Comment #1
Comment #2
Reference:
Title: Personalized News Recommendation Based on Click Behavior
Author: Jiahui Liu, Peter Dolan, Elin Rønby Pedersen
Venue: IUI 2010, Google Inc.
Summary:
This paper discusses new ways people who read their news online could get that news. Large numbers of people use things like Google News and Yahoo! to get their news. However, the previous algorithm Google News used was comparing against people who had read similar articles. This poses a problem because recently released articles that haven't been read yet may be skipped over. This system is designed to fix that problem.
Google News, as well as many other sites, use text-based classifications to sort articles into categories. This system would do the same as well, but instead of comparing what articles the user read to what others have also read, it will compare to all articles, with an emphasis on the recent. To conduct this analysis, the group used Google News users, and kept everything anonymous per the Google policies. The first thing this system does is compares what categories the user typically clicks on. This provides a "click distribution." Next it compares the click history with previous months to see if there is a change in the user's interests based on their clicks. Then the click distributions of all users on particular news items is compared to analyze the potential impacts on local, national, and global news. Using this data, a Bayesian framework is constructed to predict the user's interests. Also, the group kept a "recommended reading" section, which did not contain personalized articles, but articles that were popular or significant elsewhere. The results of their test showed that frequency of visits from the test group was 14.1% higher, with a 99% confidence interval. This is a significant improvement.
Discussion:
I thought this article was pretty entertaining. I can see lots of people using it, and only makes Google even more superior. There was lots of technical details and formulas that I didn't feel were necessary. I think a good article shouldn't contain many of those. Those should be saved for a sales pitch or something like that. It's a good way to get people coming back for more, in my opinion. I haven't ever used Google News, but I think I may try it for my news information. They proved that this method works, so it is unlikely that this would fail when taken out of testing. My one complaint is that it takes a click as a positive vote. What happens if you don't like what you clicked on after reading it?
I agree that this is a really useful idea. I am also kind of curious if they can extend this to other services that they have.
ReplyDeleteAs far as taking clicks as a positive vote, if they got you to click on it then they showed you something that was pertinent enough to your interests, so that is technically a success regardless of if you liked it afterwards.