Knowledge Spillovers in the Open Source Community ... › sites › default › files › TSE › documents › Chaire… · Knowledge Spillovers in the Open Source Community Evidence

Knowledge Spillovers in the Open Source CommunityEvidence in Github

Tong Wang

1Business SchoolUniversity of Edinbugh

Toulouse Digital Seminar, 2017

Tong Wang (Universities of Edinburgh) Knowledge Spillovers in the Open Source CommunityToulouse Digital Seminar, 2017 1 / 21

MotivationKnowledge Transfer

Knowledge transfer is prevalence in academia and programmingcommunity

Product development in community-based organisations is becomingincreasingly important, e.g. quite a few commercial softwares havetheir open-source versions.

There are basically two ways of learning:

Interaction with individuals (e.g. email exchange, brain storm, etc.)Read and study other good projects (e.g. read academic papers, etc.)

Learning (knowledge spillover) could also be direct or indirect.




















Interaction with individuals (e.g. email exchange, brain storm, etc.)

Read and study other good projects (e.g. read academic papers, etc.)

















Motivation(Cont’d)

Factors that affect the effectiveness and efficiency of learningprocedure remain largely unexplored.

In a UGC(User generating Content)-like open-source community, dothe ways of learning change?


Motivation(Cont’d)

Factors that affect the effectiveness and efficiency of learningprocedure remain largely unexplored.

In a UGC(User generating Content)-like open-source community, dothe ways of learning change?


Literature Review

Grewal, Lilien, and Mallapragada (2006) investigates how the networkembeddedness of projects and project managers influences the successof projects;

Goyal, van der Leij, and Moraga-Gonzalez (2006) studies the networkproperty of coauthorship network in economics;

Manski (2000), Sacerdote (2001), and Angrist and Lang (2004) studyneighborhood effects and spillovers in many other aspects such aslabor and education.

Fershtman and Gandal (2011) studies knowledge spillover in anenvironment where the knowledge producers and knowledgeconsumers are different.

Neil Gandal and Uriel Stettner (2016) evaluates the importance ofprogram modification and fuction additional to project success.


Literature Review







Literature Review







Literature Review







Literature Review







Contribution

Empirically examine the association between projectsuccess/popularity and network measures in a situation whereknowledge producers and consumers are the same group

Identify the importance of Activity and Effort in the learning process

Throw some light on social learning process.

One of the First attempts to use big data approach to analyzeknowledge spillover.


Contribution






Contribution






Contribution






Data Source

Github offers a unique opportunity to examine our researchquestions

Github is a social coding platform

Github is the world largest code hosting and open-source development

platform

Every action on Github is recorded and thus is obtainable.

“fork” is a good measure for popularity.

More importantly, the network structure of Github is quite clear


Data Source




platform





Data Source




platform





Data Source




platform





Data Source




platform





A sample page of a project on Github


Data Sources

We obtained our data by SQL Querying Github Archive and networkspidering.

All projects and contributor information before Jun. 2016 wasretrieved and stored in the spider server, the raw data is 300G +.

The contributors and projects are identified by unique IDs.

Project is linked to at least one contributor. Contributors areconnected by their joint projects, so finally we get a two-modenetwork.


Data Sources






Data Sources






Data Sources






Network Illustration


Descriptive Statistics: Observations

Most of the Github Projects only have 1 contributor, most of thecontributors only have 0 or 1 project.

Tong Wang (Universities of Edinburgh) Knowledge Spillovers in the Open Source CommunityToulouse Digital Seminar, 2017 10 /

21

Descriptive Statistics: Observations(Cont’d)

Most of the projects raise little attention. Being popular is generally verydifficult in the Open-Source Community.


21

The Project Network

The nodes of this network areprojects.

There is a link between twodifferent project nodes if thereare contributors who participatein both projects.

Each link may have a valuewhich reflects the number ofcontributors who participate inboth projects.


21

The Project Network





21

The Project Network





21

Varibles

Popularity/Success Measurement: the number of forks

Activity: No commit during the last 18 months.

Node degree: measurement of the direct effect, the degree is literallythe number of projects with which has a direct link

Closeness Centrality as the measure of indirect effect, define closenesscentrality as C (i) = N−1

Σj∈Nd(i ,j)

Existing Period: the number of years that have elapsed since theproject first appeared

Number of Contributors: the number of contributors who participatedin the project

Popular Language: The most popular 5 languages used in TIOBEIndex

Number of Comments: A proxy variable for the effort of contributors


21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Varibles





Σj∈Nd(i ,j)






21

Illustration of Closeness Centrality


21

Preliminary Exploration: Neural Networks


21

Basic Model Specification

Assume the initial success level is α and the final success level is Si .

If direct spillover effect makes sense, then a project with a higherdegree is more likely to be successful. That is, in the regressionfunction Si = α + βDi ; βshould be significant.

If indirect spillover effect makes sense, then the final success levelshould be:

Si = α +∑j

γ

d(i , j)+ βDi

since it is natural to assume that longer distance implies lowerimpact. But since we know C (i) = N−1

Σj∈Nd(i ,j) , so the equation above

can be transformed to :

Si = α +γ

N − 1C (i) + βDi


21





Si = α +∑j

γ

d(i , j)+ βDi




Si = α +γ

N − 1C (i) + βDi


21





Si = α +∑j

γ

d(i , j)+ βDi




Si = α +γ

N − 1C (i) + βDi


21

Main results: Basic Model

Table: Effect of Activity

Dependent variable:

lfork

(1) (2)

lCPP 0.208∗∗∗ (0.002) 0.203∗∗∗ (0.002)lyear −0.929∗∗∗ (0.005) −0.920∗∗∗ (0.005)activity 0.339∗∗∗ (0.003) 0.146∗∗∗ (0.006)ldegree 0.082∗∗∗ (0.001) 0.036∗∗∗ (0.001)lcloseness −0.106∗∗∗ (0.020) −0.057∗ (0.031)lNO Comments 0.312∗∗∗ (0.002) 0.303∗∗∗ (0.002)popular 0.019∗∗∗ (0.002) 0.019∗∗∗ (0.002)activity:ldegree 0.074∗∗∗ (0.002)activity:lcloseness −0.013 (0.041)Constant 2.687∗∗∗ (0.009) 2.783∗∗∗ (0.010)

Observations 691,582 691,582R2 0.251 0.254Adjusted R2 0.251 0.254Residual Std. Error 0.914 (df = 691574) 0.912 (df = 691572)F Statistic 33,077.240∗∗∗ (df = 7; 691574) 26,125.280∗∗∗ (df = 9; 691572)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01


21

Main results(Cont’d)

Table: Effects of Effort

Dependent variable:

lfork

(1) (2)

lCPP 0.208∗∗∗ (0.002) 0.197∗∗∗ (0.002)lyear −0.929∗∗∗ (0.005) −0.917∗∗∗ (0.005)activity 0.339∗∗∗ (0.003) 0.342∗∗∗ (0.003)ldegree 0.082∗∗∗ (0.001) 0.066∗∗∗ (0.001)lcloseness −0.106∗∗∗ (0.020) −0.056∗∗∗ (0.021)lNO Comments 0.312∗∗∗ (0.002) −0.048∗∗∗ (0.008)popular 0.019∗∗∗ (0.002) 0.019∗∗∗ (0.002)ldegree:lNO Comments 0.047∗∗∗ (0.001)lNO Comments:lcloseness 0.782∗∗∗ (0.047)Constant 2.687∗∗∗ (0.009) 2.709∗∗∗ (0.009)

Observations 691,582 691,582R2 0.251 0.256Adjusted R2 0.251 0.256Residual Std. Error 0.914 (df = 691574) 0.910 (df = 691572)F Statistic 33,077.240∗∗∗ (df = 7; 691574) 26,508.230∗∗∗ (df = 9; 691572)

Note: ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01


21

Discussion

A longer exposure helps.

More contributors, more forks.

Direct project spillover effect is significantly positive.

Knowledge spillover from inactive projects to active projects.

Manager’s effort is crucial to activate indirect spillovers.


21

Discussion







21

Discussion







21

Discussion







21

Discussion







21

Concluding Remarks

the knowledge distribution pattern on Social Coding platform is quitedifferent with that on traditional open source community.

indirect spillovers are generally weaker.

repo manager’s effort can make a lot of difference.


21

Concluding Remarks





21

Concluding Remarks





21

Future works

compare spillover effects of strong connections and weak connections

Explore the effect of bot-contributor

adding more control variables, such as project size and projectlanguage

Try other measurement of success


21

Future works






21

Future works






21

Future works






21

Documents

Knowledge Spillovers in the Open Source Community ... › sites › default › files › TSE › documents › Chaire… · Knowledge Spillovers in the Open Source Community Evidence