Predicting Popularity of Open Source Projects Using Recurrent Neural Networks

Abstract

GitHub is the largest open source software development platform with millions of repositories on variety of topics. The number of stars received by a repository is often considered as a measure of its popularity. Predicting the number of stars of a repository has been associated with the number of forks, commits, followers, documentation size, and programming language in the literature. We extend prior studies in terms of input features and algorithm: We define six features from GitHub events corresponding to the development activities, and additional six features incorporating the influence of users (followers and contributors) on the popularity of projects into their development activities. We propose a time-series based forecast model using Recurrent Neural Networks to predict the number of stars received in consecutive k days. We assess the performance of our proposed model with varying k (1, 7, 14, 30 days) and with varying input features. Our analysis on five topmost starred repositories in data visualization area shows that the error rate ranges between 19.76 and 70.57 among the projects. The best performing models use either features from development activities only, or all metrics including all the features.