Engagement is not enough!
Thinking about alternative metrics to optimize content curation in online environments
Jana Lasser | Graz University of Technology | jana.lasser@tugraz.at | @janalasser
WhoAmI
Physicist, computer scientist, modelling complex social systems
Focus 1: Quantifying "bad" behaviour on social media
Misinformation sharing: "Social media sharing of low quality news sources by political elites", Lasser et al. 2022, in PNAS nexus.
Misinformation sharing: "From Alternative conceptions of honesty to alternative facts in communications by U.S. politicians", Lasser et al. 2023, arXiv.
Hate speech: "Collective moderation of hate, toxicity, and extremity in online discussions", Lasser et al. 2023, arXiv.
Focus 2: Agent-based models of interventions in complex social systems
COVID-19 spread: "Assessing the impact of SARS-CoV-2 prevention measures in schools by means of agent-based simulations calibrated to cluster tracing data", Lasser et al. 2022, in Nature Communications.
Healthcare system resilience: "Stress-testing the Resilience of the Austrian Healthcare System Using Agent-Based Simulation", Lasser et al. 2022, in Nature Communications.
Social media platforms as information architectures
Social media platforms disseminate information from all domains of life: entertainment, educational, health, news, political, commercial, art, ...
Example Twitter: ~500 million tweets are created and ~150 billion tweets are sent to the devices of ~42 million users every day.
For every user, from a pool of ~500 million tweets about 3600 tweets (0.007%) are selected .
Traffic rules: How is the content selected? How is the content ranked?
Twitter's timeline creation pipeline
How is content selected?
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Twitter's timeline creation pipeline
Social: people a user follows
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Twitter's timeline creation pipeline
Engagement: reactions of other users to content
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Twitter's timeline creation pipeline
Similarity: content and users that are similar to the user
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Twitter's timeline creation pipeline
Final ranking: probability the user will engage with the content
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Twitter's timeline creation pipeline
Filters: "author diversity" and "content balance"
Source: Twitter's Recommendation Algorithm, Twitter, March 31, 2023.
Engagement might not be enough
The business model of large social media platforms is based on capturing user attention for advertising revenue.
This might be a good metric to optimize user entertainment. What about education, health, news, politics, commerce, art, ...?
There is mounting evidence that social media platforms threaten social cohesion [1]. Optimizing for engagement might be part of the problem [2].
[1] "A systematic review of worldwide causal and correlational evidence on digital media and democracy", Lorenz-Spreen et al. 2023, Nature Human Behaviour.
[2] "Facebook Struggles to Balance Civility and Growth", Roose et al. 2021, New York Times.
Beyond engagement
How can we replace or supplement engagement as metric for content delivery?
See also: Building Human Values into Recommender Systems: An Interdisciplinary Synthesis, Stray et al., 2022, arXiv.
Beyond engagement
How can we replace or supplement engagement as metric for content delivery?
See also: Building Human Values into Recommender Systems: An Interdisciplinary Synthesis, Stray et al., 2022, arXiv.
Beyond engagement
How can we replace or supplement engagement as metric for content delivery?
See also: Building Human Values into Recommender Systems: An Interdisciplinary Synthesis, Stray et al., 2022, arXiv.
Beyond engagement
How can we replace or supplement engagement as metric for content delivery?
See also: Building Human Values into Recommender Systems: An Interdisciplinary Synthesis, Stray et al., 2022, arXiv.
Challenges
Value trade-offs
Who decides which values count and how much?
How do short-term metrics relate to long-term outcomes?
Social media platforms are complex systems. Behaviour is hard to predict.
Many long-term outcomes are not measureable from user behaviour
Measuring these metrics requires surveys. Integration of survey signals & long-term outcomes into recommender systems is an open challenge.
Outlook
I know how to operationalise and measure observables of interest on social media platforms.
I also know how to build and calibrate agent-based models of complex social systems to test intervetions.
Next: Modelling interventions in social media platforms to improve long-term societal outcomes (ERC Starting Grant application in fall).
Come talk to me if you hate or like my ideas :)