Database Reference
In-Depth Information
tendencies. The contention of an item is frequently estimated using the
Entropy measure. However, selecting non-popular items may not be helpful
because the users might not be able to rate them. On the other hand,
selecting items that are too popular, may not provide us with sucient
information regarding the user's taste. Thus, in order to steer a clear course
between selecting non-popular items and ones that are too popular, the
appropriate way seems to be to query users about popular yet controversial
items.
Nevertheless, combining entropy and popularity is not always su cient
for learning about the user's general taste. The queried items should also
be indicative to other items.
The use of the three criteria above — popularity, controversiality and
predictability — is not sucient for obtaining a good user profile because
of two main reasons. First, the interactions among the items are ignored.
Namely, two queried items with high criteria values can also be highly
dependent. Thus, querying the user regarding both items will usually
contribute relatively little information compared to asking the user about
only one item.
Dynamic Methods and Decision Trees
16.3.2
There have been very few attempts to dynamically implement the initial
profile process, i.e. adopt further queries to the current user feedback. Most
of the dynamic methods employ decision trees. Decision trees seem to be
more successful for initial profile generation.
The Information Gain through Clustered Neighbors (IGCN) algorithm
selects the next item by using the information gain criterion while taking
into account only the ratings data of those users who match best the target
newcomer's profile so far.
While IGCN is based on a non-tree idea, the inventors of IGCN indicate
that the usage of a decision tree can be easily adopted. Users are considered
to have labels that correspond to the clusters they belong to and the role
of the most informative item is treated as helping the target user most in
reaching her representative cluster.
In the next section, we present one of the recent dynamic methods
that employ decision trees to guide the user through the elicitation process.
Moreover, in all cases each tree node is associated with a group of users.
This makes the elicitation process an anytime process. The essence of an
anytime process is that the user may answer as many questions as she likes.
Search WWH ::




Custom Search