This semester I had the opportunity to help out in a few sessions of Matt Kirschenbaum’s Digital Studies graduate seminar. Matt wanted to include some hands on exercises collecting data from the Twitter API, to serve as a counterpoint to some of the readings on networks such as Galloway and Thacker’s The Exploit. I don’t think the syllabus is online, but if it becomes available I’ll be sure to link it here.
Now ENG 668K wasn’t billed as a programming class, so getting into Python or R or <insert your favorite programming language here> just wouldn’t have been fair or appropriate. But discussing the nature of networks, and social media platforms really requires students to get a feel for the protocols and data formats that underpin and constrain these platforms. So we wanted to give students some hands on experience with what the Twitter API did and, perhaps more importantly, didn’t do. We also wanted to get students thinking about the valuable perspectives that humanists bring to data.
The Graph From 10,000 Feet
I started with a general introduction to Twitter and the graph as a data model. Much of this time was spent doing a close read of a particular tweet as it appears in the Web browser. We looked at the components of a tweet, such as who sent it, when it was sent, hashtags, embedded media, replies, retweets and favorites, and how these properties and behaviors are instantiated in the Twitter user interface.
We then took a look at the same tweet as JSON data, and briefly discussed the how and why of the Twitter API. We wrapped up with a quick look at the history of graph theory, which was kind of impossible and laughable in a way. But I think it was important to at least gesture at how networks and graphs have a long history, have accumulated a lot of theory, but in essence are quite simple.
As an exercise we all created paper graphs of some topic. There’s no need to get lost in Gephi to start thinking about the graph data model. I showed them a paper graph that modeled a small subset of MITH people and projects as an example:
I wish I took pictures of the ones all the students came up with. They were amazing and made my example seem super dull. MITH is in the middle of working on a viewer for all our projects over the past 15 years, so we have a dataset of the people and projects available. I created a hairball graph view of the full dataset just to balm my damaged pride.
Tags and TAGS
But this was all just a prelude to where we spent the majority of our time in the subsequent classes, which was in using the Twitter Archiving Google Sheet or TAGS. The thing that TAGS did well was get students thinking about the various representations of a tweet, and how those representations fit different uses. The representation as HTML is clearly good for the human user on the Web. The JSON representation was designed for API access by clients on mobile devices, and other services. And one of those other services is TAGS, which takes the JSON data and puts it into a familiar tabular layout for analysis. This complemented an earlier set of sessions in the semester where Raff Viglianti and Neil Fraistat talked about reimagining the archive, with a dive into encoding a Shelley manuscript using TEI.
We did run into some problems using TAGS, mostly centered on the Twitter API Keys. To use TAGS you need a Twitter account, and a Google account. This may seem like a no-brainer but I was impressed to see there were some students who assiduously wanted neither. Respect. Once you have a Twitter account you need to register an app, and get the keys for the app. In order to create an application you need to attach a mobile number to your user account. This is understandable given Twitter’s ongoing fight against spam, but it presents another privacy/security conundrum to students. But once the keys are in hand TAGS is mostly straightforward to enter a search query and get some data.
The wrap up exercise for the hands on TAGS experimentation was to do some data collection and then write a very basic data story about that data. I felt like I appropriated the phrase data story a little bit. It’s a new term that I never really contextualized fully because of time constraints. Data storytelling could be a whole class in itself. In fact Nick Diakopoulos offered one this semester at UMD. But we simply wanted the students to try to collect some data and write up what they did, and what (if anything) they found in the data. To get them started a provided an example data story.
My Hastily Composed Data Story
For my data story I assembled a Twitter data dataset for the poet Patricia Lockwood. I collected 1,592 of her tweets for the period from April 20, 2015 to November 22, 2015. To do this I configured TAGS settings to fetch using the API type
status/user_timeline for the user TriciaLockwood. Lockwood has tweeted 12,658 times since May 2011, so I wasn’t able to get all of her tweets because of the limitations of Twitter’s API.
I thought it might be interesting to see who Lockwood corresponds with the most for the time period I was able to collect in, so I created a pivot table (the Correspondence sheet in the Google Sheets document) that listed the people she directed tweets at the most. The top 5 users she corresponded with were:
- @cat_beltane - Gregory Erskine
I thought it was interesting that the top person she corresponded with didn’t seem to be easily googlable even though he has a somewhat distinctive name. Erskine’s profile says he lives in Louisville, Kentucky which isn’t terribly far from St Louis and Cincinnati, both of which are places Lockwood has lived in the past. According to the New York Times she seems to be currently living in Lawrence, Kansas. Perhaps they knew each other in the past and are keeping in touch via Twitter?
The http://lfriendys.com website Erskine has linked from his Twitter profile was registered in 2014 using an address in St Louis. So perhaps they knew each other there. It’s possible to search and scan all their public correspondence, which goes back to September 16, 2011.
Ok, this is starting to feel creepy now.
The resulting data stories were really quite amazing. As I said I barely gave any context for data storytelling, but the students ran with it and well surpassed my example. I can’t include them here because I didn’t ask for their permission. But if any students decide to put them on the Web I will link to them.
Matt did ask me to write up a general response to these data stories, with particular attention to where TAGS break down. I’ve included it here, but I’ve changed the students names, since they didn’t write them thinking they would be on the Web with their names attached.
Here’s what I wrote…
I really enjoyed reading the data stories that everyone was able to put together. One phrase that will stick with me for a while is Julie’s hashtag salad as a term (or neologism really) for tweets that are mostly made up of a coordination of hashtags. I’m going to use it from now on!
I was surprised to find that the stories that established a strong personal voice were particulary effective in drawing out a narrative in the data. Perhaps I shouldn’t have been, but I was :) I guess it makes sense that a narrative about data would need to have a narrator, and that the narrator needs a voice. So it stands to reason that attention to this voice is important for communicating the perspective that the data story is taking. I think this is one reason (among many) that humanists are needed in the STEM dominated field of data science. Voice and perspective matter.
Another important aspect to the stories that worked well was an attention to the data. For example Courtney’s analysis of the device used (mobile vs desktop) and time of day was quite interesting. The treatment is suggestive of the way the data is intertwined with events in the physical world, and the degree to which social media can and cannot mediate those events. Samuel’s analysis of the use of language (English, Spanish and Portuguese) by a particular individual in different contexts was another example of this type of treatment.
But there were many interesting things that were discussed, so it’s probably not fair for me to start highlighting them. The stories contained several comments about the limitations of collecting data from Twitter using TAGS. For example:
Not only does TAGS give us a very brief snapshot of a giant conversation, but we Twitter users must figure out what we need before too much time passes and the data starts to become very hard to draw out.
This is one limitation of TAGS, as it removes the physical layout of Twitter, requiring users to cut and paste hyperlinks to read the original content of tagged Tweets.
The first was that it still takes a human to connect all of the patterns together. Google spreadsheets do not contain a function to sort text by theme, so the user still needs to look at the text to see if there is a pattern to the Tweets.
The second is in the limitations of TAGS. Vince and I were both initially surprised with the seemingly low volume of results we received. After a little research, we realized that TAGS only collects seven to nine days of Tweets.
Of course, it is impossible to tell just from the data how much @Farmerboy or the other’s meant to, or not, to refer to the Little House on the Prairie franchise.
For me, temporality posed the largest problem in that it was important to time when to run the script to capture certain kinds of data.
Not only is the process capped by a maximum number of tweets archived at a specific moment, it is also difficult to explore a topic outside of the “now” because of its relegation to the last seven days, rather than allowing for a specified date range. Additionally, the more tweets I attempted to archive at one time, the greater likelihood the script would fail and yield zero results.
I thought I remembered mentioning that the search API was limited to the last 7-9 days, but even if I did I clearly didn’t emphasize it enough. The search API does restrict access mostly for business reasons since Twitter have a service called Gnip which allows people to purchase access to historical data in bulk. So, if you are interested in a topic, and don’t want to pay Twitter thousands of dollars for data, it is important to collect continuously over a period of time.
TAGS tries to do this for you by allowing you to schedule your search to be rerun, but there are limits to the size of a Google Sheet: 2,000,000 cells, or 111,111 TAGS rows. It also isn’t clear to me how TAGS deals with duplicate data, or how it ensures that it doesn’t have gaps in time. At any rate these observations about the limits of TAGS and the underlying Twitter API are great examples of getting insight into Twitter as a platform, in Tarleton Gillespie’s use of the term. If this sort of thing is of interest there is an emerging literature that analyzes Twitter’s as a platform, such as González-Bailón, Wang, Rivero, Borge-Holthoefer, & Moreno (2014), Driscoll & Walker (2014) and Proferes (2015).
Just as an aside Twitter’s web search isn’t limited to the last 7-9 days like the API. For example you can do a search for the tweets mentioning the word
twttr (Twitter’s original name) before March 22, 2006 which will show you some of the first day of tweets from Twitter’s founders.
The comments also point to another limitation of TAGS as a tool. The spreadsheet has the text of the tweet, but it is extremely data centric. To see embedded media, the users profile information, the responses and the full presentation of the tweet it is necessary to visit the twett on the Web using the URL located in the
status_url column. This can prove to be quite a barrier, when you are attempting to decode the intent or intended meaning of a message by simply browsing the spreadsheet. The additional context found in the human readable presentation of the Web page makes it much easier to get at the intent or meaning of a tweet. But how do you do this sort of analysis with thousands of messages? This raises good questions about distant reading, which is also an area where a DH perspective has a lot to offer to the data science profession.
It will be interesting to hear what the students made of the class in the reviews. But all in all I was kind of surprised at how low-tech instruments like paper graphs, spreadsheets and data stories could yield valuable thinking, discussion and analysis of social networks. Of course I expect most of that was due to the readings and lectures that went on outside of these exercises. I’m looking forward to hopefully being able to iterate on some of these techniques and ideas in the future.
Driscoll, K., & Walker, S. (2014). Working within a black box: Transparency in the collection and production of big Twitter data. International Journal of Communication, 8, 1745–1764. Retrieved from http://ijoc.org/index.php/ijoc/article/view/2171
González-Bailón, S., Wang, N., Rivero, A., Borge-Holthoefer, J., & Moreno, Y. (2014). Assessing the bias in samples of large online networks. Social Networks, 38, 16–27. Retrieved from http://papers.ssrn.com/sol3/papers.cfm?abstract\\_id=2185134
Proferes, N. J. (2015). Informational power on Twitter: A mixed-methods exploration of user knowledge and technological discourse about information flows (PhD thesis). University of Wisconsin, Milwaukee. Retrieved from http://dc.uwm.edu/etd/909/