Replication Package for ESEC/FSE 2023 Paper "How Early Participation Determines Long-Term Sustained Activity in GitHub Projects?"

This replication package can be used for replicating results in the paper. It contains 1) a dataset of 290,255 repositories; and 2) Python scripts for training and interpreting models. The GitHub repository of the paper is available at https://github.com/mcxwx123/Sustainable_projects. We recommend manually setup the required environment in a commodity Linux machine with at least 1 CPU Core, 8GB Memory and 100GB empty storage space. We conduct development and execute all our experiments on a Ubuntu 20.04 server with two Intel Xeon Gold CPUs, 320GB memory, and 36TB RAID 5 Storage. We use GHTorrent to restore historical states of 290,255 repositories with more than 57 commits, 4 PRs, 1 issue, 1 fork and 2 stars. The raw data of repositories (collected in their first 1,3,5 months(s)) are stored in `Replication Package/data/prodata_1.pkl`, `Replication Package/data/prodata_3.pkl`, and `Replication Package/data/prodata_5.pkl`. The contribution of features resulting from LIME model is stored in `Replication Package/data/limeres_m3_t2_k1.pkl`.<br> `Replication Package/data/X_test_m3_t2_k1.pkl` and `Replication Package/data/y_test_m3_t2_k1.pkl` store the test dataset for the LIME model. You can run `Replication Package/fitdata.py` to get the results in Table 3 and 4, run `Replication Package/draw_compare_variable.py` to get Figure 2 and run `Replication Package/allvari_statistics.py` to get Table 5. In `Replication Package/Variable_comparison_with_different_parameter.pdf`, we show the LIME results under different parameters. In `Replication Package/sample_pros.csv`, we also provide the list of randomly selected repositories in Section 3.1.<br> The explanations for collecting the variables, the examples of variable effects on project sustainability, and the hyperparameter setting of the machine learning models are provided in the README.md file.本复现套件可用于复现该论文的研究成果，内含两部分内容：1）包含290255个代码仓库的数据集；2）用于模型训练与可解释性分析的Python脚本。该论文的GitHub仓库地址为https://github.com/mcxwx123/Sustainable_projects。 我们建议在通用Linux主机上手动配置所需运行环境，该主机需至少配备1个CPU核心、8GB内存及100GB可用存储空间。本次开发与全部实验均在搭载两颗Intel Xeon Gold CPU、320GB内存与36TB RAID 5存储的Ubuntu 20.04服务器上完成。 我们使用GHTorrent恢复了290255个代码仓库的历史状态，这些仓库均满足提交次数≥57次、拉取请求（Pull Request，PR）≥4个、议题（Issue）≥1个、复刻（Fork）≥1次与星标（Star）≥2个的筛选条件。各仓库的原始数据（采集自其创建后的第1、3、5个月阶段）分别存储于`Replication Package/data/prodata_1.pkl`、`Replication Package/data/prodata_3.pkl`与`Replication Package/data/prodata_5.pkl`文件中。 LIME（局部可解释模型无关解释，Local Interpretable Model-agnostic Explanations）模型生成的特征贡献度结果存储于`Replication Package/data/limeres_m3_t2_k1.pkl`。`Replication Package/data/X_test_m3_t2_k1.pkl`与`Replication Package/data/y_test_m3_t2_k1.pkl`则存储了用于该LIME模型的测试数据集。 执行`Replication Package/fitdata.py`可复现论文中表3与表4的结果；运行`Replication Package/draw_compare_variable.py`可生成图2；执行`Replication Package/allvari_statistics.py`可得到表5的内容。在`Replication Package/Variable_comparison_with_different_parameter.pdf`中，我们展示了不同参数设置下的LIME实验结果。此外，`Replication Package/sample_pros.csv`提供了3.1节中随机选取的代码仓库列表。 README.md文件中包含了变量采集说明、变量对项目可持续性影响的示例，以及机器学习模型的超参数配置细节。