Skip to main content



Tracked: How Does Netflix Know What You Like?

Have you ever wondered how streaming platforms seem to have a perfect sense of what you might like to watch next after you’ve finished a movie or a show? While in today’s world we are well aware that our online activity is being tracked, we may not realize the extent to which our viewing preferences are being tracked and used. But how does this tracking occur?

We can take a look at Netflix – one of the largest streaming platforms in the world. Netflix uses something called “collaborative or content-based filtering, or a combination (hybrid) of the two” (Brams).  This specific type of filtering uses information gathered from other users who have also watched the same things you have, to generate new recommendations for you. For example, let’s say individual A and individual B both like comedy movies. Individuals A and B both saw and enjoyed the movie Mean Girls. Individual A also saw the movie Legally Blonde, but individual B has not seen this movie. The collaborative finding algorithm would recommend Legally Blonde to individual B because of individual B’s demonstrated shared interests with individual A. Content-based filtering looks instead at the content of the media consumed, and gives recommendations based on similar content. So a content-based recommender would see that Mean Girls and Legally Blonde share the same genre of comedy and teen films, and would recommend them to individuals based on those tags. We can dive deeper and ask ourselves on a large-scale, how are these connections made and represented?

To be able to really see how the recommendations come to be, we must consider the use of network graph models. In the simplified example of Mean Girls and Legally Blonde above, individuals A and B would be nodes with edges from both individuals going to a node representing Mean Girls. Individual A would also have an edge connecting it to a node representing Legally Blonde. Because of the A and B’s similar relationship to the Mean Girls node, individual B would be recommended Legally Blonde, and if B watches it and enjoys it, they can become a reference point for someone else to watch it as well. 

Getting a bit more complicated, we can investigate the use of knowledge graphs. Knowledge graphs tell us “what items are related to what properties … how they are related and impose no restriction on what can be related and how” (Brams). These knowledge graphs include components that are collaboration graphs. “Collaboration graphs record who works with whom in a specific setting” (Easley, Kleinber), with co stars of movies, shows, and other media being a primary example of such graphs. These graphs employ paths which are comprised “of nodes with the property that each consecutive pair in the sequence is connected by an edge” (Easley, Kleinber). The nodes of the knowledge graph are individuals in films, movies, shows, documentaries, and genres, and the edges tell us how those nodes are related to each other. Through this network of paths, nodes, and edges, these graphs can connect vast numbers of media based on actors who have starred in them, their genres, their content/genre, their accolades, etc.

In the graph below, we can see the paths between content, and genres and other categories that would push certain categories and nodes to be more significant in influencing the recommendation process. In this specific graph, if the user were to read I am Malala, nodes that are more closely related to the I am Malala node are pushed through the ranks, so the computation for recommendations is over a smaller number of nodes that are more related to content the user has consumed and liked.  Certain algorithms like PageRank, which is also used by Google, allow for there to be “user-specific relation weights” (Brams). If a user likes watching certain actors in movies, regardless of the genre, the algorithm “could weigh the Stars and Co-stars relations higher for that user” (Brams). 

Graph Representing Connections

In this manner, our viewing preferences are constantly tracked and recorded. Because knowledge graphs create a complex network that would otherwise be difficult to generate, our tastes also assist in recommendations for others through connections made in the recommendations graph. Through such connections, these algorithms and graphs effectively present us with recommendations we actually enjoy, and thus we spend less time searching for media that we want to consume. Hot take: maybe being tracked isn’t so bad.

Links:

https://towardsdatascience.com/movie-recommendations-powered-by-knowledge-graphs-and-neo4j-33603a212ad0

https://www.cs.cornell.edu/home/kleinber/networks-book/networks-book-ch02.pdf

 

Comments

Leave a Reply

Blogging Calendar

September 2021
M T W T F S S
 12345
6789101112
13141516171819
20212223242526
27282930  

Archives