Recommendation algorithms

Recommendation is where the role of algorithms is most visible to the public. Netflix, for instance, had a highly publicized contest to improve its recommendation algorithms,[1] while video creators on platforms such as YouTube and TikTok frequently articulate an idea of "the algorithm" as a combination agent, gatekeeper and antagonist.[2]^, [3] The lack of physical constraints on the size of their catalogues means that streaming platforms may host thousands, millions or even hundreds of millions of items; while this is one of their selling points, it can also contribute to choice paralysis. Platforms have found that audiences have little tolerance for this: Netflix, for example, reports that if a viewer hasn’t selected what to watch within 90 seconds of their search, they’re likely to leave the platform.[4] To prevent people from leaving, recommendation systems typically use at least two different algorithms: one to filter out content unlikely to appeal to the consumer and another to prioritize the remaining content to make items that are most likely to appeal to users the most prominent. As we’ll see, at this stage in the process, the recommendation system may determine what content is offered and how it’s presented.

 Some streaming services, like Disney+, use recommendation systems that rely primarily on data about the content itself – prioritizing what’s new and popular, for example. Most, however, incorporate data that they’ve collected about users to find associations between the content and different users. This is done in three ways (or any combination of the three):

1. By matching explicit attributes of an item – tags, cast members, genre, etc. – with items that the user has enjoyed in the past (a content-based approach).
2. By surveying users on their preferences and then matching items to those answers (a knowledge-based approach).
3. By identifying similar users and then recommending items that they liked (a collaborative filtering approach).[5]

Whichever approach (or combination of approaches) is used, it requires a great deal of data about both the content and the users to be successful. The difficulty in making successful recommendations for new users is known as the cold-start problem, and it’s one that different platforms resolve in different ways.[6] In general, though, it’s primarily addressed by collecting user data as quickly and accurately as possible, both explicitly (such as by tagging items or asking users to rate them) and implicitly (for instance, by recording whether or not a user finished watching an item, as a proxy of whether or not they liked it).

What viewers and creators refer to as "the algorithm" is actually many different systems. Association and classification algorithms are used to analyze user preferences, tag videos, and then recommend videos to users on that basis. User interactions are then evaluated based mostly on whether the user 'liked' the video, whether they commented on the video, how long each session lasted and how soon the user returned to the app after leaving. Those evaluations are used to further train the recommendation model for that user.[7] Finally, the entire cycle can operate as a feedback algorithm, as the insights gleaned about the audience can be applied to decisions about what content the platform should create, commission or acquire.

While digital media delivery increases the potential for diversity by making it possible for platforms to offer a near-limitless catalogue and lowering the cost for the consumer of trying something new,[8] there is evidence that algorithmic delivery actually reduces it. In general, streaming television services do better than American broadcast and cable television at representing Latinx, LGBTQ2S+, and Indigenous actors.[9] Netflix, in particular, outperforms broadcast television in terms of gender diversity, with 30% of showrunners, 31% of directors and 52% of major characters being women (compared to 22%, 19% and 45% for American broadcast TV, respectively)[10] and its films have a higher proportion of women and members of under-represented groups than those from Hollywood studios.[11] However, it isn’t clear whether these effects occur due to streaming services' algorithms, independently of them or despite them. The increased diversity on Spotify may be more attributable to its essentially infinite catalogue or to the ability of listeners to curate and share their own playlists, which allow songs and artists from diverse communities to become sufficiently popular for the service's algorithms to start recommending them.[12]

Data on diversity in open streaming services, such as YouTube or TikTok, is mixed. Canadian research on YouTube found that a significant majority (70.5%) value seeing diversity represented there, and almost half (48.2%) report seeing better diversity representation on YouTube than on traditional media. The same study found that "visible minorities" (the term used by the report's authors, following Statistics Canada usage) were represented more or less at par with their representation in the Canadian population among YouTube posters, while Indigenous peoples, women and people with disabilities were significantly under-represented.[13]

There is evidence that users posting content to social media feel pressure to conform to the recommendation algorithm's desires, or at least their understanding of them. While this is often articulated in terms of taking part in an adversarial system – "hacking the algorithm" – if there’s a mismatch between the user's identity and how the algorithm has classified their content, it will either marginalize them or minimize their diversity as they become more successful. For example, a woman who posts science-related videos to YouTube expressed her frustration that as her channel grew, her audience shifted from being equally male and female to predominantly male.[14]

Research also consistently finds that creators feel pressure to make themselves more visible to the algorithm: "This logic shapes the topics discussed in videos, genres engaged with, video lengths, titles utilized, video thumbnail design, and organization of speech." But if the content is oriented too much towards the algorithm, it runs the risk of being "clickbait" and facing backlash from viewers.[15] Creators must walk a fine line between catering to (what they perceive to be) the algorithm's preferences without being seen as trying too hard to do so – a line that is much easier to walk if what they produce is already favoured by the algorithm.

Creators are also pressured to produce and publish content more often to retain their algorithmic ranking. As one YouTube contributor noted, "the algorithm forces you to constantly produce content. So, you can't be like, I'm going to do a short film and take a break for like a month and a half because short films take time. You can't do that. You are going to lose hundreds of thousands of followers, and you are not going to make money." The same research found that YouTube creators favoured design changes that would minimize the “Matthew effect” and promote serendipity: specifically, letting viewers see what their friends are watching, showing viewers content they wouldn’t otherwise click, promoting more human-curated recommendations and promoting less popular creators.[16]

There are two ways to improve what recommendation algorithms show you: curating your feed[17] and training the AI.[18]

Curating means finding sources that you know you like and can count on. To get reliable information, find and follow sources that have knowledge or expertise on the topic, a process for verifying and correcting information, a motivation to be accurate, and that aim to be objective or are transparent about their point of view.

Training your algorithm means sending signals that show what you want (and don’t want) to see.

See this tip sheet for more specific steps on how to train your algorithm.

[1] Bennett, J., & Lanning, S. (2007). The Netflix prize. In Proceedings of KDD Cup and Workshop. Retrieved from: https://www.cs.uic.edu/~liub/KDD-cup-2007/NetflixPrize-description.pdf

[2] Hinkle, D. (2021). How streaming services use algorithms. Arts Management & Technology Laboratory. Retrieved from: https://amt-lab.org/blog/2021/8/algorithms-in-streaming-services

[3] Pedersen, E. (2019). “My videos are at the mercy of the YouTube algorithm”: How content creators craft algorithmic personas and perceive the algorithm that dictates their work. [Master's Thesis]. Department of Electrical Engineering and Computer Science. University of California at Berkeley. Retrieved from: https://digitalassets.lib.berkeley.edu/techreports/ucb/text/EECS-2019-48.pdf

[4] Gomez-Uribe, C. A., & Hunt, N. (2015). The Netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems (TMIS), 6(4), 1-19.

[5] Hinkle, D. (2021). How streaming services use algorithms. Arts Management & Technology Laboratory. Retrieved from: https://amt-lab.org/blog/2021/8/algorithms-in-streaming-services

[6] Bonini, T., & Gandini, A. (2019). “First week is editorial, second week is algorithmic”: Platform gatekeepers and the platformization of music curation. Social Media + Society. https://doi.org/10.1177/2056305119880006

[7] Smith, B. (2021). How TikTok reads your mind. The New York Times. Retrieved from: https://www.nytimes.com/2021/12/05/business/media/tiktok-algorithm.html

[8] Tan, T. F., Netessine, S., & Hitt, L. (2017). Is Tom Cruise threatened? An empirical study of the impact of product variety on demand concentration. Information Systems Research, 28(3), 643-660.

[9] Nielsen. (2020). Being seen on screen: Diverse representation and inclusion on TV. Retrieved from: https://www.nielsen.com/us/en/insights/report/2020/being-seen-on-screen-diverse-representation-and-inclusion-on-tv/

[10] Hailu, S. (2021). Streamers put more women in charge of TV shows than broadcast networks, study finds. Variety. Retrieved from: https://variety.com/2021/tv/news/boxed-in-study-2021-streaming-networks-1235063810/

[11] Smith, S. L., Pieper, K., Choueiti, M., Yao, K., Case, A., Hernandez, K., & Moore, Z. (2021). Inclusion in Netflix original US scripted series & films. INDICATOR, 46, 50-6.

[12] Dhaenens, F., & Burgess, J. (2019). ‘Press play for pride’: The cultural logics of LGBTQ-themed playlists on Spotify. New Media & Society, 21(6), 1192-1211.

[13] Berkowitz, I.S., Davis, C., & Smith H. (2019). Watchtime Canada: How YouTube Connects Creators and Consumers. Ryerson University. Faculty of Communication & Design. Retrieved from: https://audiencelab.fcad.ryerson.ca/wp-content/uploads/2019/05/YouTube-Full-Report-FINAL_V7_May21.pdf

[14] Bishop, S. (2020). Algorithmic experts: Selling algorithmic lore on YouTube. Social Media + Society, 6(1), 2056305119897323.

[15] Bishop, S. (2020). Algorithmic experts: Selling algorithmic lore on YouTube. Social Media+ Society, 6(1), 2056305119897323.

[16] Pedersen, E. (2019). “My videos are at the mercy of the YouTube algorithm”: How content creators craft algorithmic personas and perceive the algorithm that dictates their work. [Master's Thesis]. Department of Electrical Engineering and Computer Science. University of California at Berkeley. Retrieved from: https://digitalassets.lib.berkeley.edu/techreports/ucb/text/EECS-2019-48.pdf

[17] Frau-Meigs, D. (2024). Algorithm literacy as a subset of media and information literacy: Competences and design considerations. Digital, 4(2), 512-528.

[18] de Groot, T., de Haan, M., & van Dijken, M. (2023). Learning in and about a filtered universe: young people’s awareness and control of algorithms in social media. Learning, Media and Technology, 48(4), 701-713.