From: sebastianstern@wanadoo.nl
Ted Dunning:
| You are starting in pretty much the standard place, but your first
| step from there is not quite what is done in a real recommendation
| system.
|
| The thing you are trying to predict is whether somebody would buy
| something, not whether they have it or would want it. This sounds
| like a trivial difference, but it has massive implications.
No, my system is not necessarily for selling stuff; it should/would work on
_any_ kind of objects; e.g., it could predict the degree of preference for
poems based on your preferences for other poems, or the degree of preference
for paintings based on your preferences for other paintings.
| First of all, purchase is observable, preference is generally not.
I solve this by letting the user demonstrate his preference by repeatedly
selecting one object out of two, or by assiging different grades or monetary
amounts to different objects.
| Secondly, you don't really care if you are predicting future
| purchase from past purchases or the phase of the moon. Any feature
| that works is good (to quote the Butch Cassidy movie, there are no
| rules in a knife fight).
|
| Thus, you are trying to predict proclivity to purchase given
| *everything* you know about the person and *everything* you know
| about people in general. Moreover, in any given moment, the person
| is fixed and the items to be recommended are variable. This limits
| the problem somewhat.
Yes, any information I can extract from the subject (besides demonstrated
preference as described above) can be used to predict his preferences to
some extent.
| From there, the decision comes down to a simple financial decision
| of what item has a higher expected profit. Advanced systems of this
| sort also put a value on knowledge so they might recommend something
| as a probe so that they can learn about the behavior of your kind of
| people with that kind of item.
|
| In my own work on the subject (at Musicmatch where we tripled
| revenues by optimizing the sales process and recommendations), I
| found it best to take a radically different approach. The
| difference was that we took the point of view that what matters is
| not the instantaneous expected return, but rather the overall life
| cycle revenues (profits). For example, you might find that if you
| offer a cardboard cutout of a Mercedes for $1000 that you get lots
| of short term sales because on your web-site nobody can tell that
| it isn't a real car. The long-term view of the net present value
| of this offer would, however, indicate that it is a complete
| disaster due to returns, customers never returning and the cost
| of litigation. The life-cycle view accounts for these effects.
In these paragraphs you are again assuming the system should work for
selling stuff, but this is not necessarily the case. E.g. poems, paintings,
faces, sounds do not have 'revenues'.
| The overall problem is extremely difficult, especially when you
| have low data rates and limited manpower for analysis and
| optimization. There are effective solutions, however.
Although I suspect the methods you use are not exacly what I am looking for,
I would be grateful if you could point me to some references on this subject
(books, papers, etc.).
--
Sebastian Stern
Freedom is the freedom to say (= (+ 2 2) 4). If that is granted, all else
follows.
[ comp.ai is moderated. To submit, just post and be patient, or if ]
[ that fails mail your article to , and ]
[ ask your news administrator to fix the problems with your system. ]
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
