I got some great advice from a couple of my favorite bloggers recently. Daniel Tunkelang had several good suggestions, but the one that really stood out was that I should write faster. :-) And Andrew Chen suggested that I try to incorporate more real data and case studies into my writing. I’ll do the best I can with both suggestions, though the case study request may prove a bit challenging, since while my job with Netezza lets me hobnob with lots of cool digital media customers that have transformed their businesses through terabyte-scale data analysis, most of them don’t really like disclosing their sources of competitive advantage to the world (so while you would recognize their names, I can’t write about them). And besides, this is a personal blog.
So I’ll write a really fast post about this awesome Google research paper: “Google News Personalization: Scalable Online Collaborative Filtering”. It may not be quite as awesome as James T. Kirk, but it was nominated for the best paper award at the 16th International World Wide Web Conference in 2007.
Here are the things that make this paper awesome:
1) While this paper is about improving the relevance of recommended news articles in Google News, the ideas presented also apply to the challenges in increasing campaign lift for online advertising – a topic that’s of particular interest to me (and the entire advertising industry).
2) The punch line is that data scale and item churn (characteristics common to both the Google News problem and the online advertising problem) make query performance and data load performance even more important for increasing targeting precision in these domains than in other domains where recommender systems are relevant.
3) The authors’ approach resulted in a 38% increase in click thru rate (CTR) over the benchmark in a test conducted on Google traffic (millions of users) over a period of 5-6 months.
4) This diagram is pretty cool (the blue call-out boxes contain my annotations – the rest of the diagram depicts the paper’s real time recommender system). If this doesn’t make you want to read the paper, I don’t know what will. :-)
If you know of any other good papers on related topics, please leave pointers to them in the comments!
Photo credit: Foxtongue
{ 1 comment… read it below or add one }
Write faster, faster!
Seriously, thanks for the heads up to this paper, which I’d missed. I’m highly skeptical of collaborative filtering approaches to news (I’ve personally been underwhelmed with Google’s attempts), but I know better than to argue with Google over numbers! I am glad they’re attacking the challenge of item churn head-on, since that is indeed a huge difference between news and ecommerce, at least for recommendation approaches that extrapolate from past user logs. And, while I may not like the approach in general, I do respect the researchers’ attempts to incrementally improve it.