Tuesday, December 20, 2011

How we use 20% time at Adku



Experiments in the social graph
Here at Adku, we love data, and we are constantly looking for ways to leverage new sources of information to make our algorithms better. While our main line of research focuses on the relevance of signals such as demographics and click stream information, we’ve also recently been enamored with the possibilities of the social graph. A few of our engineers used 20% time to put together some hacks in this direction, and we would like to invite you to try them out.
1. Likelygifts (http://www.likelygifts.com)
Now that you’ve tried them out...
Let us explain why we built these tools in the first place.
1. Likelygifts
In spite of what privacy “features”, Facebook, Google+, and Twitter roll out, your actions on a social network are inherently public. They can be culled and managed, but a like is a publicly stated preference, and a wall post is a publicly stated fact. For real privacy, you can use the phone, or meet in person. The broadcast nature of social networks is something that our culture has not fully mastered. Hence much of the rise of social networks has been accompanied by frauds, cons, and Farmville. While social networking has exploded, useful derivatives of social networking have not.   
We built Likelygifts in hopes of changing that. If you’re like us, you’ve just started panicking about all the gifts you have to send out by the end of the month. And who knows what third cousin Bob wants? Or the college buddy who has way too many toys? Maybe we can jog your memory.
Likelygifts doesn’t post on your wall, solicit your friends, or spy on you when you’re offline. We ask for the minimum amount of information to give you accurate recommendations, and that’s all we do.

( a competitor's facebook permissions request vs. ours!)
2. Mildred
We built Mildred as a visualization tool for ourselves. After looking through rows and rows of aggregate statistics about our friends’ likes, we thought maybe there was a better way to see things.
But then we liked Mildred enough to show our friends, and they liked it enough that now we want to show you.
Here are some screenshots:

This is Carlos’s taste graph. Out of all of his friends, he alone likes ‘The Hills’. His other favorite ‘Lost’ is shared by a large number of his friends.
`
David likes ‘The Beatles’, ‘Pink Floyd’, …, and is solitary in his love for Klute.
The most interesting aspect of Mildred for us came from comparing the taste graphs for different people. We found that we could tell who the graph belonged to without even looking at the names. And though everyone seems to like The Beatles & Coldplay, the less well known tastes proved very useful for distinguishing people. Hence one of the upcoming 20% time projects: topic clustering on facebook likes. David is working on that right now, and I am sure something exciting will come of it soon. In the mean time, we'd love to hear suggestions for ways to play with this data.