Recommendation systems in social media platforms

From SI410
Jump to: navigation, search

The recommendation system has been regarded as one of the most prominent applications of information technology in recent years. In general, a recommendation system is able to predict the interests of users and makes information searching easier based on the huge volume of data. Though a well-established recommendation system utilizes many different technologies, there are two types of recommendation systems: content-based systems and collaborative filtering systems. The former relies on the content of the data while the latter examines the preferences of similar users to generate suggestions. Almost all of the social media apps, including Facebook, Twitter, Instagram, TikTok, etc., have implemented such systems to increase user interactions. 

However, in recent years, more and more criticisms have emerged toward the use of recommendation systems, especially in social media. The public has become more and more cautious about the contents of the messages they receive. For example, Facebook users are unsure whether the algorithms are secretly nudging them to vote for political candidates, as the Facebook Cambridge Analytica data scandal revealed. According to a recent Morning Consult poll[1], 56 percent of people in the United States want more regulatory control over social media corporations, and nearly three out of five every people believe that social media platforms do not go far enough to keep users secure. Regulators around the globe have tightened their regulations on social media platforms. For example, the European Union is drafting its own Digital Services Act that addresses the use of recommendation systems by social media platforms and requires them to provide users with an opportunity that is not based on profiling user information. [2]

Recommendation systems in social media platforms

Certain characteristics of social media platforms resulted in a successful recommendation for them that differed from those used for movies or shopping platforms. According to the paper "Social recommendation model based on user interaction in complex social networks", the authors imply that "most of the content recommendation processes in social platforms introduce content types that are created and shared by the users themselves, gather the feedback information of other users through the articulated relationships in complex networks, and implicitly infer the user preferences and content popularity."[3]

Mechanisms of the recommendation system

Data Input for the Recommendation System in Social Media

In a world awash in data, it is estimated that 2.5 quintillion bytes of data are generated per day in 2018. [4] The two most important types of data used by recommender systems are user-explained ratings and implicit ratings.[5] Explicit ratings are user reviews, ratings, feedback, likes, or dislikes, whereas implicit ratings are user interactions with the contents. Examples of implicit ratings include users’ time spent on viewing the item, whether a user clicks on an item, and etc. In general, explicit ratings are more difficult to acquire as many users do not provide contributed reviews or feedback. And oftentimes, the explicit ratings can suffer from sampling biases, including survivorship bias, meaning some users only leave reviews when they have an extremely terrible experience with a product, while most users who have average or good experiences do not leave their comments. In that case, the explicit ratings alone cannot truly reflect users’ feedback.

Euclidean Distance

How does the recommendation system make recommendations

After obtaining the data, the recommendation systems then try to calculate similarities using the ratings provided by users using linear algebra concepts. The algorithms project the data into vector space and find similarities between the vectors. There are several mathematical methods for the recommendation system to measure the similarities between product A and product B, including Minkowski distance, Manhattan distance, Euclidean distance, cosine similarity, and more.  

TF-IDF

There are also different ways for the recommendation system to make recommendations. One approach is to use the content based ratings to make suggestions. Recommendation systems of this kind search for the existing ratings of the product provided by the user and then make predictions on a new product. It then selects the product that has the highest similarities with the product that the user prefers and recommends that product to the user. In the context of social media, ratings can be in the form of likes or subscriptions. The advantage of this kind of recommendation system is that it performs well even if there are no existing customer reviews for a product, but the downside is that it is particularly tough to apply to huge product databases since users tend to have differing perspectives on each item.

Another approach uses the product descriptions to make recommendations. In description-based content recommendation systems, the algorithms try to capture as much information about the product that they are recommending as possible. They search for the title, summary, genre, and types of the product, store this information in texts, and use Natural Language Processing algorithms to calculate the similarity. In the context of social media platforms, this information consists of a post’s location, the categories of the contents in the image of the post, the text contents in the post, etc. Some popular methods for calculating similarities in these types of recommendation systems are the bag of words model (count vectorization), tf-idf vectorization, and recently emerged language models like BERT (Bidirectional Encoder Representations from Transformers) that leverage deep learning concepts like neural networks.

Collaborative Filtering

Some companies also utilize a concept called "collaborative filtering" to construct their recommendation systems. Collaborative filtering systems provide recommendations based only on how previous users evaluated the product (the contents of the post on social media), not on the product itself. The algorithms only require data that shows how the product was evaluated by previous consumers. In this case, the algorithms measure the similarity of users instead of the similarity of products. The researchers who authored "Social recommendation model based on user interaction in complex social networks" summarize the concept of the algorithms concisely: "the rating information of other users in social networks to find neighbors who have similar tastes to the target user and then recommend items to them." One reason social media is more likely to use collaborative filtering techniques is that they can help with some of the problems with content-based recommendations when it's hard to get item-level content information.


Facebook and Instagram’s policies on their recommendation systems

Facebook and Instagram have provided detailed descriptions of how Facebook and Instagram use their recommendation systems. They provided a webpage that lists in what areas of their applications they show recommendations, their baseline standard for using recommendation systems, what types of content are allowed and what are not allowed to be recommended, and where to find insights about the recommendations Facebook or Instagram provide. Links to Facebook and Instagram's descriptions on their policies: What are the recommendations on Facebook? What are the recommendations on Instagram?.  

Currently, Facebook and Instagram use recommendation systems on areas including Pages You May Like, "Suggested For You" posts in the News Feed, People You May Know, or Groups You Should Join. They claim that with the help of over 50 industry experts specializing in recommendation systems, expression, safety, and digital rights, they have established a strategy called "remove, reduce, and inform" since 2016 in order to manage problematic content on Facebook and Instagram. According to Facebook, "this strategy involves removing content that violates our community standards, reducing the spread of problematic content that does not violate our standards, and informing people with additional information so they can choose what to click, read, or share."


Criticism of the recommendation system on social media platforms

Instagram and young adult depression

In recent years, social media has faced increasing scrutiny from the public and regulators due to those platforms' implicit ethical issues. Many believe social media companies are misusing or abusing their technologies in order to make more profit from advertisers. Because social media platforms often use the recommendation system to get people to spend more time on their sites, the algorithms also play a role in the severity of mental health problems.

Researchers from KU Leuven found that the increases in Instagram browsing time and posting frequency are "related to adolescents’ depressed mood".[6]

What’s more, according to an article on wired, content that promotes self harm still gets recommended to Instagram users. "Despite clear rules against any content that promotes self-harm, and despite blocking specific hashtags to make that content harder to find, social media platforms continue to serve this content up algorithmically." In 2017, a British teenager, Molly Rusell, searched for suicide and self harm related topics on Instagram before ending her own life. Instagram’s recommendation systems keep promoting self harm related content on her phone, and her father has accused Instagram of assisting in Molly's death by permitting these violent photos on their platforms and pushing them onto her feed. [7]

Internal researchers at Instagram, now acquired by Facebook, were also aware of this issue despite the company's refusal to admit it in public. In May 2021, Instagram head Adam Mosseri told reporters that research he had seen suggests the app’s effects on teen well-being are likely "quite small."[8] But the documents leaked from its internal research findings suggest that the company is well aware of the tremendous negative impacts brought by its algorithms. The slides for the internal meeting suggest that "one in five teens say Instagram makes them feel worse about themselves, and teens who struggle with mental health issues say Instagram makes them feel worse." "Inappropriate advertisements targeted at vulnerable groups" were among the top reasons that Instagram negatively affects mental health. Instagram's algorithms and recommendation system are alleged to have amplified teenagers' social pressures, especially on physical appearance. For example, a Wall Street Journal report documents an interview with a teenager, Lindsay Dubin, 19, who recently wanted to exercise more. "She searched Instagram for workouts and found some she liked. Since then, the app’s algorithm has filled her Explore page with photos of how to lose weight, the "ideal" body type, and what she should and shouldn’t be eating. "I’m pounded with it every time I go on Instagram," she said."[9] In a subsequent experiment, Lindsay found that 14 ads focused on physical appearances appeared in two-minute browsing of Instagram stories, a feature on Instagram that lets users view short videos posted by their friends along with advertisements displayed based on the users' interests.

Although some researchers have suggested that recommendation systems can also be used to improve mental health situations by "providing inspiration and motivation for planning and engaging in more pleasant activities," certain constraints related to current ethical issues are still difficult to address. Researchers found that the suggestions provided by recommendation often lack explainability as to how they are recommended. It is also a complex issue to deal with the privacy/personalization trade-off and recommendation quality. And thirdly, users have no control over how companies can use their data in other areas. [10]


Criticism of Facebook’s recommendation system policy

While Facebook has published these detailed standards for what content to recommend to its users, critics have suggested that the documentation lacks deep insight into what kind of data Facebook and Instagram use to feed into the algorithm and how they actually choose what to recommend to a given user. "That’s a key piece to understanding recommendation technology, and one Facebook intentionally left out." [11]

Besides, many critics doubt how effectively Facebook can adhere to its own stated rules. It is not uncommon for users to click on recommendations from Facebook and find conspiracy theories, harmful health information, and COVID-19 falsehoods, where the guidelines were supposed to prevent these recommendations from appearing. According to a report by NBC, Qanon, an American far-right political conspiracy theory, mass political movement, and cult, quickly gained popularity due to Facebook recommendations. [12]

Aside from militia groups and conspiracies, a study claims that Facebook has failed to prevent the spread of misinformation by prominent anti-vaccine groups, according to a study from ISD Global, a U.K.-based think tank that studies polarization, extremism, and misinformation. [13] Researchers suggest that "Facebook failed to implement its own policies at a very basic level." For example, the report outlines some of the false claims from members of the World Doctors Alliance group, "a collective of pseudo-science influencers... that hijacked the pandemic to build up a significant audience online in a multitude of languages in multiple continents" that were permitted on Facebook, from statements that the COVID-19 virus does not exist to those confirming its presence but minimizing its seriousness.

Unethical use of data for the recommendation system: Facebook Cambridge Analytica data scandal

How does Cambridge Analytica's Algorithm work?

The parent company of Instagram, Facebook, has faced more serious backlash for the algorithms embedded in its app since 2016. In the 2010s, Facebook cooperated with Cambridge Analytica, a British consulting business, to conduct political advertising gatherings. It was reported that Cambridge Analytica had made use of the information of 50 million Facebook users without their consent. [14] Cambridge Analytica’s algorithms are alleged to have made a lot of extrapolations from information about users' behaviors on Facebook, including Facebook likes among initial app users, Facebook likes among friends of app users and others. Cambridge Analytica combined the data with the personality and political test results and said that using this data it was capable of penetrating voters' deepest psyches and eliciting their underlying anxieties and wants. On March 17, 2018, the scandal surfaced when Christopher Wylie, the former Director of Research for Cambridge Analytica, became the whistle blower and was interviewed by the New York Times and The Observer. Facebook’s unethical use of data led to a $5 billion penalty imposed by the FTC and new privacy restrictions on the use of data. [15]


Critics’ Suggestions on Improving the recommendation Systems’ Transparency

As provided by Article 19, an international human rights organization that works to defend and promote freedom of expression and freedom of information worldwide, provides several suggestions on how to improve the transparency of the recommendation system to better protect the users. In its article "EU: Regulation of recommender systems in the Digital Services Act", they advise regulators to require very large online platforms (VLOPs) to provide more diversified exposure, more user choice in recommender systems, and the unbundling of hosting and content curation so as to create a more equitable online environment. [16]

In the article, they believe that current recommendation systems have some serious unaddressed flaws. Article 19 first points out that VLOPs’ business model relies on profiling users and offer personalized content. In order to maximize user engagement and profit, VLOPs’ recommender systems pick content based on a variety of characteristics. As a result, companies have little motivation to expose their consumers to a wide variety of information but rather to the posts that most interest them. Thus, Article 19 suggests that promoting alternative voices to users is important in protecting their usage of social media.

Another problem is that users often lack access to their data and the transparency of the model that generates recommendations. In the example of Facebook and Instagram, these companies are very reticent in terms of the details of their recommendation systems. This is because the quality of their algorithms is often viewed as the key differentiator that sets companies apart from their competitors. Revealing any information about the details of the algorithms will render the latest algorithms exploited by their competitors, and many years of investment wasted in vain. Furthermore, users are frequently unaware of social media's use of their data, and the power imbalance makes it difficult to stand up to large corporations." Concerns have repeatedly been raised about these systems’ tendency to promote clickbait, sensationalist, false, or "extremist" content, often pushing users down "rabbit holes" of this type of content without their knowledge or consent. And users have no power to adjust what kind of content they see. Thus, Article 19 suggests that all platforms should be required to "provide clear and accessible information on how their recommendation systems are used to present, rank, promote, or demote content." An audit should be conducted by a third party to ensure the accuracy and transparency of the algorithms.

Article 19 also suggests that VLOPs should unbundle the hosting from content curation, and allow third parties to provide alternative recommendation systems. This solution will allow users to choose from a variety of recommended systems. A more competitive market with a varied range of recommendation systems can benefit users by increasing their access to a wider range of content and allowing them to have a higher degree of control over what they will see on social media.

Article 19 provides a less bold but more realistic approach, "requiring all platforms to provide options for users to modify the parameters under which content is shown to them, with recommender systems that are not based on profiling set as the default option," given that the approach will drastically change the landscape of social media.
  1. https://assets.morningconsult.com/wp-uploads/2021/10/18135638/2110047_crosstabs_MC_TECH_FACEBOOK_Adults_v1_CC.pdf
  2. https://www.article19.org/resources/eu-regulation-of-recommender-systems-in-the-digital-services-act/
  3. (Li et al., "Social recommendation model based on user interaction in complex social networks, 2022")</
  4. https://www.forbes.com/sites/bernardmarr/2018/05/21/how-much-data-do-we-create-every-day-the-mind-blowing-stats-everyone-should-read/?sh=3c2057f360ba
  5. https://builtin.com/data-science/recommender-systems
  6. Frison, E., & Eggermont, S. (2017). Browsing, posting, and liking on Instagram: The reciprocal relationships between different types of Instagram use and adolescents' depressed mood. 20 (10), 603-609.
  7. https://www.wired.com/story/when-algorithms-think-you-want-to-die/
  8. Jeff Horwitz and Deepa Seetharaman, G. W., & B., 2021, September 14 Facebook Knows Instagram Is Toxic for Teen Girls, Company Documents Show. WSJ. https://www.wsj.com/articles/facebook-knows-instagram-is-toxic-for-teen-girls-company-documents-show-11631620739.
  9. Jeff Horwitz and Deepa Seetharaman, G. W., & B., 2021, September 14. Facebook Knows Instagram Is Toxic for Teen Girls, Company Documents Show. WSJ. https://www.wsj.com/articles/facebook-knows-instagram-is-toxic-for-teen-girls-company-documents-show-11631620739.
  10. https://link.springer.com/article/10.1007/s00146-021-01322-w
  11. https://techcrunch.com/2020/08/31/facebook-partially-documents-its-content-recommendation-system/
  12. https://www.nbcnews.com/tech/tech-news/qanon-groups-have-millions-members-facebook-documents-show-n1236317
  13. https://abcnews.go.com/Technology/facebook-failing-tackle-covid-19-misinformation-posted-prominent/story?id=81451479
  14. Rosenberg, Matthew; Confessore, Nicholas; and Cadwalladr, Carole, March 17, 2018). "How Trump Consultants Exploited the Facebook Data of Millions". The New York Times. (Archived from the original on March 17, 2018.)
  15. https://www.ftc.gov/news-events/press-releases/2019/07/ftc-imposes-5-billion-penalty-sweeping-new-privacy-restrictions
  16. https://www.article19.org/resources/eu-regulation-of-recommender-systems-in-the-digital-services-act/