What you people think is the appropriate amount number of data points
necessary to protect patron privacy in a recommendation system?
One point would be a situation where, if only one user took out or
looked at both Book A and Book B, the recommendation system would
reveal this coincidence. I contend this would violate patron
privacy—if you knew one book someone took out you could discover
others. The logic of small numbers would undermine the idea of
anonymity.
I'm thinking you need at least three, and probably more. John Blyberg
went for three or more in his SOPAC recommendations
(http://www.blyberg.net/2007/01/31/dynamic-item-recommendations/). I'm
not sure if that was for quality or privacy. That was based on opt-in
data.
Tim
Received on Tue May 20 2008 - 21:04:48 EDT