Generator Microsoft Word 11 (filtered medium) Dear Colleagues,
Following-up on our discussion of the problem of reinforcement (ant navigation) in recommender systems, I asked Marcus Spiering, one of the principal researchers in the BibTip project, how the system addresses the issue. Here is his response (e-mailed to me):
<quote>
If you have a look at the JS-File [embedded in the full details view of a bib record] you can see a parameter fbt, which indicates whether a detailed view request comes from a recommendation list or from a hit list. For the calculation we take only the requests from the hit list into account, because otherwise our underlying assumptions of a poisson process would no longer hold. Without differentiation we had a mixed process, consisting of a poisson process and a markov chain ("ant navigation"). We are already working on this hybrid statistical models but currently it is not above research level.
The live version of bibtip works only with the observations coming from the hitlist. In other words, we are aware of the "ant navigation" problem and found a way to avoid it.
</quote>
Interestingly, he also notes that, in the Karlsruhe catalog, 15% of the total requests for a detailed view of a record come from recommendation lists.
Thanks,
Kevin
--------------------------------------
Kevin M. Kidd, MA, MLIS
Library Applications & Systems Manager
Boston College Libraries
Phone: 617-552-1359
Fax: 617-552-1089
e-Mail: kevin.kidd_at_bc.edu
Blog: http://datadrivenlibrary.blogspot.com/
Received on Wed May 21 2008 - 09:32:04 EDT