This page may be out of date. Submit any pending changes before refreshing this page.
Hide this message.
Quora uses cookies to improve your experience. Read more
ad by honeypot.io
The Netherlands faces a shortage of developers.
The developer exclusive job platform. Companies must approach you with clear salary expectations.
Sign Up at honeypot.io
7 Answers
Xavier Amatriain
We have talked and published extensively about this topic. Let me start by saying that there are many recommendation algorithms at Netflix. People usually refer to the "rating prediction" algorithm that was researched in the Netflix Prize as the "Netflix Recommendation Algorithm", but that is by no means the only or the most important of the algorithms in the Netflix recommendation system.

Now, if we focus in rating prediction and the outcome of the Prize, and to complement Neal's answer, there are two algorithms that are being used in production right now: Restricted Boltzman Machines (RBM) and a form of Matrix Factorization.

Restricted Boltzman Machines are, simply put, fancy neural networks. There are some tricks to make RBMs work in the context of collaborative filtering. If you are interested, read this very good paper by Hinton and some of his students: Restricted Boltzman Machines for Collaborative Filtering.

The form of Matrix Factorization in use is the so-called SVD++ developed by the winning team of the prize. This is basically an asymmetric form of SVD that can make use of implicit information (just as RBMs do also). The Factorization Meets the Neighborhood: a Multifaceted Collaborative Filtering Model paper by Koren et al. explains the details.

Those two algorithms already appeared in the 2007 Progress Prize. In production they are combined using a linear blend. So, why aren't the other 100+ algorithms that were combined with a Gradient Boosted Decision Tree used? There are several reasons that include engineering complexity and the fact that, as I mentioned before, rating prediction is not the main concern nowadays. There are many other recommendation algorithms from personalized ranking to page optimization that make up the Netflix recommendation system.

If you are interested in learning more about these start by reading our blogpost Netflix Recommendations: Beyond the 5 stars (Part 1) or take a look at some of my recent slides The Recommender problem Revisited.
112k Views · View Upvoters · Answer requested by Sizheng Chen (陈思政)
Your response is private.
Is this answer still relevant and up to date?
Neal Lathia
The algorithms that were developed as part of the Netflix million-dollar prize (which aimed to improve the movie recommendation system) are blends of a large number of different machine learning techniques.

Two of the most notable aspects that emerged from the competition were using matrix factorisation and the so-called "temporal dynamics" to perform collaborative filtering; the full details can be found on the forum page (which has links to papers written by the winning team):

http://www.netflixprize.com//com...

While these will give you insights into the result of the competition, I have no information about how many of these proposals made their way into Netflix's live system.
29.4k Views · View Upvoters
Your response is private.
Is this answer still relevant and up to date?
Garrick Saito
I don't work for the company and have no insider knowledge, but from what I can see, they attempt to predict what you will like based on the feedback you provide on the movies you've watch.


How often you watch certain movie genres build a 'movie-watching-profile' of sorts, helping them to understand what you like and what you don't.


The more movies you rate, the more information they have to go on to assess your movie-watching preferences.


They may also be assessing your rental and streaming activities.


Very likely, all of the information available about your reviewing behavior is getting chewed up and digested by their algorithms to figure out what you'll like.

This is all just a guess though.  Very likely, they'll never reveal their proprietary methods for determining their suggestions.
13.6k Views · View Upvoters
Your response is private.
Is this answer still relevant and up to date?
James Schek
Garrick Saito's answer is pretty accurate at a high level. The Netflix Tech Blog discusses some of the details of the recommendation engine in a two part series,  Beyond the 5 stars (Part 1) and  Beyond the 5 stars (Part 2). It helps to be a bit of a math or algorithm junkie.

Before I begin, let me say that I don't work on personalization or recommendations so this is all second hand and from what I can recall hearing from other people.

At a basic level, Netflix primarily uses your ratings, viewing history, and taste preferences to determine your recommendations. I think there are other factors used such as geography, preferred language, viewing device, time of day, etc, but I am not certain.

These factors are used to group customers into "clusters" with similar viewing habits. A customer can belong to multiple clusters. Based on the cluster, Netflix can then identify the movie/show characteristics that would be most appealing to the customer or specific titles that are popular within that cluster.

So by watching Kill Bill: Vol. 1 and giving it 5 stars, the algorithm might identify you as someone who enjoys "Quentin Tarantino Movies" or "Movies with Strong Female Leads who Kick Butt" or something. (I'm making up those categories since I have no idea what the real ones are).

Those categorizations would influence the recommendation categories and titles you see, such as "Campy Action Movies with Strong Women" or "Good Quentin Tarantino Movies". So this you might find Pulp Fiction and The Fifth Element under your recommendations because they fit into those categories.

Through some additional data mining, the algorithms may also find that clusters of people who enjoy those categories also tend to watch and complete the TV show House of Cards. So this might make House of Cards show up in your "Popular on Netflix" list--because it is popular among people like you.
20.2k Views · View Upvoters
Peter Szanto

This question was pretty much answered by experts, and you can find links to the most important publications about the topic in the previous comments.
 
However, if you’re still interested check out this post, I wrote about the Netflix recommendations a few days ago. It sums up many of the reasons why this system is so efficient, and supplements the previous insights with more recent developments.

You can read it here: How Personalization Made Netflix The Biggest Video Streaming Service in the U.S.

5.2k Views · View Upvoters
Mahmoud Akl
There is an interesting article that was published on the Atlantic one year ago titled How Netflix Reverse Engineered Hollywood that will give you some insight.
10.7k Views · View Upvoters
In addition to the other answers, which explain how Netflix looks at your rankings and viewing habits, they also give preference based on what it costs them to let you watch a program. I don't have the source handy, but I read this somewhere, and it makes sense. They want to encourage you to watch the less costly content, to keep their costs down.
1.6k Views

Related Questions

Related Questions