Research On Third-party Library Recommendation Algorithm Based On Open-source Platform

Life is full of all kinds of software.The continuous growth of the software market has brought great convenience to users,but has also brought more challenges to developers.In order to better serve users,the competition among software service providers is also increasingly fierce.Developers need to develop software products that better meet user needs in the shortest possible time,and at the same time ensure product stability,so as to satisfy users.In the software development process,third-party software libraries have become an indispensable part.Developers can reuse their functions through the application programming interfaces they provide,thereby completing application development efficiently.However,due to the ever-increasing number of reusable third-party libraries,finding a suitable third-party library is a tedious and complex task for developers.Therefore,it is of great significance to help developers quickly find available thirdparty libraries.Aiming at this problem,this thesis firstly proposes a new interactive informationbased third-party library recommendation algorithm SELib Rec.The network structure is reconstructed by using the graph embedding model to capture the high-order nonlinear structural features of project nodes.After obtaining the representation of projects,the similarity measurement algorithm is used to find the neighboring projects,and then the third-party libraries that the neighboring projects depend on are recommended for the target project.Experiments show that the algorithm has better performance than baselines,but it cannot achieve better recommendation results in the user cold-start scenario.Aiming at the user cold-start problem,this thesis proposes to use graph neural network and collaborative filtering to recommend third-party libraries.First,textual information of projects as well as third-party libraries is extracted using a pre-trained language model,which is then fed into a graph neural network as initial features.The high-order neighbor features of nodes are aggregated in a graph neural network model,and the thesis optimizes the network model using a Bayesian personalized ranking loss.After obtaining the embedded representation of projects,the similarity algorithm is used to find the neighbor projects,and then the third-party libraries used by the neighbor projects are recommended for the target project.This module is called CELib Rec.In the aggregation module,the recommendation results of the above two modules are weighted and summed,and then sorted according to their scores to obtain the final recommendation list.In this thesis,the fusion algorithm is named GELib Rec.This thesis uses the interaction information and text description information of projects and third-party libraries to achieve hybrid recommendation,and extracts two datasets from the App Brain official website and the open-source platform Libraries.io respectively.Experiments are carried out in the common recommendation scenarios and user cold-start recommendation scenarios.By comparing with several advanced third-party library recommendation algorithms on multiple evaluation metrics,the effectiveness of the proposed method is verified.