Netflix Personas
My girlfriend and I share a Netflix account. By mail, this probably happens a great deal: a couple or a household of people share one account, all getting movies sent to them.
Netflix held a contest that awarded $1 million to any person or team that could improve the quality of their recommendation algorithm by 10%. That is, given any particular individual's viewing history and the viewing history of every other person as inputs, can you tailor the recommendations to this particular individual in way that increases their delight?
There are some "simple" ways of doing this. Take, for example, the Pearson Correlation Coefficient (nerd alert!), which plots data on an Cartesian graph and gets the distance between the preferences of two users. Compare every user in the system to every other one and you can match the closer viewers together.
The biggest problem lies in the second input: everyone else's viewing history. Because, for a family of four sharing a Netflix queue, the viewing history isn't a single point on the chart, but the murky haze between several scattered ones. I'm no math guru, but this is bound to screw up the margin of error when making comparisons between families, couples, and individuals.
Someone won this contest. And my guess is that the winning algorithm took this into account in some fashion. Maybe it made guesses about how many individuals were using the account based on how scattered the viewing history was for each account.
But about a year ago, we hooked up Netflix to our Wii and started watching a lot of stuff on Instant Play. The Netflix delivery has become digital. More importantly, the Instant Queue that my girlfriend and I share has become cluttered with each other's items. Let's ignore for the moment that the Queue itself quickly grows very large. I get recommendations for the television show "Table for 12" and she has to see the recommendations that result from my watching "South Park".
Even between the two of us, the recommendations we're getting are cross-contaminated. And I guarantee that even though the demographics between the audience for "Table for 12" and "South Park" barely overlap that it's affecting the recommendations for all of the people who watch either show.
Simple fix: take a cue from the Wii console itself, in which each user can create a "Mii" (this of it as a cheesy avatar) that looks just like a cartoon version of themselves.
Netflix could do this. When you first start up the app, you can go to your own personal Instant Queue to watch stuff and to get recommendations. You can put your name in, and there's an omnipresent "switch users" option at the bottom right of the screen. If I log in and it's under "Tara", could just switch to "Jim" and be taken right to my own personal Instant Queue.
I haven't done any exhaustive research on this, so it's pretty likely someone at Netflix has thought of this, and they're probably building it. This small addition wouldn't involve any scary mathematics. If not, I'm willing to bet that this simple addition to their app would probably boost the quality of their ratings by at least 10%. Wouldn't it?
If nothing else, it would make their engineers' job easier, as well as make the experience of using their service better for me. I think it's a win-win.
Netflix held a contest that awarded $1 million to any person or team that could improve the quality of their recommendation algorithm by 10%. That is, given any particular individual's viewing history and the viewing history of every other person as inputs, can you tailor the recommendations to this particular individual in way that increases their delight?
There are some "simple" ways of doing this. Take, for example, the Pearson Correlation Coefficient (nerd alert!), which plots data on an Cartesian graph and gets the distance between the preferences of two users. Compare every user in the system to every other one and you can match the closer viewers together.
The biggest problem lies in the second input: everyone else's viewing history. Because, for a family of four sharing a Netflix queue, the viewing history isn't a single point on the chart, but the murky haze between several scattered ones. I'm no math guru, but this is bound to screw up the margin of error when making comparisons between families, couples, and individuals.
Someone won this contest. And my guess is that the winning algorithm took this into account in some fashion. Maybe it made guesses about how many individuals were using the account based on how scattered the viewing history was for each account.
But about a year ago, we hooked up Netflix to our Wii and started watching a lot of stuff on Instant Play. The Netflix delivery has become digital. More importantly, the Instant Queue that my girlfriend and I share has become cluttered with each other's items. Let's ignore for the moment that the Queue itself quickly grows very large. I get recommendations for the television show "Table for 12" and she has to see the recommendations that result from my watching "South Park".
Even between the two of us, the recommendations we're getting are cross-contaminated. And I guarantee that even though the demographics between the audience for "Table for 12" and "South Park" barely overlap that it's affecting the recommendations for all of the people who watch either show.
Simple fix: take a cue from the Wii console itself, in which each user can create a "Mii" (this of it as a cheesy avatar) that looks just like a cartoon version of themselves.
Netflix could do this. When you first start up the app, you can go to your own personal Instant Queue to watch stuff and to get recommendations. You can put your name in, and there's an omnipresent "switch users" option at the bottom right of the screen. If I log in and it's under "Tara", could just switch to "Jim" and be taken right to my own personal Instant Queue.
I haven't done any exhaustive research on this, so it's pretty likely someone at Netflix has thought of this, and they're probably building it. This small addition wouldn't involve any scary mathematics. If not, I'm willing to bet that this simple addition to their app would probably boost the quality of their ratings by at least 10%. Wouldn't it?
If nothing else, it would make their engineers' job easier, as well as make the experience of using their service better for me. I think it's a win-win.