MoRe: A USER CONTROLLED CONTENT BASED MOVIE
RECOMMENDER WITH EXPLANATION AND NEGATIVE
FEEDBACK
Oznur Kirmemis and Aysenur Birturk
Computer Engineering Department, METU, Ankara, Turkey
Keywords: Movie Recommendation, user model, open user profile, trust.
Abstract: Recommendation systems have become a popular approach for accessing relevant products and information.
Existing approaches for movie recommendation systems are insufficient, because they do not provide
transparency to the users through enabling them to view and edit their profiles. In addition, negative
feedback, which is an important clue for the recommender, is not taken into account. In this paper we
concentrate on the ideas of automatically generating user profiles from the user’s item preferences, and
enabling users to view and edit their profiles to get satisfaction. In addition, taking negative feedback for
specific values is examined and discussed, which is observed to produce more accurate recommendations.
The system also provides the explanations for the produced recommendations and allows users to modify
their profile accordingly and see their modifications’ effects on the results directly. Initial experimental
results demonstrate that the system produces accurate recommendations and gets user trust and satisfaction
with the transparency and explanation facility.
1 INTRODUCTION
Recommender systems help people to overcome
information overload problem through providing
personalized suggestions based on previous
examples of a user’s likes and dislikes. Several
approaches have been developed in both the research
and business fields (J. Alspector, A. Kolcz, N. &
Karunanithi 1997, P. Resnick & H. R. Varian 1997).
Two main effectively used approaches are the
collaborative filtering (D. Billsus & M.J. Pazzani
1998) and the content-based approach (M.
Balabanovic &Y. Shoham 1997, R. J. Mooney & L.
Roy 2000). Several methods have also been
investigated for integrating both methods in order to
combine the advantages of each approach.
It is not possible to draw recommendations
without taking sufficient user feedback. However,
the information required for drawing
recommendations should be as least as possible so
that users will not be overloaded. We observe that
recommendation systems that require the least input
from the user while providing useful
recommendations are rated the most satisfying. The
recommendation process should be automated
enough to draw a user profile directly from the
user’s preferred items, because this is the easiest
information that the user can supply. Users can
guide the system accordingly afterwards, however if
the recommender will not allow them to view and
update their profile, they would feel frustrated and
give less trust to the recommender.
We believe that transparent user profiles and
allowing users to edit them can provide more trust in
recommenders. Open and editable user profiles
allow users to change their profile to provide
missing information and to correct errors in their
profiles. In addition, negative feedback is also an
important correction mechanism. If users are
allowed to provide which data they would not want,
this will be used to filter out many candidate
suggestions.
There are very few examples of open and
editable user profiles and almost no reported studies
of open profiles for the movies domain that
combines automated profile construction mechanism
with transparency. MetaLens, a movie recommender
system, where user feedback is taken for different
dimensions is investigated in (Schafer, J.B.,
Konstan, J.A., & Reidl, J 2002). They take the
values the user prefers for different dimensions, but
271
Kirmemis O. and Birturk A. (2008).
MoRe: A USER CONTROLLED CONTENT BASED MOVIE RECOMMENDER WITH EXPLANATION AND NEGATIVE FEEDBACK.
In Proceedings of the Fourth International Conference on Web Information Systems and Technologies, pages 271-274
DOI: 10.5220/0001515702710274
Copyright
c
 SciTePress

not a negative feedback for different values of every
dimension.
The incremental-critiquing approach proposed by
McCarthy at (Kevin McCarthy, James Reilly,
Lorraine McGinty, & Barry Smyth 2005) describes a
system where the user has the option of directly
updating the query of selection for the candidates.
However, the system builds the implicit user model
incrementally through taking user feedback, and it
does not include an automated user profile
construction mechanism through item similarities,
which is a desirable feature for end-user satisfaction.
In this paper, we describe MoRe, a movie
recommendation system, with particular emphasis
on how to draw and update the user profile
automatically and correct them through the user
control. The remainder of the paper is organized as
follows. Section 2 introduces our adaptive movie
recommender. Section 3 presents the evaluation
mechanism. Section 4 contains the concluding
remarks and discusses topics for further research.
2 MoRe: A USER CONTROLLED
MOVIE RECOMMENDER
2.1 Presentation of User Controlled
Movie Recommendations
MoRe is a system for personalized movie access.
Like many other content-based recommenders, the
system constructs a user model representing user’s
interests from a set of evaluations. The evaluations
are presented to the system in the form of ratings.
MoRe assembles its content from IMDb
(www.imdb.com) and it periodically gathers new
movie data and forms item profiles accordingly. The
user preferences are kept in the form of ratings in the
user profiles. A new user has to rate at least 10
movies initially. The ratings are presented as one to
five stars. From those ratings, the system constructs
a user model. After this step, there are 3 different
views, namely, Ratings View, Profile View and
Recommendations View.
From the Ratings View, users can see the movies
that they have rated, and rate new movies. The
personalization in the system is fully dynamic, so
whenever the user rates new movies, user model is
updated accordingly.
From the Profile View shown in Figure 1, users
can observe and update their profile. In this view,
first the dimensions and their scores are presented.
Users can provide negative feedback for the
dimensions and update dimension scores. Details of
every dimension are displayed under the dimension
scores.
In the Recommendations View shown in Figure
2, users can observe the produced movie
recommendations. Explanations that describe the
reasons for why those movies are recommended are
also provided. Users can view the explanation,
update their profile data accordingly, and can
immediately see the new adapted recommendations.
2.2 Construction and Update of User
Profiles
User profiles are constructed from the ratings and
updated through the User Profile view. Content-
based approach is used to form user profile. The
content of each selected movie is represented as a
weighted vector of dimension values. By processing
the content data of the rated movies content matrix;
}1,1,1;{
,,
FskDniNmmcC
kim
<
<<<
<
<
=
is formed. Here, Nm is the number of movies
rated,
Dn is the number of dimensions that are
taken into account in the calculation process, and
Fs is the number of values for each dimension.
Users’ interest profiles are also represented as
content matrixes. The content matrix is modified
whenever users modify their profile model or rate
new items. When it comes to produce
recommendations, the cosine similarity measure (B.
Sarwar, G. Karypis, J. Konstan & J. Riedl 2000) is
used to calculate the similarity between user profile
and the item profiles.
2.3 Negative Feedback and Open User
Profiles
The user model in MoRe is not only open but also
editable. Each score for every value of dimensions
and dimensions their selves can be modified. In
addition, users can add new values and dimensions if
the rated movies do not fully reflect the users’
interests. Negative feedback facility is integrated
into MoRe which increases accuracy and user
satisfaction. For instance, if casting is a dimension in
a user profile, then its data including specific scores
for actors and actresses are shown in Casting tab.
User can delete an actor here in order to provide
negative feedback for him. In addition to this, he can
add new actors and actresses together with their
scores. Negative feedback is very valuable for the
system accuracy and performance. For instance,
think about a scenario where the user likes Mel
WEBIST 2008 - International Conference on Web Information Systems and Technologies
272

Figure 1: MoRe interface with the User Profile View.
Figure 2: MoRe interface with the Recommendations View.
Gibson a lot, and rates all Mel Gibson movies with
high scores. However, most of those rated movies
are comedy films. In that case, the recommender
will think that the user likes comedy movies and will
base its recommendations on this figured out fact,
which will lead to inaccurate results. In order to
solve that problem, negative feedback feature is
integrated in the system. For the scenario above, the
user can provide negative feedback for comedy
feature of genre dimension. In addition to this, he
can specify that he does not want a dimension to be
taken into account while producing
recommendations. The personalization in MoRe is
fully dynamic in the sense that each change in User
Profile View or Ratings View (e.g. rating a new
movie, updating the score of a dimension, etc.)
causes the recommendation list to be updated on the
fly. Thus users can examine the effects of the
changes immediately after an update to the profile
data, which we expected should lead to the
improvement of the whole recommendation process.
2.4 Explanation Facility
Explanations of recommendations can be examined
from Recommendations View. Explanations are
provided in natural language so as to make this view
more user-friendly. Dimensions and dimension
values are displayed as explanation. For instance, in
Figure 2, first recommended movie “Red Dragon” is
suggested because of the Genre and Casting
dimensions, where specific values of these
dimensions are Horror, Antony Hopkins and Edward
Norton. In this view, when the user clicks one of the
explanations, details of the explanations are
displayed in terms of dimension and value scores of
these recommendations.
MORE: A USER CONTROLLED CONTENT BASED MOVIE RECOMMENDER WITH EXPLANATION AND
NEGATIVE FEEDBACK
273

3 EVALUATION
For the evaluation of MoRe, we used MAE (Mean
Absolute Error) (Thomas Hofman 2003) metric and
conducted a user study. We used MovieLens dataset
(www.grouplens.org) in order to conduct our tests.
Discussion of this evaluation process is given in
section 4. We automatically measured the accuracy
of the pure content-based recommender with MAE,
where no user interaction took place. Then we
evaluated both this simple version and full version
through a user study.
In order to find out the value of MoRe’s open
user profile and negative feedback features, an
experimental study is performed. In conducting our
user study, we examined the ideas presented in
(Kirsten Swearingen & Prof. Rashmi Sinha 2000).
At the evaluation phase, we concentrated on the
usability and usefulness factors in user satisfaction.
We attempted to confirm three hypotheses in the
study:
H1: Recommendation systems that require the
least input from the user while providing useful
recommendations (according to the user) are rated
the most satisfying.
H2: Users prefer transparency of their profiles
through the recommendation process.
H3: Negative feedback increases the accuracy of
the produced recommendations.
4 CONCLUSIONS
In this paper, a content-based approach to movie
recommendation, with open user profiles and
negative feedback facility is presented. Our main
concern was to increase user satisfaction and trust to
the recommender through providing transparency of
their profiles, explanations of the produced
suggestions and allowing them to update their
profile information with negative feedback. We
believe that transparency and editability of profiles
can be applied to address the problems of trust and
control in adaptive systems.
Currently, we established and examined which
metrics we should use, and how we should conduct
our user study. We make initial tests which have
satisfactory results.
With the MAE metric, we observe that our full
version with the user control performs better than
the simple version, which is as expected, since the
users correct errors in their profiles from the Profile
View. Users find the explanation facility and open
and editable user profiles very satisfying. We
examined the time spent for examining profiles and
explanations and the action taken after this process.
In order to get more valid data to make
comparisons with the existing systems, we have to
examine the system with more subjects, and get user
feedback, after a usage of a long period.
REFERENCES
J. Alspector, A. Kolcz, N. & Karunanithi 1997, Feature-
based and clique-based user models for movie
selection: a comparative stud., User Modeling and
User-Adapted Interaction, Vol. 7, No. 4, pp. 279–304.
M. Balabanovic &Y. Shoham 1997, Fab: content-based
collaborative recommendation, Comm. of the ACM,
Vol. 40, No. 3, pp. 66–72.
Schafer, J.B., Konstan, J.A., & Reidl, J 2002, Meta-
recommendation Systems: User-controlled Integration
of Diverse Recommendations, Proceedings of CIKM-
02 pp. 196-204.
D. Billsus & M.J. Pazzani 1998, Learning collaborative
information filters, In Proc. of the 15th Int. Conf. on
Machine Learning, pp. 46–54.
Kevin McCarthy, James Reilly, Lorraine McGinty, &
Barry Smyth 2005, An analysis of critique diversity in
case-based recommendation., FLAIRS Conference,
eds., pp. 123-128. AAAI Press.
Kirsten Swearingen & Prof. Rashmi Sinha 2000,
Recommending Recommenders: A Comparison of
Human vs. Automated Book and Movie
Recommendation Systems.
P. Resnick & H. R. Varian 1997, Recommender systems,
Comm. of the ACM, Vol. 40, No. 2, pp. 56-58.
R. J. Mooney & L. Roy 2000, Content-based book
recommending using learning for text categorization,
In Proc. of the 5th ACM Conference on Digital
Libraries, pp. 195–204.
Thomas Hofman 2003, Collaborative Filtering via
Gaussian Probabilistic Latent Semantic Analysis,
Proceedings of the 26th Annual International ACM
SIGIR”.
B. Sarwar, G. Karypis, J. Konstan & J. Riedl 2000,
Analysis of recommendation algorithms for e-
commerce, In Proceedings of ACM E-Commerce.
WEBIST 2008 - International Conference on Web Information Systems and Technologies
274