How algorithm bias leads to shadow banning, systematic discrimination.
Shadow banning refers to limiting or obscuring contents of some users or user communities in such a way that it is not directly apparent to these users.
Problem: These affected users usually belong to minority group and have fewer followers to begin with. Limiting their visibility leads not only leads to polarization of ideas but also systematic discrimination and suppression of minority groups.
Why does shadow banning occurs:
- All social media apps use content moderation algorithms.
- These algorithms essentially learn user preferences by monitoring engagement such that users will mostly see/recommended posts or brands that they are expected to engage with.
- Additionally, these algorithms are also trained on historic dataset to look out for particular feature sets. Which potentially introduces the bias, because within software fraternity there is less diversity and in-adeptly the models coded and data collected to train is usually based on the dataset and experience of prominent community (i.e. white, male and American).
- Further, goal of these platform is to make money, and these algorithms are profit driven. On the downside the users who do not generate engagement or profit will eventually become less visible. As a result, users from minority who have smaller community that share same viewpoint will have consequently less followers, and run at risk of being obscured.
- On surface this content moderation may appear benign or even good recommendation system, but these algorithms have been in scrutiny for banning under-represented groups such as several black, Hispanic, plus-sized women, pride, and minorities. Because these groups have fewer people who agree to their voices and therefore generate lesser engagement. Suppressing voices of underrepresented groups leads to systematic discrimination and polarization of ideas.
Notable examples:
- In 2018, after rigorous media coverage, Facebook as part of a legal settlement with civil rights groups had to disable it’s tool allowed options to advertisers to filter out exclude multiple ethnic groups, religious groups, and other protected classes from not just housing ads.
- During summer of 2020, Instagram removed pictures of Black plus-size model Nyome Nicholas-Williams for ‘explicit images’ despite the fact that Nyome’s pictures were completely concealed at least more than that of Kardashians or any thin influencer. Nyome herself argues that the “algorithms on Instagram are bias against women, more so Black women and minorities, especially fat Black women. When a body type is not understood I think the algorithm goes against what it’s taught as the norm, which in the media is white, slim women as the ideal.” Similarly, there has been rise in accusations against Instagram for banning the accounts of plus size women, and other minority groups.
Ongoing research in area:
While there have been countless allegations regarding the shadow banning, almost all companies have denied these mishaps usually as “bugs”. There has been very less public research to investigate and quantify these incidents.
I discuss here very briefly one recent research (link can be found here)
This research collects the public profiles and interactions of millions of Twitter users, as well as their shadow banning status.
This paper tests two hypothesis:
- H0: shadow banned users are uniformly spread among Twitter users, which corresponds to Twitter’s random bug claim. And through statistical analysis shows that this hypothesis is wrong, i.e the claimed “bugs” are not random.
- The paper then proposes and tests second hypothesis, H1: the topological hypothesis i.e shadow bans occurs for specific users with certain feature sets.
For second hypothesis, the research is built on the idea “that if one can predict with some reasonable accuracy if a profile is shadow banned by only looking at its features, then these features are encoding a part of the cause of a profile being banned.” For this they use there machine learning classifier algorithms a random forest algorithm (RF), the AdaBoost algorithm (AB), and a decision tree (DT).
And finds that these models have reasonable accuracy, i.e these models allow to precisely pinpoint the influence of features on the classification accuracy.
The paper concludes:
That the hypothesis, H1 is correct and that the bans appear as a local event, impacting specific users and their close interaction partners, rather than resembling a (uniform) random event such as “bug”.
What can be done to fix Shadow ban?
- More research in examining the underpinnings and limitations of machine-based systems for online content moderation and recommendation.
- More research on building robust models and training samples to avoid systematic discrimination.
- Inclusion of more people from diverse background in the entire development process from algorithm and model development. Diversity is one issue that many organizations still struggle with, as a result these platforms are developed by predominantly homogeneous group (white, male, American). Further these potential issues are never thought of during development or during training stage. Algorithms are trained on data gathered is mostly reflective of the history and experience of developers who are representative of fewer demographic profile. Thus, the unconscious bias of developers remains embedded in systems they create.
- Government policy for research and fixes for these gaps and make social media fairer for al groups and communities.
Reference:
https://arxiv.org/pdf/2012.05101v2.pdf
https://techxplore.com/news/2021-01-exploring-underpinnings-shadowbanning-twitter.html
https://toronto.citynews.ca/2021/04/05/the-growing-criticism-over-instagrams-algorithm-bias/
https://www.bbc.com/news/technology-57306800
https://en.wikipedia.org/wiki/Shadow_banning
<a href=’https://www.freepik.com/vectors/design'>Design vector created by macrovector — www.freepik.com</a>
#artificialintelligence #neuralnetworks #machinelearning #machinelearningmodels #randomforest #clustering #algorithmicbias
