Open Access

ARTICLE

Graph Convolutional Network-Based Repository Recommendation System

Zhifang Liao¹, Shuyuan Cao¹, Bin Li¹, Shengzong Liu^2,*, Yan Zhang³, Song Yu^1,*

1 School of Computer Science and Engineering, Central South University, Changsha, China
2 School of Information Technology and Management, Hunan University of Finance and Economics, Changsha, China
3 School of Engineering and Built Environment, Glasgow Caledonian University, Glasgow, UK

* Corresponding Authors: Shengzong Liu. Email: ; Song Yu. Email:

Computer Modeling in Engineering & Sciences 2023, 137(1), 175-196. https://doi.org/10.32604/cmes.2023.027287

Received 23 October 2022; Accepted 16 December 2022; Issue published 23 April 2023

Abstract

GitHub repository recommendation is a research hotspot in the field of open-source software. The current problems with the repository recommendation system are the insufficient utilization of open-source community information and the fact that the scoring metrics used to calculate the matching degree between developers and repositories are developed manually and rely too much on human experience, leading to poor recommendation results. To address these problems, we design a questionnaire to investigate which repository information developers focus on and propose a graph convolutional network-based repository recommendation system (GCNRec). First, to solve insufficient information utilization in open-source communities, we construct a Developer-Repository network using four types of behavioral data that best reflect developers’ programming preferences and extract features of developers and repositories from the repository content that developers focus on. Then, we design a repository recommendation model based on a multi-layer graph convolutional network to avoid the manual formulation of scoring metrics. This model takes the Developer-Repository network, developer features and repository features as inputs, and recommends the top-k repositories that developers are most likely to be interested in by learning their preferences. We have verified the proposed GCNRec on the dataset, and by comparing it with other open-source repository recommendation methods, GCNRec achieves higher precision and hit rate.

---

Keywords

---