The company has shared very few details about its algorithms. So far we know that they use DeepBird (based on Lua Torch), a deep learning system that predicts which tweets users will find interesting and engaging.
A new research study done by Jack Bandy from the Computational Journalism Lab reveals new findings of Twitter’s timeline algorithm and how it directs the attention of its more than 150 million daily active users.
To get these insights, Jack and his team used a group of automated “puppet accounts,” comparing the puppets’ chronological timelines (“latest tweets”) to their algorithmic timelines (“top tweets”). You can read how he did it here and his research papers here.
Here are some of the insights based on Twitter timelines(chronological and algorithmic) collected from a group of puppet accounts from April 10th to May 11th, 2020.
Fewer External Links
Twitter’s Algorithmic Timelines greatly favoured text-only tweets as opposed to tweets with external links. On average, 51% of tweets in Chronological timelines contained an external link, compared to just 18% in the Algorithmic timelines.
Lots of “Suggested” Tweets
Less than half of all tweets in the algorithmic timeline came from accounts that the puppets actually followed. On average, “suggested” tweets (from non-followed accounts) made up 55% of the algorithmic timeline.
This finding comes from data collected from the timelines twice per day. Jack adds that they only analyzed the first 50 tweets that appeared and that suggested tweets would probably be less prevalent if they looked at the first 200 tweets that appeared instead of just the first 50.
Increased Source Diversity
Twitter’s timeline almost doubled the number of unique accounts in the timeline, from 663 in the average chronological timeline to 1,169 in the average algorithmic timeline.
The ten most-tweeting accounts made up 52% of tweets in the chronological timeline, but just 24% of tweets in the algorithmic timeline.
The “filter bubble” metaphor is popular, but this recent study adds to the growing counter-evidence that Twitter’s algorithm increased the number of accounts in the timeline with accounts that would not appear in a chronological timeline. The effect was more nuanced for topic exposure and partisanship.
Slight Shift in Topics
Jack analyzed four tweet clusters related to the COVID-19 pandemic to get a glimpse of how the timeline algorithm may shift the topical makeup of user timelines.
- A cluster of tweets containing political information (e.g. about the president’s response to the pandemic)
- A cluster containing health information (e.g. about risk factors)
- A cluster containing economic information (e.g. about GDP or job loss)
- A cluster about fatalities (e.g. death toll reports)
Overall, Twitter’s algorithm reduced exposure to each of these topics except for the political cluster:
This evidence suggests that social media algorithms might sometimes reduce exposure to important information (e.g. health and fatality information about COVID-19), while elevating other topics.
In his report, Jack adds that the effect did not constitute a lock-tight “filter bubble,” but more of an “echo chamber” where some topics became louder while others were drowned out.
Slight Partisan Echo Chamber Effect
In analyzing partisanship, they did see a slight echo chamber effect across the puppets. For example, the algorithmic timeline decreased exposure to accounts that were classified as bipartisan.
Importantly, this was not a test of whether the algorithm had a “political bias,” but rather how the algorithm affected exposure to partisan accounts compared to users’ chronological timelines.
For left-leaning puppets, 43% of their chronological timelines came from bipartisan accounts (purple in the figure below), decreasing to 22% in their algorithmic timelines:
Right-leaning puppets also saw a drop. 20% of their chronological timelines were from bipartisan accounts, but only 14% of their algorithmic timelines:
The research reports that their scoring system for partisanship was pretty generous to Twitter as it labelled Barack Obama and Fox News as “bipartisan” (because these accounts were commonly followed both in right-leaning communities and left-leaning communities). The echo chamber effect may have been different under an alternative scoring system.
One of the most shocking findings is the reduced exposure to external links that majorly affects links to news websites since Twitter users won’t be exposed to high-quality journalistic media and impact news sites that rely on traffic.
The research adds a few things that can be done to improve how news and information is shared online. Suggestions not only include potential tweaks to the algorithm but also media literacy, user interfaces, structures of governance, and regulatory measures which all present ways to improve the ecosystem.