Professional Documents
Culture Documents
Henry Blanco Hblancolores@stud-Inf - Unibz.it Francesco Ricci Fricci@unibz - It
Henry Blanco Hblancolores@stud-Inf - Unibz.it Francesco Ricci Fricci@unibz - It
satisfying 𝚽u.
A query is represented as a Boolean vector: q = (q1 ,..., qn )
5. The AdviseSet is finally filtered by just keeping the top K queries
A user profile is defined as: w = ( w1 ,..., wn ) with the largest expected utility.
The utility of a query q for a user w is computed as: 6. The above steps are repeated until no more suggestions can be
n
Utilw (q ) = ∑ wi × f i (q ). ( fi(q) is a value function on qi ) generated or the user does not want more suggestions.
i =1
Results
Identifying a set of profiles close to the true user’s profile is
harder when the users are clustered in several groups and when
the users in each group are more diverse.
Acquiring few well selected profiles every time the system fails
to generate query suggestions yields a faster convergence to better
performances (see Figure 3).
The computation time for generating query suggestions is
roughly 4 times smaller than when a predefined set of uniformly
distributed profiles is used.
Future Work
Bootstrapping the system with a better a priori distribution of
user profiles.
Devising a solution for removing profiles that are not implicitly
shown by the users approaching the system, i.e., managing the
dynamic change of the distribution of the users’ preferences.
Figure 3: Average utility shortfall (difference in utility) in simulated interactions.
ACM Conference on Recommender Systems. October 12-16, 2013, Hong Kong, China.