Choice-Based Preference Elicitation

Collaborative Filtering Recommender Systems

approach to interactive recommending combines the advantages of algorithmic techniques with the benefits
of user-controlled exploration
the exploitation of latent fac- tors is a promising means to generate interactive, choice- based
recommendation dialogs, even when no data about the user are available yet.
this allows for producing useful recommendations even in cold-start situations
achieve a higher level of user-control without sacrificing interaction efficiency.

Automatic & Interactive recommender

systems
(Fully automated) You tube / Netflix / amazon help users finding
interesting content at a very low interaction cost because users directly
select items from recommendations shown.
Problem :Lack of transparency for user and it highly depends on
existing user profile(interest) & LESS TRUST in system
On the other hand Interactive search and filter techniques : entirely
user controlled
Problems : Huge search and navigation effort

Ineractive choice based preference

elicitation for CF
New method to achieve trade off between system support and user control

Aim to provide a solution that does not depend on the metadata like tags . They can be used as addition to
just improve items presentation.
User item matrix is required
Does not require previously provided ratings by targe user and also supports cold start situation.
Extracted latent factors from UI matrix and assign all items latent feature values vector
We utilize the vector space spanned by the latent factors as a basis for automatically
generating interactive dialogs that elicit the users preferences with respect to these factors.
a dialog consists of a series of choices between two alternative sets of movies (or a dont care op- tion),
where each decision determines the users position with regard to a single factor.

Comparison with other methods

New methods wins in fit of the resulting items with the users interests (H1), and to the perceived novelty
the methods are also differently well suited if the user does not have a search goal in mind
With respect to control , new approach was superior to automatic . As expected, the condition Manual
achieved a slightly better result here, but the difference was not significant. Conversely, Int scored
significantly better than Manual (and, also Automatic) in terms of effort
Reegarding trust ,,the results were similar to those of user control .
The results suggest that the participants appreciated the new method, and the increased interactivity and
controllability in the recommendation process it offers