The Knowledge Graph-based Github Repositories Recommendation Research And Implementation

Based on the spirit of open source and sharing of software development,open-source software code hosting and software development knowledge Q&A platforms represented by GitHub have rapidly developed and been favored by many developers.These platforms have accumulated a large amount of high-quality software development domain-related code and knowledge,but this knowledge is too scattered and difficult to solidify,and the efficiency of knowledge sharing and dissemination is low,which affects developers’ utilization of knowledge.In order to solve the information overload problem of open source knowledge,this thesis investigates GitHub project repository recommendation based on knowledge graphs,personalizes the recommendation of project repositories based on user preferences,and also improves the recommendation effect by using hierarchical knowledge in knowledge graphs to realize active dissemination of open source software knowledge.However,the following problems still exist in the research of GitHub project repository recommendation based on the knowledge graph:(1)There are a large number of guide nodes in the GitHub knowledge graph,and there are rich correlations between this node and other entities,however,the existing methods do not take into account the positive effect of the guide nodes on the recommendation effect,resulting in the process of capturing user preferences being easily influenced by other entities and this results in low-quality and biased personalized representations of users.(2)The traditional single-structure network learning model based on the knowledge graph does not fully utilize the time-dependence and fallibility of user-item interaction,and cannot effectively fuse the temporal and structural information hidden in the entities,while not considering the interaction mechanism between temporal and structural information of the user-item dichotomous network as a time-varying system,which reduces the recommendation accuracy.Therefore,to solve the problems,for the knowledge graph-based recommendation of the GitHub project repository,this thesis conducts the following research.1.A Guided Node Graph Convolutional Neural Network(GNGCN)model based on guided node graphs is proposed.This part of the study explores the active role of guided nodes in the knowledge graph recommendation task,and the model effectively captures the connections between entities by mining the influence of related nodes.The neighborhood samples of each entity in the knowledge graph are extracted as their guide nodes,and then the information and bias of the guide nodes are combined in computing the representation of an entity and extended to multiple hops to achieve convolution and aggregation of the model.2.The Dynamic Attribute Enhancement of Graph Convolutional Neural Network(DAEGCN)model is proposed based on the bootstrap node graph convolutional neural network.In this thesis,the hybrid recommendation model DAEGCN proposed in this section describes the identification of temporal patterns as a prediction problem of item popularity growth,which is solved by using time series analysis and combines it with the knowledge graph-based graph convolutional neural network to fully mine and fuse the structural and temporal information.3.In this thesis,we design and implement a GitHub recommendation platform based on the constructed recommendation model based on a knowledge graph.The platform can automatically crawl GitHub entity knowledge and provide datasets for model experiments.The platform makes recommendations based on the results of user interaction behavior modeling calculations and returns recommendation results.The platform provides users with a knowledge retrieval function and returns knowledge based on user input.In this thesis,the GNGCN model and DAEGCN model are evaluated experimentally based on the GitHub-SKG dataset and the GitHub-SKGT dataset,respectively.The experimental results show that the models proposed in this thesis have different degrees of enhancement and improvement in recommendation performance and model size compared with other baseline models.