Reducing Algorithmic Amplification on Social Media
EXECUTIVE SUMMARY
The recommender algorithms used by social media companies amplify biases and polarize already-divided public opinions. However, their exact impacts are unknown and there is no consensus on how algorithmic amplification can be reduced. The social media NUDGE act is a timely proposal that establishes studies on previously unknown negative impacts of social media recommender algorithms. And it would help regulators to design regulations that mitigate these impacts. Nonetheless, the scope of regulation and the amount of information a social media company is required to disclose can be more precisely defined. Overregulation or too much transparency in algorithms would damage existing business models, and become unsustainable to achieve their goals in the long run. Regulatory efforts should be content-agnostic and leave business models intact for them to be sustainable in the long run. The NUDGE act could be strengthened by solely focusing on the outcome of algorithms and data collection processes.
INTRODUCTION
Due to the rapidly growing amount of information online, a recommender algorithm is the only feasible way for users to extract relevant information from social media. However, the algorithm social media companies employ to increase user engagement inadvertently create echo chambers. In particular, the recommender algorithm many social media platforms use select content based on the user’s personal preferences, which helps users to find relevant information quickly. However, it can also reinforce the user’s misconceptions and political ideology. There is an enhanced probability to choose individuals whose opinions are already close to each other, thus mimicking the behavior of online media which suggest interaction with similar peers. (Sîrbu) On the other hand, recommender algorithms also exacerbate the spread of misinformation. By recommending social media users information or viewpoints they align with, recommender systems are also responsible for creating echo chambers. As a result, people within the echo chamber are less likely to believe in alternative views and become vulnerable to misinformation. Consequently, political opinions become increasingly polarized and consensus could not be reached (Choa, Ahmedb, Hilberta, Liuc, Luu 2020). This phenomenon is known as algorithmic amplification. As social media become increasingly significant factors of the quality of life of Americans, policymakers start to recognize the potential negative impacts of social media and how algorithms can amplify these effects.
The first step
The Nudging Users to Drive Good Experiences on Social Media (NUDGE) act was a bipartisan bill introduced by senators Amy Klobuchar and Cynthia Lummis. The act established the profound impacts social media can have on its users can be results of design flaws of recommender systems. The act then proposes both internal and independent studies that examine the effects of social media algorithmic amplifications. Based on study findings, the legislation would require social media platforms to implement content-agnostic changes that do not affect user content and core function of social media platforms.
The social media NUDGE act increases transparency of the effects of recommender algorithms, and it is necessary towards managing the negative impacts of social media. Given that the impacts of recommender algorithms are unknown until more research is conducted, I believe such a law is a timely step in the right direction. However, the bill is vague in defining how changes should be implemented and what metric should be used. It can be better defined and extended to alleviate the negative impacts of algorithmic amplification.
DESIGN PRINCIPLES
Regulations on recommender algorithms should focus on these two principles:
- Content-agnostic
The First Amendment to the U.S. Constitution prohibits censorship by the government. Regulations should only alter user-experience and not any user-generated content.
- Minimizing user base impacts
For these regulations to be effective, the basic business models should be unaffected. If the regulations were too heavy-handed, users would leave and switch to different platforms.
ALTERNATIVE PROPOSALS
The European Union is often seen as the global leader in the regulation of digital technologies. On October 19 2022, the Digital Services Act (DSA) was formally published in the Official Journal of the European Union. The DSA encompasses a set of regulations that tackle disinformation and harmful information online. It prescribes two main ways that the recommender algorithm can be regulated. However, I believe these two strategies are ineffective. The proposals in the DSA should not be incorporated into the future of digital regulation in the US.
- Turning off the algorithm
First, the DSA proposed a requirement for social media to include a toggle that allows people to switch off recommender systems. The users who continue to use recommendation algorithms will receive a notification that those recommendations are based on inferences generated by their personal data. A similar bipartisan bill “Filter Bubble Transparency Act” was also introduced in senate 2021 which also involved the ability to remove the recommender system. By removing the recommender algorithm, social media posts will be sorted in a reverse chronological order, and amplification effects can be reduced. However, it also makes large social media platforms unappealing. While a reverse chronological feed is possible at a smaller scale, it would eradicate larger social media platforms’ business models. As social media platforms grow in scale, the information they host increases exponentially. Assuming there is no recommender system in place, the time taken for a user to find relevant content would increase exponentially too. Recommender systems were proven to improve user experience (Knijnenburg, Bart). Thus, the removal of recommender systems would make large social media significantly less appealing to users. This goes against the principle that regulation should preserve core functionality of individual social media platforms. Therefore, a reverse chronological feed is ineffective in reducing algorithmic amplification.
- Tweaking the algorithm
Another regulatory tool discussed in the DSA is to modify certain features of the algorithm to reduce amplification effects. In a similar vein, Senator Edward J. Markey and Congresswoman Doris Matsui introduced the Algorithmic Justice and Online Platform Transparency Act of 2021 to prohibit harmful algorithms. Online platforms may not use features such as race, age, gender, ability and other protected characteristics for their recommender algorithms.
For such a regulation to be successful, the algorithm itself has to be transparent to independent organizations that conduct tests on the recommender system. The studies would allow engineers to glean a better understanding of the negative impacts of the algorithm and what to be done to amend it. However, if the changes were to become a legislative requirement, the time lag between approval and implementation of the modified algorithm would be problematic. Recommender algorithms are growing at a steady pace. In 2018 the deep learning algorithm AlphaZero had 500 million parameters. As of today, the algorithm behind Facebook’s newsfeed has 12 trillion parameters (Mudigere). The sheer size of these algorithms make understanding their ranking of information impossible to comprehend. Social media companies can use strongly correlated parameters to replace any banned parameters. Rendering regulation ineffective. Moreover, the ever-increasing scale of recommender algorithms would render any modification ineffective. If a change were to be made, it has to go through rounds of studies and bureaucratic approval. The change may no longer be relevant by the time it is implemented.
Moreover, studies on algorithms involving retraining of models are not sustainable. The growth of computing needed for retraining is outpacing computing power. Since the 2010s, AI systems are becoming increasingly computationally expensive to train as computing power used in training AI systems doubles every 3.4 months (Cross 2020). This growth dwarfs Moore’s Law, which states that the speed and capability of computers would increase every two years. The divergence between the complexity of recommender systems and computing power would become increasingly stark. Therefore, the efforts on studying the algorithms themselves could only be sustained with increasing amounts of investments. On the other hand, the studies would take longer to conduct and changes can only be implemented at much later times. The time lag between proposal and implementation of change is expected to increase. The proposal of requiring social media to disclose their algorithms would be a short-term solution at best. It is unrealistic that regulatory forces have the resources to compete with multiple large tech-companies. Attempts on studying details of recommender algorithms would be a diversion of resources from achieving long-run goals. In order for regulations to be sustainable, resources should be utilized on designing solutions that translate well into the future.
On the other hand, the government can demand the algorithms to update less frequently and hence grow at a slower pace. Such regulatory efforts on algorithms themselves would have to put a halt on technological progress on recommender systems, which may in turn impede the social media platforms’ competitiveness. As previously mentioned, damaging the attractiveness of individual social media platforms can be counter-productive. Tweaking the algorithm is a band-aid solution at best.
RECOMMENDATIONS
The NUDGE act can be strengthened in several ways. To reduce the negative impacts of recommender systems, it is important to recognize that both inherent design flaws and existing biases in dataset can hinder the performance of these algorithms. Thus, I propose a simultaneous approach that targets both aspects, and is compatible with the design principles. The recommendations can be included in future digital regulation bills.
- Issue 1: Biases in dataset
A common notion among data scientists is that algorithms merely reflect the biases in the dataset. In real life, datasets often have long right tails. Minority groups are less represented in the dataset, therefore algorithms can have biased prediction results on these groups (Hooker 2021). The cause of bias and unfair predictions can be independent of recommender model learning process (González, Ortega, Pérez-López, Alonso). Hence, eliminating these biases involves using new datasets. Collecting more data is expensive and does not eliminate the unequal effectiveness in algorithms. Moreover, biases in algorithms can also jeopardize business decisions.
There are two possible methods to mitigate biases in datasets. First, reducing sampling bias in data collection processes. This issue is being comprehensively studied, and there are a multitude of possible solutions to this problem such as dataset resampling (Li, Yi, Vasconcelos). Since this topic is well covered and the solutions can be more nuanced based on different context, the discussion cannot be meaningfully covered in a few paragraphs. Hence, I choose to focus on the proposal to reduce dataset bias by creating synthetic data.
Second, biases can be minimized in the dataset using data augmentation techniques. Data augmentation is the process that artificially generates more data points using existing data. By adding synthetic data points, the dataset would represent that of an “ideal world”. Therefore, data augmentation can significantly reduce sampling bias and prejudice based bias while keeping model accuracy constant (Sharma, Zhang, Aliaga, Bouneffouf, Muthusamy, Varshney 2020). However, this is a relatively new technique and the standards are different across companies. More studies should be conducted on social media companies’ data augmentation practices. Nevertheless, recommender systems should be required to be trained on sufficiently augmented datasets. A random sample of the augmented data can be used as a metric in measuring the companies’ commitment in combating algorithmic biases.
- Issue 2: Design flaws in algorithms
While dataset hinder the effectiveness of recommender algorithms, it is not the only reason causing algorithmic amplification. Some recommender algorithms are more likely to amplify negative impacts than others (Hooker 2021). Thus, clearly improvements can be made on the design of recommender systems.
After following Twitter accounts from the opposing political spectrum, both republicans and democrats became more entrenched in their beliefs (Bail, Argyle, Brown 2018). Evidently, simply exposing social media users to alternative viewpoints would not reduce opinion polarization. Instead, such instances can be perceived as confrontational. Social media users are more likely to believe information from people they feel a personal connection in (Saveski, Gillani, Yuan, Vijayaraghavan, Roy 2021). These individuals can be people acquainted in real life or people who share similar beliefs. By altering the recommender in such a way that increases exposure to alternative views from unfamiliar accounts, while it may increase user engagement, opinion polarization would worsen. Therefore, regulation should aim at decreasing the harmful contents shown to the users.
As previously mentioned, changing the details of recommender systems in hope to bring about targeted outcomes is unfeasible. Thus, recommender systems should be trained based on their algorithmic impacts. Researchers at Meta proposed a theoretical framework that allows recommender systems to dynamically react to algorithmic amplification. First, to identify the extent of amplification, the change in user preferences towards recommended content can be measured and quantified. If the users start “drifting” towards problematic content, the recommender system would limit the exposure of these contents. Solutions derived from this framework were successful in mitigating amplification while increasing user engagement by 2% (Kalimeris, Bhagat, Kalyanaraman, Weinsberg). Borrowing from this concept, the “drift” in user preferences can be used as a metric mentioned in the NUDGE act. By requiring social companies to reduce the “drift” towards extremist content below a certain threshold, algorithmic amplification can be mitigated. While this method is front-loaded as it requires re-training of recommender systems initially. Verifying its effects is relatively easy. On the other hand, it does not require social media companies to disclose the details of their algorithms. Therefore, they would not lose their competitive advantages to foreign platforms unaffected by the regulation.
CONCLUSION
While more transparency and regulation are needed to understand and mitigate algorithmic amplification. These regulations must be narrowly tailored to the problem to avoid undesirable side-effects. Algorithms should remain at the hands of social media companies. Regulations should not be fixated on changing the algorithm itself, as algorithms are too complex to implement a change with predictable effects. Biases in algorithms either come from the algorithm design or dataset. The NUDGE act should instead target both flaws at the same time: data augmentation and user preference drift. I believe the effectiveness of the NUDGE act can be significantly bolstered by my proposals.
References:
Bail, Argyle, Brown “Exposure to opposing views on social media can increase political polarization” 2018 https://www.pnas.org/doi/full/10.1073/pnas.1804840115
Choa, Ahmedb, Hilberta, Liuc, Luu “Do Search Algorithms Endanger Democracy? An Experimental Investigation of Algorithm Effects on Political Polarization”Journal of Broadcasting & Electronic Media, Volume 64, 2020 — Issue 2, 2020 https://www-tandfonline-com.libproxy1.usc.edu/doi/full/10.1080/08838151.2020.1757365
Cross “The Cost of Training Machines is becoming a Problem” The Economist, Technology Quarterly, 2020 https://www.economist.com/technology-quarterly/2020/06/11/the-cost-of-training-machines-is-becoming-a-problem
González, Ortega, Pérez-López, Alonso, “Bias and Unfairness of Collaborative Filtering Based Recommender Systems in MovieLens Dataset,” in IEEE Access, vol. 10 https://ieeexplore-ieee-org.libproxy2.usc.edu/document/9808125
Hooker. “Moving beyond “algorithmic bias is a data problem””, Patterns, Volume 2, Issue 4, 2021, https://doi.org/10.1016/j.patter.2021.100241
Kalimeris, Bhagat, Kalyanaraman, Weinsberg “Preference Amplification in Recommender Systems” Conference on Knowledge Discovery and Data Mining (KDD) 2021 https://research.facebook.com/publications/preference-amplification-in-recommender-systems/
Knijnenburg, Bart “Explaining the user experience of recommender systems.” User Modeling and User — Adapted Interaction, vol. 22, no. 4–5, 2012, pp. 441–504. ProQuest, http://libproxy.usc.edu/login?url=https://www.proquest.com/scholarly-journals/explaining-user-experience-recommender-systems/docview/1019903659/se-2
Li, Yi, Vasconcelos “REPAIR: Removing Representation Bias by Dataset Resampling.” Cornell University Library, arXiv.org, 2019. ProQuest, http://libproxy.usc.edu/login?url=https://www.proquest.com/working-papers/repair-removing-representation-bias-dataset/docview/2211209936/se-2.
Mudigere “Software-Hardware Co-design for Fast and Scalable Training of Deep Learning Recommendation Models” 2022 https://arxiv.org/pdf/2104.05158.pdf
Saveski, Gillani, Yuan, Vijayaraghavan, Roy “Perspective-taking to Reduce Affective Polarization on Social Media“ 2021 https://www-proquest-com.libproxy1.usc.edu/docview/2581624944?pq-origsite=primo&accountid=14749
Sharma, Zhang, Aliaga, Bouneffouf, Muthusamy, Varshney. “Data Augmentation for Discrimination Prevention and Bias Disambiguation.” AAAI/ACM Conference on AI 2020. https://doi-org.libproxy2.usc.edu/10.1145/3375627.3375865
Sîrbu, Alina, et al. “Algorithmic Bias Amplifies Opinion Fragmentation and Polarization: A Bounded Confidence Model.” PloS One, vol. 14, no. 3, 2019, pp. 1. ProQuest, http://libproxy.usc.edu/login?url=https://www.proquest.com/scholarly-journals/algorithmic-bias-amplifies-opinion-fragmentation/docview/2188588723/se-2
