Recommendation Systems
We talked in class today about how networks can be used to form a model of recommendations. Indeed, the ability to predict an individual’s preferences about an item has become an increasingly important in the world of online content providers. For instance, video streaming services such as Netflix and Amazon Prime rely on being able to feed their customers new shows and movies to watch. This is why in 2006 Netflix launched a $1 million prize to a team that could predict a user’s movie rating 10% better than they could. It took 2 until 2009 for a team to finally complete the task, using a combination of 100s of predictive models [1]. The importance of recommendations systems goes beyond videos, however, as they are used to recommend everything from news articles to romantic partners.
There are two main types of recommendation algorithms. The first is content-based filtering systems, which only rely on data about the content being recommended. For example, Pandora builds up a station by analyzing the attributes of the seed music and finding other songs that share many of the same attributes. When a Pandora user rates a song, this data is then used to further refine which attributes are important and which are not. This approach requires a method to determine the attributes of a piece of content. In this particular example the song attributes are provided by the Music Genome Project, a service which catalogues songs by about 450 characteristics [2]. The second type, collaborative filtering, is finds other users who are similar to a given individual, and then recommends the content they enjoyed. While this system does not require classification of the content by attributes, it does require a large user database to drive the predictions off of, in addition to large amounts of data about which content these users liked. Today, most major recommendation systems use a combination of these two types so that it can take advantage of the strengths of each.
[1] http://techblog.netflix.com/2012/04/netflix-recommendations-beyond-5-stars.html
[2] http://www.pandora.com/about/mgp