Decision Trees and Recommender Systems - Data Mining with Decision Trees: Theory and Applications - page 261

While pairwise comparison is considered to be accurate, it is time
consuming and thus hardly used in real-time applications. Instead of
comparing items, we suggest clustering the items, for the purposes of
acceleration and simplification of the pairwise comparison. In this way we
can trade accuracy with time consumption.
When new users signup to a recommendation service, the system is
oblivious to their preferences. Therefore, the initial recommendations the
system provides are of relatively low quality. This phenomenon is known
as the user cold start problem. A common approach to solve the cold start
problem and to learn the user's preferences is to ask her to rate a number
of items (known as training points). Preference elicitation is the initial
acquisition of this profile for a given user.
It would seem apparent that a user's preferences could be elicited
by simply asking her to define her preferences on various aspects (such
as film genre in the movie domain). However, this simplistic approach
usually does not work, mainly because users have problems in expressing
their preferences and may not have the domain knowledge to answer the
questions correctly. Thus another approach is to ask the user to provide
feedback regarding several items presented to her. Roughly speaking there
are two types of item-based methods: static and dynamic methods. With
static methods, the system manages a seed set of items to be rated by the
newcomer. This set is preselected regardless of the feedback provided by the
current during the elicitation process. The methods are described as static
because they use the same items for all users. On the other hand, with
dynamic methods, the questions are adopted to feedback from the current
user.
This is the place to note that several researchers criticize the idea of
initial questionnaire. According to these researchers, users may lack the
motivation to answer initial elicitation questions prior to any discerned
and concrete benefits. Moreover, user preferences are context-dependent
and may be constructed gradually as a user develops self-awareness. If the
recommender system imposes a prolonged elicitation process just when the
user signs up, the preferences obtained in this way are likely to be uncertain
and erroneous.

16.3.1 Static Methods

Recently, published researches discuss the criteria for selecting the items
about which the user should be queried. Among the ideas that have
emerged are the use of controversial items that are also indicative of their