Running on Netflix: How Machine Learning is Fueling Your Netflix Binge-Watching Problem

In an era of decreased attention span, heightened competition and high content creation costs, Netflix is countering churn with innovative machine-learning algorithms which aim to provide subscribers with the most personalized entertainment experience but would this be enough to maintain a sustainable competitive position in the long run?

Netflix’s combats its main threat, subscriber churn, by creating a personalized experience that customers are willing to pay for. This strategy has allowed the company to become the largest streaming company in the world with 137 million subscribers throughout 190 countries providing them with a library of 14,835 worldwide[1] (5,579 in the US)[2] original and licensed titles. But how can Netflix provide a personalized streaming experience to its millions of subscribers with its vast library? The answer lies on the introduction of machine learning to the company’s innovation and improvement processes.

Machine learning has been instrumental at creating a competitive edge for Netflix on two main avenues: (1) content suggestion and (2) content acquisition and creation.

(1) Content Suggestion

Two years ago, Netflix started experimenting with sophisticated supervised (e.g. classification and regression) and unsupervised (e.g. clustering and compression) machine learning algorithms[3] that aimed at aggregating large sets of data and identifying patterns in consumer behavior to improve user experience while decreasing monthly churn rate (lowest in the streaming industry at 9%/year in 2016).[4]

Netflix has estimated that users spend 60 to 90 seconds browsing on its interface for new shows to watch before they lose interest.[5] These machine learning algorithms help users navigate through Netflix’s vast library, translating into 80% of watched content coming from algorithmic recommendations[6] and annual savings of well over US$1 billion from decreasing churn rates[7].

Historically, Netflix relied on customer reviews (ranking from 1 to 5 stars) to predict consumer preferences. After the introduction of machine learning algorithms on their streaming content, Netflix started collecting and analyzing a wide arrange of data on many metrics considered to be better predictors of subscriber behavior including:

  • How many users watched an episode and an entire series
  • Episode and series rating
  • In-title behavior (pause, rewind or fast-forward)
  • Churn rate per episode or series
  • User browsing and scrolling habits[8]

Netflix is investing heavily on providing customers with the largest and most relevant content library in the market. Therefore, it is instrumental to provide subscribers with an effective medium to facilitate the navigation and selection process of this library which is considered by Netflix to be the “moment of truth”[9].

(2) Content acquisition and creation

In 2013, the company started producing original content which supports the company’s long term strategy of relying less on outside studios[10]. The data they gather using machine learning has been widely used internally to acquire and create relevant content that users will find exciting. As traditional studios shift towards their own online streaming models (e.g. Disney+), content supply will likely see a dramatic decline, which drives Netflix’s need to strengthen its internal creation efforts and acquire the right material to provide its subscribers with a relevant and exciting library.

In the short-term, Netflix is expected to increase investing in original content (US$ 8 billion in 2018)[11] and continue developing machine learning to guide subscribers to relevant content. In the medium term, Netflix is experimenting with a groundbreaking interactive “Choose-Your-Own Ending” format to provide customers with a tailor-made title experience while learning new insights from their users.

Machine Learning Meets Open Innovation: Choose-Your-Own Ending

In an unprecedented move, Netflix is going to provide its users the ability to control the narrative of the show “Black Mirror” and let them pick the ending for some of the episodes in its 2019 season. Even though content creation costs will increase with this format, the potential benefits of outsourcing innovation to subscribers and getting insights from their behavior when they are most active can create value for the company outside the “Black Mirror” universe.[12]

What Else Can Netflix Do?

Streaming content is largely a solo activity (60% of Netflix activity)[13] but there is still a significant number of hours in Netflix that occur in more social settings. Deciding which title(s) to watch can be a very daunting task for a subscriber to decide alone and can get even harder with company. In order to maximize user engagement and satisfaction, Netflix must find ways to aggregate data from different users and develop models that could predict preferences of a viewer when that person is watching content alone vs. with friends, significant others and family.

Challenges Ahead

  1. As Amazon is increasingly collecting more data on its users beyond their watching time (e.g. marketplace, grocery shopping), how can Netflix improve its machine learning capabilities or where else can the company gather more data to provide a better customer experience than competitors?
  2. For international audiences, their viewership data is biased towards local content or foreign content that has been translated into local language. Can Netflix develop a model that would allow them to predict if some foreign content (if translated) could be successful in a new market leading to significant savings in local content creation costs?

(792 words)


[1] “Amount Of Original Content Titles On Neflix 2017 | Statistic”. 2018. Statista., accessed November 2018

[2] Clark, Travis. 2018. “New Data Shows Netflix’s Number Of Movies Has Gone Down By Thousands Of Titles Since 2010 — But Its TV Catalog Size Has Soared”. Business Insider., accessed November 2018

[3] Gomez-Uribe, Carlos A., and Neil Hunt. 2015. “The Netflix Recommender System”. ACM Transactions On Management Information Systems 6 (4): 1-19. doi:10.1145/2843948

[4] “Parks Associates Announces Update To OTT Subscriber Churn Rates For Netflix, Hulu, And Amazon Users”. 2016. Parksassociates.Com., accessed November 2018

[5] Gomez-Uribe, Carlos A., and Neil Hunt. 2015. “The Netflix Recommender System”. ACM Transactions On Management Information Systems 6 (4): 1-19. doi:10.1145/2843948.

[6] “This Is How Netflix’s Top-Secret Recommendation System Works”. 2018. Wired.Co.Uk., accessed November 2018

[7] McAlone, Nathan. 2016. “Why Netflix Thinks Its Personalized Recommendation Engine Is Worth $1 Billion Per Year”. Business Insider., accessed November 2018

[8] “How Netflix Uses Analytics To Select Movies, Create Content, & Make Multimillion Dollar Decisions”. 2018. Neil Patel., accessed November 2018

[9] Netflix, Q4 2015 Letter to Shareholders, p. 5.,accessed November 2018

[10] Netflix, Q3 2018 Letter to Shareholders, p. 3.,accessed November 2018

[11] Morris, David. 2018. “Netflix Is Expected To Spend Up To $13 Billion On Original Programming This Year”. Fortune., accessed November 2018

[12] Shaw, Lucas. 2018. “Netflix Is Planning A Choose-Your-Own-Adventure ‘Black Mirror’”. Bloomberg.Com., accessed November 2018

[13] Truong, Alice. 2018. “You’Re Not Alone: Binge Watching Is A Solo Activity For Most People”. Quartz., accessed November 2018


The Road to Autonomous Driving at Fiat Chrysler Automobiles


Can Machine Learning Save Us From Being Stuck in Traffic?

Student comments on Running on Netflix: How Machine Learning is Fueling Your Netflix Binge-Watching Problem

  1. Great post! As an avid Netflix viewer, the “choose-your-own-ending” format really appeals to me and seems like an innovative use of both machine learning and open innovation. However, do you think it only works because the story lines for Black Mirror are contained to one episode each? Would creating too many decision points become too costly? Additionally, creating these alternate versions only tests a very limited number of scenarios. Creating one episode shows and seeing what people gravitate to may provide similar insights. Alternatively, Netflix could run a crowd sourcing campaign where they ask viewers how they would respond and pick the most popular / interesting ones to film (though turnaround time here could be very slow).

  2. I did not know that Netflix was letting viewers control the narrative of the show “Black Mirror” and allowing them to pick some of the endings for the 2019 season. I am curious to see whether this strategy increases viewer engagement – when I watch a movie, I don’t necessarily want to know or predict what happens. I like being surprised by unusual endings, and it seems that placing the ending in viewers’ control might be counterproductive to that.

    More generally, I think it is great for Netflix to leverage big data and machine learning to better predict user preferences. However, by using past behavior to predict future preferences, Netflix will not be able to “surprise” users with new genres of movies that are completely unrelated in characteristics to movies watched in the past. Another potentially intriguing application of machine learning might be to track user behavior to monitor for multiple people using the same Netflix account.

    1. Thanks for reading. I agree with your concern of killing the suspense of an ending by having you pick it but I think Netflix’s strategy is more towards giving you the ability to get in the mindset of the characters, transport yourself to that world, make decisions and see what the consequences would be (without the actual recourse).

  3. I absolutely agree on the suggestion that Netflix should leverage their data of foreign markets in order to fill the funnel of upcoming movie and TV show ideas. Hollywood studios have already done this partly in the past where they copied Asian and South American Foreign Box office hits and adapted them to the preferences of the US viewers. With all the data, Netflix can do this as well and might even just go one step further in straight taking a foreign movie and provide them to audiences in other countries with the help of subtitles.

  4. Awesome post! Well-written and clearly captures the important ways in which Netflix employs machine learning. The “choose-your-own-ending” idea is particularly interesting as it combines machine learning with open innovation, though I wonder how cost intensive such an endeavor would be for Netflix. In addition to Netflix’s move into content creation, I wonder how they can use machine learning to better advertise their new content? In my experience, and as your essay outlines, it can be tedious and time-consuming to sift through Netflix offerings to find a new show. I know they address this through content suggestion, but there are still so many options. It seems that most interest in shows is generated from word-of-mouth. If they partnered with social media platforms, for example, they could generate ads for shows based on your search results or influencer preferences.

    1. Tom Riddle thanks for reading and completely agree with you on the social media platforms piece (many of the shows that I watch were recommended by someone that I deem has good taste on TV shows/movies). There are many applications to these algorithms and sets of data that are worth exploring.

  5. Very interesting! I’m also interested to learn about the controls Netflix puts in place to guard against merely repeating the same types of content over and over. In other words, if Netflix uses customer information to see what types of shows a certain person likes, how to they think about introducing new types of content rather than just showing the same type of content? Similarly, I wonder how they control for this same issue when green-lighting new shows.

  6. This was a fantastic post, thanks for submitting it! I would be curious to hear if you fear that as Netflix gets better and better at predicting viewers interests and giving them more control, if this will continue to grow into unhealthy viewing habits of individuals, particularly children. I know this wasn’t the perspective of this post, but just a question I had as I was reading through.

    1. Thanks for reading! I agree with your concern in that these ML algorithms will only take our necessity of getting leisure increasingly to an extreme. Hopefully they are able to invest in creating/acquiring relevant educational content for kids which could lead to a better societal impact than just merely entertainment.

  7. I echo the sentiment in Michael Scott’s post above – it concerns me that Netflix is able to use this data to manipulate viewing behaviors – specifically binge watching (which I believe they encourage and strive for). Just as social media companies, mobile application developers, and smart phone manufacturers are trying to make their products as addictive as possible, Netflix is using our viewing behavior data to do the same. This strive for addiction (perhaps a harsh word, but fair in this case) will have long-term negative consequences for Americans and people across the world, and machine learning is making this quest ever easier.

    I also worry about Netflix’s ability to understand our viewing behaviors will assist them in influencing our opinions, as we have seen nefarious users of social media accounts do recently.

    1. Thanks for reading and for the great insight. Independence in media is one of the main problems that we must fight for today to maintain our democracy. I would say that some balanced political commentary in my series is OK but when they introduce these sophisticated ML algorithms they can actually start very subtly influencing our ways of thinking. I think this is more a corporate governance matter of making sure Netflix is not significantly controlled by an interested party (just like they do for newspapers and other publicly-traded media outlets) and that creative departments have strong ethical sense.

  8. Great post! I really can’t wait for the “choose your ending” functionality!

    Addressing your concerns, your thoughts on how Netflix can obtain more data is very interesting. We are moving to an era where data is one of the most valuable resources needed for a company’s success and, undoubtedly, the more data the merrier. I think that in order to obtain this data, Netflix has to leverage its size and either partner with another large online company and share their databases (e.g. with Walmart) or to diversify its business, in order to observe the customers’ behavior in other aspects of their lives. This way, Netflix will be able to determine what are the general preferences of its customers (e.g. do they like adventure and/or sports? Do they have kids? etc) and offer them an improved customer experience.

    Regarding the international audiences, I think that Netflix should focus on the content of the movies that a particular viewer prefers. As a non-US viewer, usually the country of origin of a movie doesn’t really make a difference to me!

  9. Great post! In reading this, I thought it was really interesting to see how Netflix has successfully leveraged machine learning to suggest and create content. Though these were major successes for the company, as TV purist, I think the “choose your own ending format” gets a little dangerous and really removes most of the “art” from television. In thinking about some of the greatest TV shows of all time (e.g. Sopranos, Breaking Bad… Gossip Girl), part of what made them so interesting was the creative choices the directors took in developing characters and ending story lines. By removing the element of surprise, I fear TV will suffer and Netflix’s content library could be increasingly viewed as “gimmicky.” Further, I wonder how changes like this could impact of creative control they give their creators and how that may impact their ability to attract talent which has historically been a real differentiator for the company.

  10. Often when scrolling through Netflix, I am disappointed by the recommended movies that usually populate for me. Although the films are usually within my desired genre, I never have felt that the algorithm has read my preferences as well I would have liked. Perhaps Netflix should consider letting users add input to refine their algorithm. I’ve seen similar examples where users can search for music genres and artists and from that listing, playlists are formulated that do a better job of pinpointing a relevant starting point. Film preference however, may have a higher standard deviation from a user’s “medium” genre choice. Allowing for input would decrease variation and allow machine learning to begin at a more accurate starting point.

    1. Thanks for reading and for the insight. I agree with you and that is why they have a review feature where you can cross out the TV shows and movies that you watched that you definitely didn’t enjoy. They can improve this by asking you periodically which types of movies you are interested in watching or which genres are most appealing to you.

  11. I love this article – super interesting read.
    I agree with Camille’s comment above- I am curious to know how they link new content to user preferences or user profiles.
    Additionally I always thought the suggestions of Netflix are somewhat on point – at least for me. However, as mentioned in the article, sometimes the viewing happens in social settings which is not necessarily something you want to drive your suggestions – so these searches skew the suggestions in a certain way. Maybe it would be useful to maybe have the option to select whether its social or individual?

    1. Thanks for reading. I completely agree with you in that a toggle for social vs. individual could be very helpful to make more accurate suggestions given that they are able to create a control for several people ghosting other accounts (multiple users in one account can create a lot of noise in the data leading to Type 1 errors that can erode customer experience).

  12. I think a critical input into Netflix’s engine is it’s smart tagging system in the categorization of “micro genres”. You touched upon how Netflix collects viewership data in how users engaging with the browsing and watching stage — which is key — but the crux of their recommendation engine where titles are surfaced is around big data tagging. Here are a few super interesting articles on that topic, and why your personal categories may seem so strangely specific and strangely relevant based on your previous selection behavior…For example, “Romantic Indian Crime Dramas”, “Time Travel Movies starring William Hartnell”, “Visually-striking Foreign Nostalgic Dramas”, and “Cult Evil Kid Horror Movies” are just a few genres out of 76,897 that The Atlantic author was able to mine from Netflix’s database:

  13. Interesting post! I am shocked that Netflix saved over $1B just through personalized content to lower churn rates. As a customer who has used Netflix, Prime Video, and Hulu, I can easily say that Netflix does the best job at surfacing titles that I want to watch. I am curious how Hulu would be affected by these algorithms. My hypothesis is that Hulu’s canceled subscriptions are driven by commercials on a paid product (and un-willingness to pay on the premium subscription with no commercials). The second thing I’m curious about is how much of that savings is actually attributed to the model vs. having better original content or more selection on their platform.

    1. Completely agree with you! Price points of Netflix and Hulu are not that different yet the experience in Netflix is significantly better (that might explain to some extent the difference in churn). Hulu has more relevant TV content from other networks that Netflix is strategically shifting away from but it seems like it is not enough to retain customers at a meaningful level.

Leave a comment