Another big problem for academia is that it is relatively easy to get OK performance with a recommender and once you have that, the big wins are often outside the recommender algorithm. Often these external factors are things UI changes to set user expectations and such. These factors are essentially impossible to extract into a reproducible academic setting and thus the academic research is necessarily pushed into a less important part of the problem.
It isn't even clear that there is a conceivable solution to this issue. On Tue, Jul 23, 2013 at 9:23 AM, Sebastian Schelter <ssc.o...@googlemail.com > wrote: > Yes, the big problem for academia is the lack of access to real production > systems which would give much more diverse signals to learn from and the > possibility to evaluate on real users. Fortunately, this is slowly > beginning to change, Berlin-based company plista for examples offers a > recommender contest, where you can test your algorithm in their live > system: > > http://contest.plista.com/ > > --sebastian > > > 2013/7/23 Ted Dunning <ted.dunn...@gmail.com> > > > On Tue, Jul 23, 2013 at 6:07 AM, Jayesh <jayesh.sidhw...@gmail.com> > wrote: > > > > > > > > > > > I have been reading about CF algorithms. Everyone seems to be taking > the > > > preference value as ratings, or any singular attribute. However, in a > > > typical ecommerce scenario the entire clickstream data is important ( > > with > > > varying weights) to determine the affinity of the user vs item. > > > > > > > Yes. This is the literature, but it is the opposite in practice. > Ratings > > rarely convey as much information as the much richer and more voluminous > > stream of implicit data. > > > > Even worse, almost all academic research ignores the fact that multiple > > kinds of behavior is involved in a real system. > > > > Check out my talk at Buzzwords for a possible solution for you. > > > > > > > If we consider many parameters, do we use any kind of a regression to > > > formulate the affinity score (that takes into consideration all the > > > features and their respective weights that impact the users liklehood) > > and > > > run any CF algorithm over these scores? > > > > > > > Bayesian bandit is what I would recommend. > > >
