15 Mar

Playing with algorithms 3: But what about Twitter?

I don’t use Twitter’s Discovery timeline, because although it is ‘curated for you’, I don’t find it curates what I’m looking for. However, it IS using my data, and it is  different between my desktop version and the mobile app that I use more often.

I selected the the top 10 tweets in my Twitter Discovery timeline in both Mobile and Desktop versions, at 1.30pm Sunday 15 March. (They are listed below, under Appendix). I carried out a content analysis of the tweets to see why Twitter might serve them to me.

***

Content analysis

Content analysis: coding

Analysis.

All the tweets included some visual component: a screet shot, picture or gif. Most of the accounts are large, established accounts belonging to publishing platforms or authors.

I did a quick content analysis, where I coded the text, image, embedded content, or information from the link (but not following the link) in each tweet into:

  • Geek culture
  • Gender or cultural Diversity
  • Privacy/online security
  • Technology
  • Hard News
  • Science
  • Pop culture
  • History
  • Writing
  • Visual attraction (pictures that accompanied the tweets were pretty, rather than funny or screen texts)
  • Food
  • Academia
  • Humour
  • Inspiration

(These categories are in order of identification, not anything more significant).

Tweets were coded with as many of the above catagories as might be relevant. Saladin Ahmed is a fantasy writer and geek/pop culture critic, he writes about racial, gender and religious diversity, and he often posts humourous tweets. He appears at the top of both lists, once more on desktop and twice more on my mobile list. His top tweet registered in 7 of the above catagories (the highest of any tweet).

Twitter thinks this is my ideal Tweet.

Twitter thinks this is my ideal Tweet.

There were 5 tweets coded Humour on mobile and 4 on desktop. 4 tweets coded Geek culture (Desktop: 3) and Diversity (Desktop: 2) on my mobile timeline. 3 tweets coded Privacy/online security (Desktop: 2) and History (Desktop: 1). Overall, mobile Discovery tweets registered 34 categories, where as desktop tweets only fitted 28.

Therefore, categories that met my interests were more likely to be served to me in the Discovery timeline of the mobile app than the desktop web version. 

However, this left a large gap–what was influencing the desktop version to rate these particular tweets highly?

I had included the information Twitter shares in grey above a tweet to tell me why it’s in my timeline in the descriptions (See appendix). One clear difference stood out so I added another emergent category:

  • High Retweet count

Of the tweets in the desktop Discovery timeline, 5 had a high retweet count (66-197). Of the tweets in the mobile app, there was only 1. This accounted for most of the gap between desktop and mobile.

In the absence of individual engagement data (suggested by what I favourite, RT, click on, reply to etc), Twitter uses general popularity as a category to decide what is going to appear on my Discovery timeline.

And yes, I did make pie for Pi day.

I made pie for 3.1415 day.

I made pie for 3.1415 day.

***

Appendix 1: The top 10 tweets

1. Saladin Ahmed (who I follow, RTd by another person I follow) on the Telegraph pointing out that now, gasp, white men are being targetted for online abuse too. [Desktop & mobile]

2. Slate (who I follow, RTd by 147 other people) on Pi day 3.1415.  [D]

New York Times (who I follow, RTd by William Gibson) on how the ‘tech titans’ of Silicon Valley who created platforms that require and harvest personal information are protecting their own privacy. [M]

3. Bibliophilia (who I follow, RTd by 66 other people) on an 18th century double bible. [D]

Newsweek (followed by someone I follow) on whether the Antropocene started with the Native American genocides [M]

4. The Economist (who I follow, RTd by 197 other people) on the increased risk of nuclear war.

Classicpics (RTd by someone I follow) on the ‘perfect body’ in 1955 (the woman is a healthy weight, with muscles and curves). [M]

5. New York Times (who I follow, RTd by William Gibson) on how the ‘tech titans’ of Silicon Valley who created platforms that require and harvest personal information are protecting their own privacy. [D]

Slate (who I follow, RTd by 147 other people) on Pi day 3.1415.  [M]

6. Saladin Ahmed (Rtd by 2 people I follow) with a Star Trek reaction gif about Gizmodo, diversity. [D]

Science headline of the week about snail sex (RTd by someone I follow). [M]

7. Science headline of the week about snail sex (RTd by someone I follow). [D]

Saladin Ahmed (Rtd by 2 people I follow) with a Star Trek reaction gif about Gizmodo, diversity. [M]

8. Wall Street Journal (who I follow, RTd by 166 other people) on the return of the business phone call. [D]

Endgadget (followed by someone I follow) on the fact that new HBO still ‘has strings attached’. [M]

9. Advice to Writers (who I follow, RTd by 82 other people) with a quote from Maya Angelou. [D]

Person RTd by someone I follow on 11th century limestone head discovered in Norfolk.[M]

10. An academic I follow (favourited by another academic I follow), joking television marathons should count towards tenure. [D]

Saladin Ahmed, with a reaction gif from Lord of the Rings about having too much work to do. [M]

7 thoughts on “Playing with algorithms 3: But what about Twitter?

  1. Really great ‘reverse engineering’ of the Twitter timeline Katherine.

    ‘Of the tweets in the desktop Discovery timeline, 5 had a high retweet count (66-197)’

    So, in other words, the assumption here is that popular tweets are going to be more important and relevant to you? Is that the case? Why would Twitter assume that this *is* important. I wondered if you could say a bit more here about what the implications of you discoveries here might be. You’ve done a great job of exposing the automated processes at work behind the façade of the discovery timeline, but what are the consequences, and how significant are they?

    I must have missed Pi day, what a shame…your pie looked rather delicious though.

    • Thanks Jeremy!
      That’s a good question. My guess is that, based on more limited data, Twitter is not going to personalise it’s curation quite so individually. Instead, it will make assumptions that I will behave like ‘most’ people. Popularity therefore is going to be significant–‘lots of people like this, so you are more likely like this’. This is a reasonable assumption to make for a generic user.
      The app, on the other hand, is able to weigh up more data points (for me, because of how I use it), and therefore was able to be more specific. Because I have a lot of niche interests, the distinction is visible on my two timelines. If I engaged more with popular tweets, I doubt that would have been so clear.
      That’s my guess, but I’m about to go off and do some more research, as Ed suggests!

  2. Katherine, yours is definitely the most ‘appetizing’ blog site. A very instructive and informative series of blogs with analysis and synthesis on Twitter, tweets and tweeting. I am still coming to grips with Discovery timeline and Twitter audit that you’ve introduced. Cheers, PJ

  3. A pie for PI day, looks delish! Twitter also captures geolocation data, although this may be not necessarily be factored in for the discovery timeline. The differences in desktop and mobile timelines suggest that Twitter also makes assumptions about how people interact with content based on devices (desktop or mobile) . I think you’ve uncovered many interesting things here. In fact, when I googled for “twitter discovery timeline algorithm” this page was in the top 10 results.

    • Woah! That’s very surprising! thanks Ed!
      Yes, I suspect that geolocation data would be used, but I live pretty close to where I work, so I wouldn’t expect that to factor in the differences.
      I’d be interested in what the differences are between typical desktop and mobile use (I would assume something to do with ‘work’ vs ‘leisure’, but I’ll have to search!).

      • Okay, did some research (my two blogs are now the top results for that search term!).

        Most of all, mobile users are usually in a different frame of mind than desktop users. We can’t really make any precise assumptions about the context of the device usage, but compared to desktop usage they quite often will be standing up and walking around or driving, and sometimes just scanning their handset or tablet. This might call for less detailed content than what you would use for more focused desktop users. Instead of presenting all of the day’s news items, for example, tracking unread items or the most popular trending items might be more important. Users might also want different content depending on their location.Smashing Magazine

        Interestingly, this suggests that mobile users might want more “popular trending items” rather than less.

        But the article goes on to say:

        Other times, such as when waiting for a connecting flight, mobile users will have more time on their hands and attention to spare.

        Therefore, the app might be responding to how most people are now using their mobiles (at home, while commuting, with time to spare), or at least how I use my mobile (definitely desktop Twitter is used between other work, where as mobile Twitter is regularly used instead of television as an extended leisure activity).

Leave a Reply to Ed Guzman Cancel reply

Your email address will not be published. Required fields are marked *

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>