Thread by @informor

I used my 30 mins to make a pitch for thinking (and shifting) the unit of analysis or unit of observation used in our research, with a quick survey of common UoAs and sample papers. A thread: #icwsm2020

https://twitter.com/icwsm/status/1270003351393009664

When talking about online safety/abuse/misinformation one can image about content- and people-driven units of investigation; and then their mix. In each, we can start from the smaller units and work our way up. 2/

1) *Sub-resource* units: e.g. claims, statements, images. These are often not uniquely addressable, may appear in multiple locations. Sample paper:

ranger.uta.edu/~cli/pubs/2014… 3/

2) *Resources* like YouTube Video, News Article, Social Media Post, Website Comment. These are addressable; possibly compound/multi-faceted; w/ metadata; often w/ an identified author. Sample paper:

dl.acm.org/doi/abs/10.114…

4/

3) *Cascades and conversations*: networks and diffusion trees of content items, e.g. Tweet and retweets; Posts and re-shares; Image and its derivatives/manipulations; Comments and responses to it. Sample (famous) paper:

science.sciencemag.org/content/359/63…

5/

4) *Narratives*: general storylines that are broader than claims but more specific than topics, e.g. “Antifa is planning riots in your suburb” or “Disinfectants can cure Covid-19”. One of my favorite papers (and title formatting!) is from @tanmit

people.cs.vt.edu/tmitra/public/…

6/

5) *Topics* has been a popular unit of analysis, incl. both temporal and consistent topics, e.g. Election 2016, Black Lives Matter, Notre Dame fire, Covid-19. Hashtags, Trending topics, events are important subclasses. Example paper from @katestarbird:

dl.acm.org/doi/abs/10.114…

Moving to people-driven units of analysis/observation:

1) *Users/authors/Content posters/creators*. Can be formally authenticated (e.g. Twitter user) or identified (media journalist), or informal (contributor on blog). In the context of abuse, lots of work on trolls/bots

8/

Early summary of bot-related work here:
cacm.acm.org/magazines/2016…

And a recent example from our work (@yiqqqing) looks at the users who attack political candidates on Twitter: dl.acm.org/doi/10.1145/33…

9/

2) *Publishers* -- organizations or websites spanning multiple authors, contributors (and resources). YouTube channels are somewhere between users and publishers (from @manoelribeiro and @cervisiarius):

dl.acm.org/doi/abs/10.114…

10/

And @grinbergnir used publisher as one of the units of observation in their election 2016 work:

science.sciencemag.org/content/363/64…

11/

3) *Targets* are useful to think about in the context of online safety; accounts/individuals targeted/mentioned in the content, e.g. entities/known individuals (e.g. Bill Gates) or specific users (e.g. @aoc). Our paper in #ICWSM2020 (@yiqqqing):

aaai.org/ojs/index.php/…

12/

Related: @amyxzh excellent paper on targets of email abuse (can't help but mention it though not quite a data-driven analysis):

homes.cs.washington.edu/~axz/papers/sq…

13/

4) *Groups/communities* is the next level, including formally defined ones (FB Group, Subreddit) or not, e.g. classes of individuals (e.g. police; members of an administration) and the study of differences wrt race/gender. Example:

dl.acm.org/doi/10.1145/31… 14/

5) *Campaigns*: Finally, I built on @katestarbird argument that we need to develop a better understanding of information campaigns, but this requires multi-platform, mixed-method, and context-drive approach. A paper from Kate's team you should all read:

dl.acm.org/doi/10.1145/33…

We have an in-submission paper that takes a similar approach, led by @maurice_pj and @gvrkiran -- stay tuned or hit me in a few weeks for an early version!

16/

17/17 All this in case it's useful to think about the units you use in your work + how you can expand by thinking about others. If I have time I'll post about my other takeaways later! #icwsm2020

[Note: used some ACM links but papers should be publicly available, try Googling]

WHAT DID I MISS? I am sure I have made some glaring omissions, even in this context of data and safety/abuse/misinfo -- please advise.