laid the foundation for the profound application of research management data.On the one hand,with the increasing growth of research data,researchers face a serious problem of information overload and need more efficient tools to help them filter information.On the other hand,researchers want to access personalized research information and cooperate with people with the same research interests.Research projects are highly valuable as a centralized representation of the research content of researchers in their research fields.In this paper,we develop a project recommendation system to help researchers efficiently discover research projects that match their interests and learn about the research directions of researchers in their fields,thereby promoting exchange and collaboration among members of different projects.In this paper,the recommendation model based on knowledge graph is extensively investigated,and the KGIN(Knowledge Graph-based Intent Network)model based on knowledge graph and neural graph network is selected as the basic model.KGIN achieves excellent recommendation performance and interpretability by user intent modeling and relational path-aware aggregation.Based on KGIN,an improved model WKIU(Weighted KGIN with user-interest-boundary loss)is proposed in this paper.WKIU generates the final representation by combining embeddings of different depth aggregation layers through the attention mechanism,which effectively distinguishes the importance of high-and low-order information.Furthermore,the personalized interest boundary is introduced into the loss function,which alleviates the problem of sample gradient disappearance in training and helps the model distinguish positive and negative samples more efficiently.Through the comparative experiment on the research dataset,it is proved that the recommendation performance and training efficiency of WKIU are better than KGIN,which has considerable research and application value.In addition,the optimal independence intent modeling scheme,the granularity of intent division and the number of relational path aggregation layers of the model are determined by the ablation experiments.Based on the WKIU model,the scientific research project recommendation system is implemented in terms of function and technology.The system consists of three modules:data preprocessing,recommendation scoring and recommendation display.In particular,the recommendation display module provides researchers with personalized project recommendations along with visualized recommendation reasons based on recommendation scoring results,intent relationship distribution,and knowledge graph association relationships.更多还原