[Paper Review] Collaborative similarity analysis of multilayer developer-project bipartite network

Dengcheng Yan, Bing-Hong Wang|arXiv (Cornell University)|Mar 9, 2017

Complex Network Analysis Techniques15 references3 citations

TL;DR

This paper proposes a multilayer bipartite network model to analyze collaborative relationships between developers and projects on GitHub using watching and forking interactions. It applies collaborative similarity to measure interest diversity, revealing that developers' collaborative similarity decreases with degree in the watching layer (indicating broader interests) but increases in the forking layer (indicating focused, specialized engagement), while projects show no clear correlation in the forking layer.

ABSTRACT

To understand the multiple relations between developers and projects on GitHub as a whole, we model them as a multilayer bipartite network and analyze the degree distributions, the nearest neighbors' degree distributions and their correlations with degree, and the collaborative similarity distributions and their correlations with degree. Our results show that all degree distributions have a power-law form, especially, the degree distribution of projects in watching layer has double power-law form. Negative correlations between nearest neighbors' degree and degree for both developers and projects are observed in both layers, exhibiting a disassortative mixing pattern. The collaborative similarity of both developers and projects negatively correlates with degree in watching layer, while a positive correlations is observed for developers in forking layer and no obvious correlation is observed for projects in forking layer.

Motivation & Objective

• To understand the multifaceted relationships between developers and projects on GitHub beyond single-layer models.
• To model watching and forking interactions as a multilayer bipartite network for a holistic view of developer-project collaboration.
• To investigate the diversity of developers' interests through collaborative similarity, a measure of clustering behavior in bipartite networks.
• To examine how degree and network structure correlate with collaborative similarity across layers.
• To reveal behavioral differences between watching (information tracking) and forking (development engagement) on GitHub.

Proposed method

• Construct a multilayer bipartite network with two layers: watching (notification-based) and forking (code-branching) relations between developers and projects.
• Use real GitHub data from the GHTorrent Project (January 2015), filtered to PHP projects, to extract edge lists for both layers.
• Apply power-law fitting to degree distributions using the powerlaw Python package with log-spaced bins to preserve tail accuracy.
• Analyze nearest-neighbor degree correlations to detect disassortative mixing patterns in both layers.
• Define degree-dependent collaborative similarity as the average Jaccard similarity between projects selected by developers of the same degree, using log-binning to visualize trends.
• Compare collaborative similarity distributions across degree ranges, distinguishing between watching and forking layers.

Experimental results

Research questions

• RQ1 How do the degree distributions of developers and projects differ across the watching and forking layers?
• RQ2 What is the nature of degree correlations between nodes and their nearest neighbors in both layers?
• RQ3 How does collaborative similarity vary with node degree in the watching and forking layers?
• RQ4 Why do watching and forking behaviors lead to different patterns in collaborative similarity?
• RQ5 What insights do these patterns provide about the diversity of developers’ interests?

Key findings

• All degree distributions follow a power-law form, with projects in the watching layer exhibiting a double power-law distribution.
• Negative correlations between node degree and nearest neighbor degree are observed in both layers, indicating disassortative mixing for both developers and projects.
• Collaborative similarity of developers in the watching layer is negatively correlated with degree, suggesting broader interest diversity among highly active watchers.
• In contrast, collaborative similarity of developers in the forking layer is positively correlated with degree, indicating increasingly focused, specialized interests among frequent forkers.
• Projects in the forking layer show no significant correlation between degree and collaborative similarity, with values fluctuating around the global average.
• The results are explained by the functional differences: watching promotes exposure to diverse projects, while forking reflects sustained, field-specific engagement.