When YouTube recommends videos that match our current interests, or when Amazon suggests products we might find intriguing, what mechanisms are at play? Recommendation algorithms. These are highly intricate systems developed to further personalize the user experience, though with the potential risk of producing sometimes undesirable effects of polarization. They also spark debates regarding the extensive cross-referencing of our personal data from various sources...
As companies like Amazon, Google, or Facebook ventured into the internet sector and established their dominance where tens of thousands of startups had faltered, they encountered a specific challenge: How to keep a visitor engaged? How do they ensure that an individual is enticed to revisit, and in the case of news or streaming platforms, to remain engaged for as long as possible?
In navigating this uncharted territory, the tech giants devised innovative strategies to captivate visitors through the use of recommendation algorithms.
These algorithms were formulated with several goals in mind:
- to make the user experience as enjoyable as possible;
- to present visitors with content that aligns closely with their interests;
- to improve the performance of various key metrics (such as video viewing time, reading duration, average shopping cart value, etc.);
- to filter content in a manner that is customized for each individual.
The crucial keyword here is: personalization. For a media company like Facebook, achieving this is quite remarkable. Over two billion users experience a unique news feed that is constantly refreshed every second.
Consider YouTube as another example. As the leading online video platform, it houses billions of videos and is updated with new content daily. Nonetheless, upon logging into YouTube, you are presented with merely a few dozen video recommendations, a selection that changes with each visit but consistently – albeit not always subtly – aims to provide a lineup that will grab your attention. How does YouTube manage to offer content that makes you eager to return and explore more? Primarily by merging topics that you find interesting at any given moment with those of other users who share a similar profile to yours.
All the major digital entities quickly realized the potential of tapping into the vast data reservoir of internet users. Today, these recommendation algorithms are employed by a large array of websites. They are geared towards identifying the most suitable new content at any given time. Among the most familiar examples are:
- the post or friend suggestions on Facebook, Instagram, or Twitter;
- the products recommended by Amazon or Alibaba;
- the video clips suggested by YouTube or TikTok;
- the featured articles on news portals;
- the travel destinations and activities proposed by sites like Booking and TripAdvisor.
The success of this approach is undeniable. So much so, that at Netflix, 80% of the films watched the most originate from recommendation algorithms.
How do recommendation algorithms work?
How do these internet behemoths fine-tune their approach to target each of us with increasing accuracy? Their recommendation algorithms utilize filtering methods to recognize patterns. The principal strategies they employ are as follows.
Collaborative Filtering
Collaborative filtering stands as one of the most employed and effective recommendation algorithms. It operates on the theory that if two individuals have appreciated similar content in the past, they will likely enjoy the same content in the future.
The strength of collaborative filtering lies in its complete lack of directiveness. It relies entirely on the user’s history. Yet, this method has the drawback of diminishing the diversity of content accessible to the user. Users end up being exposed not to a variety of perspectives but predominantly to information that echoes their prevailing beliefs.
Content-based Filtering
The content-based technique analyses a collection of content whilst disregarding the preferences of other users. It concentrates on similarities to recommend content. The essence of content is determined by cataloging key words and then comparing these to the words in other pieces of content.
Popularity
This algorithm presumes that if you frequently visit a specific website, you are likely to enjoy the content that attracts the most visitors. In essence, it recommends the most popular content. An advantage of this technique is that it can be applied to new users of a site.
The risks of polarization
While recommendation algorithms provide numerous benefits to users through personalized recommendations that are frequently relevant, they can also present several disadvantages, some of which have societal implications.
For example, there has been a notable increase in the number of individuals expressing extreme views on subjects such as politics or climate change. This phenomenon of “polarization” can appear dangerous, as it has the potential to erode critical thinking or, more simply, the ability to explore beyond conventional paths.
Similarly, SACEM disclosed that 99% of the most streamed tracks on Spotify pertain to just 10% of the catalog. Even more concerning is that 20% of the tracks are never suggested to listeners.
