Written by Devyn Hinkle

What is an algorithm?

An algorithm includes the steps or parameters used to solve problems or perform calculations. Algorithms perform integral functions in our everyday lives. While most of the time algorithms can benefit our lives by making our decision making processes easier, it is not as simple as that. Algorithms are not static. Instead, most algorithms use machine learning, statistics based on gathered data, and AI to make them more advanced or specified. They are constantly evolving, and as useful as they are, they are missing one key factor: common sense. Even still, most people cannot use the internet without using an algorithm. They are in our search engines, emails, apps, video games, dating services, travel sites, GPS, and our streaming services.

Algorithms in streaming

As streaming platforms have emerged as leading sources for entertainment media — particularly for film, television, and music — access to content has increased to an overwhelming amount in many cases. In December of 2015, consumer research from Netflix estimated that a subscriber loses interest after 60 to 90 seconds of browsing before they choose something or abandon the streaming platform. This is where algorithms step in.

Actor Marc Maron theorizes that everyone had one person in their life that influenced their interests by introducing them to new films, music, books, clothes, or anything else. This person came to us at an influential time and changed the trajectory of our lives. One could argue that that role is now filled by algorithms.

Amazon has been using recommendations from algorithms for its service since 1998. Netflix has been using algorithms for recommending entertainment since 2007 when it was only sending DVDs. Now, nearly every major streaming platform uses its own unique algorithm that combines human knowledge and machine learning to create a decision making process that will guide the viewer. To do this, they use algorithms. All of the streaming platforms are competing against each other for subscribers, hours of view time, and notoriety. Algorithms are part of that. Netflix and other platforms like Hulu, Disney+, and HBO Max are looking to reduce their churn rate (the number of subscribers who cancel the service within a certain amount of time), increase viewing time, and keep viewers’ interest. However, how each platform does this is different. This article will break down these four companies’ approaches to algorithms as well as discuss the implications of algorithms on streaming sites.

The basics of streaming algorithms

Most of the recommendation systems can be sorted into three different types:

1. A Content Based algorithm uses the attributes of an item, such as its metadata, tags, or text, to make recommendations that are similar to items the user has previously interacted with.

2. A Collaborative Filtering method uses the interests and behaviors of other users with similar tastes to make recommendations.

3. A Knowledge Based method uses the attributes of an item and correlates them with the preferences of a user to make recommendations based on similarities. This method is the foremost for discovering new content because it isn’t based on past behaviors.

Many streaming platforms use a hybrid of these three systems. For example, knowledge based systems work well for new users that don’t have any history on the platform while other systems may work better for other types of users.

Netflix

Netflix is the most transparent about its algorithms. In fact, in 2006 Netflix hosted a competition with a $1 million prize for a recommender system that was better than its current system. While one team was able to create such a system, it’s unclear if Netflix implemented it. Regardless, in 2007, Netflix pivoted from its original DVD delivery model to streaming, changing its entire system of recommendations. Then, in 2014 Netflix, invested 3% of its revenue at the time to develop a recommendation engine that eventually included data from all countries and could be altered in real time.

In 2015, Netflix stated that choices made based on the recommender system account for around 80% of the hours streamed. The other 20% are the result of searches that utilize an entirely different set of algorithms. For its algorithm, Netflix uses data on what users watch, what they search, how they rate, when they watch, and more to customize page construction, genre rows, trending videos, order of videos, and even the icons. Netflix claims that these systems have decreased its churn by several percentage points and have saved $1 billion per year.

One of the things that makes Netflix’s page layout so interesting is its very unique genre titles like “Reluctant Adults” or “Gritty Movies” (real genres currently featured on my Netflix page). These are what Netflix considers “altgenres.” According to an article published by The Atlantic, in 2014, Netflix had nearly 77,000 altgenres. These altgenres were created through a combination of machine learning and human intellect and are part of how Netflix tailors its recommendations through algorithms.

Netflix has published much of the details on its algorithms. The first use of algorithm series is in the matrix-like layout and content of each user’s homepage:

• The Personalized Video Ranker (PVR) takes the catalog of videos and orders them for each member within the genre rows.

• The Top-N Video Ranker generates recommendations for the Top Picks row, personalizing the selection for a member.

• The Trending Now is similar to Top-N in that it focuses on trends, but this one particularly focuses on short-term trends like Valentine’s Day or a weather event while still including consideration of personalization.

• The Continue Watching sorts recently viewed content using an estimate of what a user is most likely to continue watching or re-watch.

• The Video-Video Similarity or Because You Watched develops recommendations based on a piece of content the user interacted with. These recommendations are not personalized, but the titles that the Because You Watched features are personalized depending on estimations of a user’s preference for content.

• Finally, the Page Generation algorithm uses all of the algorithms above to personalize which rows show up and in what order.

The other uses of algorithms are:

• The Evidence algorithm helps users decide if the content they have clicked on is what they want. These metrics include predicted star rating, the information on the content, and the icons shown for content in rows. Examples for this include whether to show if the content has won an award or which icon the user will most likely choose.

• The Search is made up of multiple algorithms that help determine which results are most relevant for a user by including partial queries (“frie” showing “friend”) or concepts (showing French movies for “French” instead of titles including “French”).

While the algorithms alone are impressive, Netflix also runs about 250 A/B tests every year on around 100,000 users. The goal is to not rely on the algorithms but to use real data to figure out what is working. More recently, it has also incorporated the data from its new addition, Netflix Party, a browser extension that allows users to watch with other users. The data from this allows them to expand the knowledge of a user’s preference outside of what they are watching.

Netflix’s impressive algorithm system is only one of the many ways streaming platforms are developing personalized recommendations.

HBO Max

While Netflix’s recommendation system focuses mainly on machine learning and algorithms, HBO Max has tried to take an alternate approach that utilizes a hybrid of algorithm and human curated content but with a focus on that human touch.

HBO Max launched over a decade and a half later than Netflix’s streaming platform, in May 2020. However, it cemented its competitive advantage beyond content in its recommendation system. While HBO Max still includes algorithms and machine learning recommendations (albeit much less documented than Netflix’s) they are quick to highlight the "pockets of the platform" that engage human curated recommendations. It intends to have both HBO Max employees and celebrities alike creating lists for users. Like Netflix, its goal is to be as personalized as possible but in a way that is not so isolated.

HBO Max’s Senior Vice President of Product, Sarah Lyons, believes this advantage is as important as the actual content. In the future, HBO Max hopes to include even more human-focused curation by connecting users with other human recommendations like friend-to-friend recommendations. The Verge equates its practices with a model closer to Spotify than Netflix where the shared experience is the focus.

Hulu

In late 2020, Hulu redesigned its experience to follow a layout more similar to Netflix with horizontal, scrollable rows. With the redesign, it turned its recommendation system to personalize the rows with either a human-curated selection, an algorithm selection, or a combination. Hulu focuses on algorithm selections for its search engine and general recommendations but on human curation for the trending areas. The platform seems to be taking the happy medium, looking for a balance between human and algorithmic curation.

Other platforms

HBO Max is pushing its human curation, Netflix has its algorithms, Hulu is a hybrid, and even still, other platforms have their own strategies. Here are some of the other platforms that stand out in their recommendation systems. 

Disney+ launched only half a year before HBO Max. However, its focus is very much on algorithms and Natural Language Processing (understanding the content and its emotions). It is focusing on not only the history, but how people are using the platform. For example, it studies the sequence in which people are interacting with the platform, what they watch multiple times, and what they don’t click on. The goal is to create a more engaging platform to help users understand what is available.

Amazon Prime centers on its algorithms and machine learning from its other services. It mainly uses a collaborative filtering method that looks at what other users are doing.

The Criterion Channel is perhaps the most off-the-cuff all the streaming platforms. It has strayed away from algorithms entirely to focus on showing what it believes is important and not something it believes a user is necessarily interested in.

The downfalls of algorithms

Algorithms may seem like the perfect solution for recommending users content that they will enjoy. However, there is a reason platforms like HBO Max and Criterion are veering away from algorithm-only recommendations.

Recently, famed director Martin Scorsese came forward to voice his concerns that streaming platforms are devaluing the art of film to “content” because of algorithms. While he has nothing against platforms like Netflix (in fact, he is in business with Netflix), he sees the suggestions as limiting, only encouraging the user to watch the same kinds of things. One example he brings up is if you watch something by accident or if someone else picks the movie for a night, you are suddenly being fed titles you have no interest in. He believes that people should amount to more than their data.

Scorsese’s fears may be based on his interest in maintaining the art of film, but they actually allude to a larger issue with recommendation algorithms: feedback dynamics. Algorithms are inherently built to become more and more specialized to focus on personalization. For example, when Netflix first began recommending things, the algorithms were based on the data it had, which wasn’t much. Now, Netflix’s algorithms have years of user data to build its recommendations on.

While this sounds ideal, it creates Algorithmic Amplification and some content is pushed more at the expense of other content, narrowing viewpoints and creating an echo chamber. There is an optimized level where the quality of recommendations are the strongest. Once  you pass that, you enter into a Filter Bubble where users are only seeing content that aligns with what they’ve already seen, not allowing them to discover or explore new interests. Feedback loops then take it a step further, causing the output to be the input. What that means is that the algorithm is learning from the user’s choices and the user’s choices are based on what the algorithm is recommending to the user. For example, if I watch a lot of Romantic Comedies from the year 2015, the algorithm would feed me more Romantic Comedies from 2015 and I would therefore chose a Romantic Comedy from 2015. Now the algorithm learned that I like Romantic Comedies from 2015, but that was just what it originally showed me. The Feedback Loop only increases the effects of Filter Bubbles, as users are encouraged to behave within their bubble and the bubbles start to shrink.

The other issue platforms face with algorithms is their inherent bias. As Seth Godin mentions, algorithms cannot be neutral. They are built on human data and pull habits or instincts from individuals to further develop the process. Netflix has received criticism for gearing racially inclined images towards certain users. While Netflix does not use demographics in its algorithm, the algorithm can be biased and make inferences. 

Marcus du Sautoy is a British mathematician whose studies focus on algorithms and AI. One of the questions he asks in his book The Creativity Code is whether “AI can complete creative tasks better than humans.” While recommendation system algorithms are not inherently creative, the task of asking them to continually transform and learn from themselves is. What he addresses is the idea that when the algorithm is fed faulty data, it creates a faulty algorithm, particularly with human biases or data that can “trick” an algorithm.

In streaming, we return to the idea that no human’s preference is linear. If I am truly interested in Romantic Comedies made in 2015, but one day decide to watch a horror movie with my friend, have I fed the algorithm faulty data? If others are engaged in similar actions, are they changing my algorithm? Furthermore, how do we avoid the bias issues Netflix has run into, inherent by inferences based on preferences?

Sautoy concludes that AI can assist in creativity, but is not able to be creative itself. Recommendation system algorithms are the same: they can help us come to a decision, but they do not make the decision themselves. However, we need to be aware that in order to allow a user to be “creative” in their decisions, we can’t allow the algorithms to spiral into Feedback Loops perpetuating AI’s lack of creativity.

Netflix, in particular, is at risk of this with its interacting algorithms all based on personalizing the experience. The overarching issue with the Feedback Loop is the recommendation systems not allowing users to broaden and discover new content. The concurrent problem is that users aren’t normally inclined to choose something new. So, it is hard to say if the human curation models HBO Max and even Hulu are using will break the feedback loops. However, HBO Max’s idea of including peer suggestions or attaching celebrity names to these curations should help. People need human suggestions to nudge them towards  something new because users tastes are not simple enough for an algorithm to interpret on its own.

In a 2018 study, nearly 70% of respondents claimed that it is too hard to find content on streaming services or that their service continually recommends the same content. Only 21% agreed that their services know what they want to watch better than they do. While the study is a few years old, it illustrates the inherent issue with algorithms on services. There will always be a level of dissatisfaction in recommendation systems when content is so readily available.

Streaming services should not only focus on pleasing their customers with personalized recommendations but also on how to provide that personalization without pigeon-holing the user. A user’s tastes are not as simple as an algorithm makes them out to be no matter how complex the algorithm is. Platforms need to find ways to mimic the way people discover new interests offline, with suggestions from friends, influencers, or even just happy accidents. HBO Max is on its way to including some of these options, but there needs to continually be an effort to recognize the isolation a personalized screen creates.

Furthermore, Deloitte estimates that users increased the number of streaming services they subscribe to from three before Covid-19 lockdowns to five post Covid-19 lockdowns. While algorithms within services are extremely important, users are beginning to face the issue of how to pick what to watch across multiple platforms. In the future, individual streaming service platforms and maybe streaming players like Roku need to start exploring how to create algorithms or recommendation systems across multiple services.

The future will include algorithms and machine learning, but as platforms proceed, they need to recognize the downfalls of algorithms. If platforms do that, we as users need to pay attention to how we are utilizing recommendation systems.