Thursday, June 4, 2009

Personalized Recommendation on Dynamic Content Using Predictive Bilinear Models

Dynamic Content Using Predictive Bilinear Models is a Yahoo! paper about suggesting relevant news stories for Yahoo's "Today" module The paper describes a bi-linear model leveraging both instantaneous dynamic click-through information and static content based features. The objective function adopted is concave and minimized with a gradient descent approach.

The authors collected about 40 million click/view events by about 5 million users from the random bucket before a certain time stamp for training. One evaluation metric adopted is the number of clicks in each rank position. The intuition is that a good predictive model should have more clicks on the top-ranked positions and lesser clicks on the lower ranked position. Results are quite interesting: the model shows the benefits of adopting dynamic click-through data.

Unluckly, no information is given about the time needed to process the data. A rather important point when you process (near) real time data.

No comments:

Post a Comment