Big Data? Big Issues: Degradation in Longitudinal Data and Implications for Social Sciences
1. Matthew S. Weber
Hai Nguyen
Rutgers University
WebSci 2015
Oxford, UK
BIG DATA,
BIG ISSUES
2.
3. 3
Dataset Research Potential Dates Captures Unique URLs
Hurricane Katrina Online networks and organizational
resilience (Chewning, Lai and Doerfel,
2012; Perry, Taylor and Doerfel, 2003) in
the wake of disasters; information
dissemination
2003 – 2012 1,694,236 663,740
Superstorm
Sandy
2003 – 2012 41,703,112 20,013,455
US Senate Study the growth of political activity in
online environments (Adamic & Glance,
2005; Bruns, 2007; Chang & Park, 2012);
polarization & media discourse
109th – 112th
Congresses
26,965,770 8,674,397
US House 51,840,777 12,410,014
Occupy Wall
Street
Previous research on NGOs in the online
environment (Bach & Stark, 2004;
Shumate, 2003, 2012; Shumate, Fulk, &
Monge, 2005); use of hyperlink data to
study the formation and role of alliances
between SMOs
2010 – 2012 247,928,272 11,3259,655
US Media
Previous studies of news media
organizations (Greer & Mensing, 2006;
Weber, 2012; Weber & Monge, In
Press); focus on evolutionary patterns
2008 – 2012 1,315,132,555 539,184,823
17. • Scale out across multiple datasets:
– US House – 2005:2013:
– US Senate – 2005:2013
– Hurrican Katrina – 2003:2012:
– Occupy Wall Street – 2010:2012
17
18. 0 5 10 15 20 25 30
050000010000001500000200000025000003000000
Potential vs. Actual URLs
CountofPages
18t
CountofURLs
Potential
Actual
Difference
19. 19
0e+002e+064e+066e+06
Changes in Crawl Completeness
CountofPages
t
CountofURLs
OWS
House
Senate
Katrina
existing
potential
b =
set a unit of time for analysis, c
choosing n perios across a total time T
20. In the ideal case, it would be possible to create a factor that corrects
for data degrade:
bt
How does this help?
Each of the illustrated cases fits against an
exponential function ~ b
• Senate: 0.13
• House: 0.13
• Katrina: 0.02
• OWS: 0.10
20
ebt
24. Lessons Learned
• Degradation is a factor in working with available large-scale data
– In part, degradation is related to the provenance of the data
– In turn, there is a need to record the origins of datasets (provenance)
• Patterns of degradation prove problematic for statistical analyses
– Ex: network analysis with snowball samples vs. whole network
• Continued work needed to develop research guidelines as more
scholars engage with this data
24
25. Get in contact with us:
– matthew.weber@rutgers.edu
– @mediareinvented
The Team
– Kris Carpenter, Vinay Goel, Internet Archive
– David Lazer, Katherine Ognyanova, Northeastern University
– Allie Kosterich, Hai Nguyen, Rutgers University
Research supported by NSF Award #1244727 and the NetSCI Lab @ Rutgers
Editor's Notes
There are many types of large-scale data… only talking about Internet based data… focusing on datasets that are re-used.
- Markus - “social scientists are used to fine-grain, well-controlled data, and that doesn’t exist on the web”
20th Century Collection = 9TB of metadata
Media Seed List = 4,891
For instance, researchers have proposed focusing archival efforts on capturing data that changes the most frequently, in order to capture the majority of new content [36]. Elsewhere, researchers have suggested that crawling strategies should prioritize archival efforts based on the size and relative position of websites within their larger ecosystems [37].
Driscoll and Walker (2014) For instance, a comparison of Twitter data collected via a public API and data collected from a “fire hose” provided by GNIP PowerTrack, found significant differences between the two datasets. In most cases the PowerTrack data proved to be more powerful,
3 month windows of time…
Also looked at the size of the webpages, and estimating out size… wasn’t as reliable.