What you choose to present on the absence of history is a very similar
problem as what you present when you have just a little information.
The best case is when the recs engine always gives you the best it can
as well as giving you some indication of how much it knows. Then the
presentation layer can make decisions about what to show.
One way to look at recommendations is as a form of inference from the
behavior of the mass of users and the possibly limited information
about the user.
Regarding the best recommendation with empty history, no I don't think
that the most popular items are a good choice. You do much better to
consider the value of eliciting interesting information from the user
by giving a variety of items. The most popular items tend to be much
too homogeneous to be a good recommendation. If they don't like the
first they won't like the rest. Better to just show one item.
On Jan 22, 2009, at 11:26, Sean Owen <sro...@gmail.com> wrote:
Interesting indeed... I suppose I had thought of that as something a
bit separate from recommendation. You're really just giving, what, the
top-rated items or something? perhaps something more appropriate given
the domain. Either way it feels like something the caller can come up
with outside the context of a recommendation engine. I had thought it
less surprising for the engine to say nothing when it can't actually
recommend something meaningfully.
I suppose DataModel could expose getTopRatedItems or something?
On Thu, Jan 22, 2009 at 7:23 PM, Ted Dunning <ted.dunn...@gmail.com>
wrote:
ANother option is to actually recommend something for new users.
