## Netflix’s Unparalleled Algorithms

Netflix has grown from a small dot-com era venture to a titan in the tech industry over the past two decades. Netflix has thrived by recruiting employees that create high-quality systems to support both streaming and direct-mail services. Furthermore, Netflix has been able to succeed by leveraging the skills and talents of outside groups to optimize their search algorithm to improve the user experience. How did Netflix manage to lure talented outside groups to work on these projects? They offered the Netflix Prize, a \$1,000,000 cash award. The Netflix Prize started in 2006 as an open competition to “improve the accuracy of predictions about how much a user is going to enjoy a movie based on their movie preferences” (Netflix, 2009). Groups assembled from across the world to beat out Netflix’s Cinematch algorithm by over 10 percent to claim the \$1,000,000 prize.

The winning team, BellKor’s Pragmatic Chaos, won the prize by creating an algorithm that built off the principles of a bipartite graph. In this example, Netflix users were a set of vertices and movies/TV shows were a set vertices (Blelloch, 2011). An edge was created if a user had watched a movie or a TV show and a weight was assigned to that edge depending on the user’s rating. This overall set-up of the problem would satisfy the definition of a bipartite graph as its “vertex set could be partitioned into two cells such that every edge joins a vertex in one cell to a vertex in the other cell” (Gross, 2013). The groups then used predictive analytics to predict future edges between users and movies/TV shows. While there are well-established ideas on predicting links in unipartite graphs, these assumptions do not often hold (if they hold at all) for bipartite graphs (Kunegis, De Luca, Albayrak, 2010). In fact, “the special structure of bipartite graphs makes common link prediction algorithms ineffective”. Thus, the teams could have employed algorithms such as “algebraic link prediction methods that restrict spectral transformations to odd functions” or a “variant of the von Neumann kernel” to maximize predictive analytic power (Kunegis, De Luca, Albayrak, 2010).

Ultimately, while complicated methods were used to create the algorithm that eventually won the Netflix Prize, the overall set-up was grounded in the idea of Netflix being one giant bipartite graph between users and TV shows/movies.

http://www.netflixprize.com/