Business Case: Deciding What to Watch—Video Recommendations at Netflix

Netflix is the undisputed leader of video streaming services, accounting for more than half (53%) of U.S. video streaming subscriptions. Amazon Prime Video (25%) and Hulu (13%) are the company’s largest competitors. Netflix is also the oldest company in this group, having originally started as a DVD by mail rental service. Unlike other companies that dominated the DVD rental business, Netflix successfully made the transition to on-demand video streaming by investing in new technology and redefining its business model. The service is now available in 190 countries and claims over 90 million subscribers globally.

Netflix executives credit the company’s recommendation system for driving the “Netflix experience” and boosting profitability (Gomez-Uribe & Hunt, 2015; Raimond & Basilico, 2016). Surprisingly, the origin of the recommendation system dates back to 2000, when Netflix was still a DVD rental service. Recommendations during these early days were based largely on members’ movie ratings. Ratings often reflect how people want to be perceived as opposed to how they act. For instance, rating data will tend to overemphasize how much people like documentaries and foreign language films, whereas behavioral metrics provide more accurate measures of how subscribers use the service. Today, when Netflix subscribers use the online service, they see recommendations generated by multiple algorithms that use descriptive information about the subscriber and their past viewing behavior (Gomez-Uribe & Hunt, 2015). Netflix claims that 75% of the activity on the service is a result of the recommendations it offers subscribers.

 Netflix Analytics

Netflix enjoys a significant advantage over traditional television channels because the company collects information about how subscribers use the service. Netflix can make marketing and product decisions based on several behavioral metrics. You might be surprised at the details Netflix collects:

• The device you use (tablet, Roku, smart TV, etc.)
• Where (zip code) you watch from
• The days and times you watch
• When you pause, rewind, or fast-forward during viewing
• How you search—the words and phrases used, how long you search, etc.
• Whether or not you watch the credits following a show
• How many episodes of a series you watched
• Whether or not you watch all episodes in a series
• How long it takes you to watch all episodes in a series
• How many hours you spend using the service
• What movies and television shows you watch
• How often you use the service

In addition to making recommendations, Netflix uses the information to do the following:

• Identify subscribers who are likely to cancel the service
• Select new movies to add to their catalog
• Decide if a television show should be renewed for another season
• Identify movies and television shows to drop from the catalog
• Determine the days and times to recommend certain movies or shows
• Determine what to recommend immediately following the viewing of another movie or show
• Determine how to describe movies and shows (i.e., long vs. short descriptions)

 Recommendation Algorithms at Netflix

The Netflix home screen can offer up to 40 rows of recommendations to a subscriber. Each row is generated by a different algorithm designed to personalize recommendations as well as determine the order in which movies and shows are listed. Each row is based on a different theme or rationale for the titles appearing in the row. Netflix even uses a Page Generation Algorithm to personalize the type of row-level recommendations and their order when creating the page. Some examples of the different recommendation rows include the following:

 Genre Rows

Several of the rows appearing on the home page are based on movie or television show genres that Netflix believes the subscriber will be interested in based on past viewing behavior. Genre rows are generated by what Netflix calls its Personalized Video Ranker (PVR). The rows reflect three levels of personalization: (1) the selection of the genre, (2) the selection of specific titles within the genre, and (3) the ordering of the titles.

 Continue Watching

Titles appearing in the Continue Watching row highlight episodic content that Netflix thinks a subscriber might want to return to. The Continue Watching ranker evaluates recently viewed videos for signals that a subscriber intends to resume watching or is no longer interested in the title. These signals include things like time since last viewing, point of abandonment (mid-program, end of program), if other titles have been viewed since, and type of device used.

 Because You Watched

The Because you watched (BYW) row is based on the similarity of recommended videos to past videos watched by the subscriber. The BYW row is determined by the Sims Ranker, which generates an ordered list of videos, based on similarity, for every title in the catalog. Various personalization cues are then used to further refine the subset of videos that actually appear in the row on the home page.

 Top Picks

The goal of the Top Picks row is to feature Netflix’s best guess as to the videos in its catalog that are most likely to be of interest to the subscriber. The Top Picks algorithm uses cues from the individual subscriber along with viewing trend information to recommend titles from among the most popular or top-ranked videos in the catalog.

Netflix believes that its recommendation system plays a significant role in user satisfaction and customer retention. A team of workers regularly updates the system with new algorithms and modifications to existing ones. Their ultimate goal is to generate such high-quality recommendations that subscribers will rarely have to search for videos to watch.

Questions

1. You read about four different types of recommendations that Netflix features on their home page. Think of a new type of recommendation row that Netflix could use and the kind of information or behavioral metrics that would be needed to generate your recommendations.
2. Based on the information in this case, would you say that Netflix primarily uses content-based filtering, collaborative filtering, or both? Explain your answer.
3. Netflix is expanding globally. When Netflix first enters a market, the recommendation system can face “cold start” or “sparsity” problems. Explain why this happens and suggests ways that Netflix might deal with these challenges.
4. What metrics do you think Netflix could use to identify subscribers who are likely to cancel the service?
5. Visit Netflix’s Technology Blog http://techblog.netflix.com . Identify three challenges that the company faces in generating recommendations for its subscribers.