Crowd Crawling: Towards Collaborative Data Collection for Large-scale Online Social Networks Cong Ding, Yang Chen*, and Xiaoming Fu University of Göttingen.

Crowd Crawling: Towards Collaborative Data Collection for Large-scale Online Social Networks

Cong Ding, Yang Chen*, and Xiaoming Fu

University of Göttingen*Duke University

Significance of social network data crawling

•Understanding user behaviors

•Improving SNS architectures

•Handling privacy/security issues

•and so on...

Current data collection methods (1)

•ISP-based measurement [Schneider IMC’09]

Only ISP companiescan do that

Current data collection methods (2)

•Cooperate with SNS companies [Yang IMC’11]

Most research groupsdo not have chance

Current data collection methods (3)

•Crawl data by a single group (and share them to others)

[Gjoka INFOCOM’10]

Suffering requestrate limiting

Shortages of crawling by a single group

•Waste computing andnetwork resources

•Introduce overhead toservice providers (andmay lead stricter rate limiting)

•Lack of ground truth forthe research community

A new thought

Why not collect data collaboratively?

System overview

Coordinator

Crawlers

System design

•Fetching UIDs (BFS, etc.)

•Handling crawling failure (timeout)

•Bypassing request rate limiting (massive IP addresses)

•Data fidelity (redundant crawling)

Implementation

•A proof-of-concept prototype (without the data fidelity part)to crawl in Weibo

•472 PlanetLab servers as crawlers

Evaluation

•In 24 hours, we have crawled 2.22M users’ data from Weibo,including user profiles, all the posts, all the social connections

•Comparison:

•Fu et al. (PLOS ONE 2013) get 30K user’s data in 6 days•Guo et al. (PAM 2013) get 1M user’s data in 1 monthCrowd

CrawlingFu et al. Guo et al.

#UIDs/day 2.22M 5K 33K

Evaluation

Evaluation

Conclusion and Discussion

•Data sharing may violate some providers’ terms of servicesoTwitter does not allow to share data (even for

research)oWeibo allows to share data among researchers

•Unlimited data sharing might cause ethical issuesoThe data should be anonymized

•We will publish the data crawled in the evaluation