Discussion Notes

Jan 29, 2001

(courtesy Rob Capra with some help from Saverio Perugini's notes)

Ways of Targeting

      Individual  |  Topic   |  Location
    --------------+----------+------------
        single    |  single  |   single
                  |          |
        group     |  group   |   group
                  |          |
         all      |   all    |    all

Some examples:

individual, group: amazon.com (probably, from what we know)
location, group: pizza hut, movie listings that are routed based on zip code

Why Target Users?

cross-sell (i.e. sell beer to diaper buyers)
up-sell (i.e. just $2 more for two extra)
fleece-sell (i.e. sell tires cheap, but charge a lot to install them)
turn browsers into buyers

Business people may be referring to "versioning" when they talk about targeting. Versioning has to do with selling different versions of the same basic product (i.e. hardcover, softcover, and eventually bargin books).

One approach to targeting is to think of recommendation as filling in missing values to a people-artifacts matrix; this gives rise to the functional modeling viewpoint. An alternative viewpoint is:

Data Mining

Looking for things that go together
It is sometimes difficult to find the causality - may find that beer and diapers are purchased together, but is it beer->diapers or diapers->beer?
Interpreting results of data mining can be tricky
Useless info: `taxicabs return a lot of brake pads'
Obvious info: `everyone who gives birth is female'
Suspicious info: `people good at knitting tend not to have the Y chromosome'
Dangerous info: incorrect inferences (see coffee and tea example below)

Need to make sure we have all the information before drawing conclusions

Example about coffee and tea (from Brin, Motwani, and Silverstein)


                   |  drinks coffee    |  no coffee   |  sub-totals
   ----------------+-------------------+--------------+--------------
     drinks tea    |         20        |      5       |      25
   ----------------+-------------------+--------------+--------------
       no tea      |         70        |      5       |      75
   ----------------+-------------------+--------------+--------------
      sub-totals   |         90        |      10      |     100

80% of tea drinkers also drink coffee, but the apriori probability of being a coffee drinker is 90%. So being a tea drinker actually decreases the chances that a person is a coffee drinker when compared to the prior probability of being a coffee drinker. Rather than target tea drinkers for a coffee promotion, we could have done better by not targeting at all. Moral of the story: we need to look at all the information.

Caution in making constructive inductions

Complex relationships/meta-information -- the at least two example:


       beer    chips    diapers  |    salsa
     ----------------------------+------------
         1       0         1     |      1
         1       1         0     |      1
         0       1         1     |      1
         1       0         0     |      0

Q: What predicts purchasing salsa?
A: Purchasing at least two of {beer, chips, diapers}

Computers are not good at this type of meta-reasoning. They typically only look for what they are told. If they are not told to look for a type of relationship, they will probably not find it.

Example from biology: Crick and Watson looking at how combinations from 4 base pairs form 20 amino acids (which then go on to encode protiens). They correctly inferred that there must be at least 3 base pairs to get 20 > 16 combinations, but incorrectly inferred other aspects of the encodings. Considered a challenge problem for constructive induction.

Are there RS designs that work for all ways of targeting?

(Can we design an RS independently of a (decided) way of targeting?)

Is this desirable, first of all?
Note: This doesn't imply that we don't think of the user (or user class)
Is there such a system out there? Is there a system that can work at multiple levels of the `populace'?
Observation: It is hard to design a system without thinking of (and centering the design around) a specific user class.
Enumerate all the scenarios ("ensemble", "voting") - this allows the user to identify themselves. Examples include the VT web site (for students, for faculty, etc.) and the Dell computer web site (for Home users, for Business users, etc.).
Non-mutually exclusive classes: To really achieve this, we need the ability to put something in multiple clusters at the same time. The clustering algorithm used should allow for one thing to truly be in more than one cluster, simultaneously.
Roles: users may use the system in many different roles at different times. They may be working in an "individual single" role at one point, but then change to being in a "topic group" role at a later point. Determining what role the user is in at a particular time can be challenging.
Dynamic: the system needs to be able to dynamically adapt to new information. For example, the system may start out with the user in an all category and then move the user into a "group" after gaining more information.
Context: using context can help determine what role the user is in for the current interaction.
It appears that if you are not doing enumeration, you need these last four (ability to put something in multiple clusters, roles, dynamic, and context).