POI ENHANCED VIDEO RECOMMENDER SYSTEM USING
COLLABORATION AND SOCIAL NETWORKS
Alessandro da Silveira Dias, Leandro Krug Wives and Valter Roesler
PPGC, Informatics Institute, Universidade Federal do Rio Grande do Sul, Av. Bento Gonçalves, 9500, Porto Alegre, Brazil
Keywords: Video Recommender System, Points of Interest, Collaboration, Social Networks.
Abstract: Every day, the number of videos available in the world increases. For instance, there is a vast amount of
video websites, like Youtube and NetFlix, VOD services, as well as PVR devices that automatically record
videos, 24 hours a day. Apparently this situation allows a large possibility of choice for the user, on the
other hand, it creates an overload problem, i.e., a difficulty to find the correct content for the user needs.
One of the ways to treat such an overload is the use of recommender systems, which filter the content in
order to deliver what is most interesting to the user. This paper presents an approach that allows the
annotation of points of interest on videos on the Web. Through this, users can mark their most interesting
points on the videos. This information can thus be used in conjunction with the user profile and interests to
provide recommendations. The differential of this paper is to show how points of interest can be used to
enhance video recommender systems and how to design social networks of users with common interest
points.
1 INTRODUCTION
Every day the number of videos available in the
world increases. For example, there is a vast amount
of video websites (e.g. Youtube, NetFlix, TerraTV,
etc.), and VOD services (Video On Demand), as
well as devices that automatically record videos,
known as PVR's (Personal Video Recorders), 24
hours a day. Just on YouTube, the leader website of
online video sharing on the Web, 48 hours of video
were daily added every minute on the Website in
2010 (Google, 2012).
It poses an important issue for the user: the
overload of video content. One way to treat such an
overload consists on the use of recommender
systems, which filters the content in order to deliver
what is most interesting to the user.
The typical approach used by filtering systems
consists on a hybrid recommender system, i.e., one
that uses both content-based and collaborative
filtering, minimizing the problems that these
approaches have individually. Additionally, in order
to create new recommendation models or to improve
recommendation, new approaches have been
proposed, such as the use of context information
(Naudet, Mignon, Lecaque, Hazotte and Groues,
2008), information from social networks (Golbeck
and Hendler, 2006), annotations of content with tags
(Hung, Huang, Hsu and Wu, 2008), among others.
In the video recommendation area, recent works
have focused on the annotation of Points of Interest
(POI). Through this, users can mark points on the
video that are more interesting for them.
This paper presents an approach to use POI to
enhance video recommender systems, and shows
how a social network of users with POI in common
can be created and used to perform
recommendations. In such approach, the user
participates more interactively and actively through
collaboration.
The rest of this paper is organized as follows.
Next section presents related work. In section 3,
recommender system's theory and concepts are
presented. Section 4 gives an example of a video
recommender system that is extended in sections 5
and 6 in order to use POI to enhanced
recommendation and to build social networks of user
who have POI in common. In the last section, we
present our conclusions and discuss future works.
2 RELATED WORKS
There are many works about video recommender
717
da Silveira Dias A., Krug Wives L. and Roesler V..
POI ENHANCED VIDEO RECOMMENDER SYSTEM USING COLLABORATION AND SOCIAL NETWORKS.
DOI: 10.5220/0003961307170722
In Proceedings of the 8th International Conference on Web Information Systems and Technologies (WEBIST-2012), pages 717-722
ISBN: 978-989-8565-08-2
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
systems. For instance, Qin, Menezes and Silaghi
(2010) extract information about video’s relationship
using a network formed from reviews left as
comments in videos from YouTube. A network of
videos is created, and it is used as the basis of a
recommender system. Analogously, this paper
shows how to create a network using POI.
Nathan, Harrison, Yarosh, Terveen, Stead and
Amento (2008) present a system named
CollaboraTV, which was made to study new ways to
watch TV and new interaction approaches among
viewers. In this system, the user can put temporal
linked annotations on videos or TV programs. One
kind of annotations is POI. When the user is
watching a video, he/she can mark positive or
negative interest points in the current position of the
video stream. These points are used to build interest
profiles for users and communities of users. The
paper suggests that the most popular positive interest
points of different TV programs can be used to
produce a new program. The system offer
recommendation of the most popular TV programs
and encourages the exchange of recommendations
between users. Their authors suggest that
annotations could be used to generate
recommendations in a future work. In this paper, is
actually presented how to use interest points in
recommendation.
Chakoo, Gupta and Hiremath (2008) exploit the
fact that users tend to like more of particular
segments of the viewed content than the rest. The
extraction of these segments and their information is
used to enrich user experience, improving the
quality of video recommender systems. The user
must mark each segment pointing its beginning and
ending in the video stream. Additionally, the user
can rate each segment. With this information, an
user profile of scene interest is built. Their authors
focus on development of a method to extract scene
segments from the viewed content and on the
presentation of a framework capable of presenting
recommended content visibility. Differently, this
paper focus on the classic item recommendation and
presents and analyzes a generic way of marking
interest points, which can be done not only in
viewed content, but also in content being viewed.
The aforementioned works showed that
interactivity and collaboration can be used to
improve video recommenders systems. In addition,
current research in this field shows that the
combined use of content and collaboration is better.
Therefore, the most-used approach is to combine
content-based filtering (CBF) and collaborative
filtering (CF) in a hybrid approach. For instance,
Lekakos and Caravelas (2008) present experiments
showing that the accuracy of a hybrid approach is
greater than the one using these base filtering
approaches alone. Based on this fact, this paper
shows how to enhance these hybrid recommenders
using POI.
3 RECOMMENDER SYSTEMS
Recommender systems help users identify items of
interest. These recommendations are generally made
by two types of filtering: collaborative filtering (CF)
and content-based filtering (CBF).
CBF takes the descriptions of the previously
evaluated or currently accessed items by the user to
calculate the similarity between items, and, then,
recommend items of interest to the user. This type of
filtering enables personalized recommendations for
users, however it has two main disadvantages
(Nguyen, Rakowski, Rusin, Sobecki and Jain, 2007):
(i) it depends on one objective description of the
items; and (ii) it tends to overspecialize
recommendations.
CF calculates the similarity between users and
recommends items that are liked by similar users.
This uses ratings given by users in the past to find
the best item. It has two main disadvantages
(Sarwar, Konstan, Borchers, Herlocker, Miller and
Riedl, 1998): (i) the early-rater problem that occurs
when an user is the first from his/her neighborhood
to rate an item; and (ii) the sparsity problem that is
caused when there are few ratings for the items.
These filtering types can be combined in a hybrid
recommender system that takes the advantages of
both in order to overcome their disadvantages alone.
4 VIDEO RECOMMENDER
SYSTEMS
Often called movie recommender systems, the most
popular ones apply a hybrid approach. The CBF
component uses descriptive features of the videos.
These features can be, for instance, title, genre,
language, duration, producer, actors and plot
keywords. These features can also be used to
describe the user's preferences in an user profile. In
this case, the recommendation task consists on
matching item features and user preferences.
To make the recommendations a CBF
component evaluates the similarity between the
current item that the user is accessing (or items
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
718
accessed in the past) and other existent items.
Similarity can be measured in different ways. A
simple one checks if the genre of the video is in the
user profile. If so, the item is a candidate to be
recommended. More complex ways use similarity
metrics as the cosine measure. In this metric each
item is represented by a vector, the length of which
is equal to the number of non-unique features of all
available items. The elements of the vector state the
existence or non-existence (boolean) of a specific
feature in the description of the item. This metric,
which is presented bellow (1), is used by Lekakos
and Caravelas (2008).
sim
,
=
.
|
|
∗
|
|
=
∑
∑
∑
(1)
In this equation, a
i
and b
i
are the values of the i
th
elements of vectors a' and b'. The result is a numeric
value used to calculate the of CBF component
prediction.
The CF component uses user ratings of items.
The typical approach used is a neighbourhood-based
algorithm, divided in 3 steps: (a) computation of
similarities between the current user and other users,
(b) neighbourhood development, and (c)
computation of prediction based on weighted
average of the neighbours' ratings on the target item.
For the first step, typically the Pearson's
Correlation Coefficient is used. In (2) presented the
adapted Pearson's Correlation Coefficient utilized in
Lekakos and Caravelas (2008) is presented.
,
=
∑
∑
∑
∑
∑
∑
∑
(2)
In this equation, X
i
and Y
i
are the ratings of users X
and Y for item i, and X, Y refer to the mean values
of the available ratings for the users X and Y. At the
neighbourhood development step, neighbours with
positive correlation to the active user are selected.
Finally, to compute an arithmetic prediction for an
item, the weighted average of all neighbours' ratings
is computed using following equation (3).
=
+
∑
∈
∑
|
|
(3)
Here, K' is the average mean of usr’s ratings, J
i
is the
rating of neighbour J for the item i, J is the average
mean of neighbour J’s ratings and r
Kj
is the Pearson
correlation measure for the user and neighbour J,
and K
i
is the prediction for item i in the CF
component.
A hybrid approach can be applied by combining
the CBF and CF components using some of the
hybridization methods. For instance, the Switching
method (Burke, 2007) can be used. Both components
can make its predictions and a switching criterion
choose the prediction that will be used to make the
recommendation. This switching criterion can be, for
instance, the number of CF predictions: if it is above a
threshold the CF component will be used, otherwise
the CBF will be used.
5 POI VIDEO RECOMMENDER
SYSTEM
Natan et al. (2008) and Chakoo et al. (2008) showed
that interactivity and collaboration of users can be
used to improve a recommender system. The first
allow users to mark their current position on the
video stream with positive or negative interest
points. The second allows users to mark segments on
a viewed video. The user tends to like particular
segments of the video more than the rest. In the
remaining of this paper, a segment of video will be
called as POI.
A POI of an user in a video can have intersection
with a POI of another user or users. When the
number of intersections within a single video
exceeds a certain threshold, and it occurs between
different videos among couples of users, it is
suspected that these users have common interests, or
similar taste, about videos. Faced with such
evidence, the following proposition is released:
"interest points can be used to find similar people, or
who have common interests, or taste like, about
videos, and this similarity can be used to enhance a
video recommender systems". One of the goals of
this work is to verify if this proposition is true.
Figure 1 shows how the POIs can be arranged
along a video timeline. In this example, four users,
identified by U
1
, U
2
, U
3
and U
4
, share their interest
points about one video. The interest points of some
users have intersections with the interest points of
other users. Above a threshold of intersections, it is
said that users are similar. Although to have more
certainty on this similarity, that analysis should be
extended to different videos, also considering a
certain number of videos.
Figure 2 shows the analysis on an extended set of
videos. In this figure, it is observed that users U
2
and
U
4
are the most similar in the group of users for a
minimum of three videos and two intersections (i
m
)
in each video (video
n
).
It is also possible that users have not marked any
interest points in a video or have never seen the
video. It is also possible that when re-watching a
POIENHANCEDVIDEORECOMMENDERSYSTEMUSINGCOLLABORATIONANDSOCIALNETWORKS
719
video, the user mark new interest points or want to
clear existing ones.
Figure 1: users’ POI in a video and their intersections
showing common interests between users.
Figure 2: users' POI in different videos.
5.1 Enhancing a Video Recommender
System with POIs
To calculate the similarly between users, a utility
function will be used. This function will use the
intersection on interest points about video of a
couple of users to calculate a value between 0 and 1
that correspond to the degree of similarity between a
couple of users. When the value is 0 the two users
don’t have similar interest, or taste, about any video;
when the value is 1 the two users have the highest
degree of similarity of interest, or taste, about a
video. This function will be used to enhance the
similarity calculated by the CF component, using
equation (4).
,
=
,
∗1+
,
(4)
Here,
is the enhanced similarity of CF
component,
is the similarity of the CF
component and
is the utility function. If the
value computed by the utility function is 0 the CF
similarity doesn't change; if this value is greater than
0 it acts increasing the
as percentage value.
5.2 Marking POIs
POIs are marked by the user while he/she watches
the video. The user interface of the system must
provide a widget with options to collect POIs.
For instance, the webpage showed in the Figure
3, that was built to make a preliminary study about
POI marking in this work, can be used to mark POIs.
In this page when the user starts the presentation of
the video a new area is displayed, where there are
options to mark POIs. It can be done through 3
buttons ("10 seconds", "20 seconds" and "30
seconds"). When the user click in one of this
buttons, for instance "20 seconds", a POI is marked
from the current position on the media stream until
the point 20 seconds before the current position of
the media stream. For instance, if the current
position of the media stream is 50 seconds and the
user click on the button "20 seconds" the POI will
start at position 30 seconds and finish at the position
50 seconds. The POI can be logged using an
asynchronous HTTP request running on background.
This approach to mark POIs is fast and easy to do
and it doesn't need to come back the media stream.
Figure 3: A webpage where the video is presented, and the
user can mark POIs.
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
720
(In the preliminary studies made in this work about
POI marking people marked interest points seconds
after it occurs on the video stream and with no long
duration, and without losing the attention on video
due to interaction. So, it is expected that 10, 20 and
30 seconds could be adequate to a complete study.)
6 POI SOCIAL NETWORK AND
COLLABORATION
Social networks are composed of users who have
common interests and share information among
themselves. Users can share experiences, interact
with others, learn and disseminate knowledge
(Zanda, Menasalvas and Eibe, 2011).
Through the similarity between users calculated
by the intersection of interest points, users with
similar interest or taste about videos can be found.
Based on this evidence the following proposition is
released: "the similarity between users based on
interest points, which must be expressed in
numerical value, can be used to build a social
network. Through this network the person can meet
others similar to him/her and receive (or exchange)
more precise recommendations, person to person".
Other goal of this work is to verify if this
proposition is true.
A POI social network is presented in the Figure
4. The nodes are users; an edge is established
between two users if they have at least one
intersection of interest point. The user is the central
node, the rest of nodes are the existing users similar
to he or she. They are connected by edges. Each
edge has a numeric value, calculated by the utility
function (Section 5.1), and its value correspond to
width (weight) of the edge between the nodes. Users
with higher degree of similarity are more closer, and
with lower degree are more distant at the social
network graphic presentation.
An user viewing your social network can meet
users like he/she, and seek advice, or even, exchange
recommendations directly with another user. For
example, U
1
could access the U
16
and see the "Top
10 rated videos by U
16
," could "exchange
recommendations by chat or e-mail," or even "see
what the user is watching now" if the user U
16
is
online (and permits it). This approach employ a
simple yet powerful mechanism, follow the
Informational Social Influence that tells to user that
when he or she do not know what to do, he or she
often times copy other users (Aronson, Wilson and
Akert, 2005).
Figure 4: A POI social network formed by users with
interest points in common about video.
For instance, the webpage showed in the Figure
5, that was built to make a preliminary study about
POI social network presentation in this work, can be
used to present a POI social network. When the user
access other user the system shows a tooltip with its
user profile (photo, name, age, gender, country, city,
status) and some options to interaction: see the "Top
10 best rated videos", "chat" and "e-mail" to
exchange recommendation of videos, and "watching
now" to see what video the user is currently
watching. This three last option are active when the
user is logged in the system.
Figure 5: A webpage with a POI social network and
options to access or exchange video recommendations
with other user with similar interests about video.
7 CONCLUSIONS AND FUTURE
WORKS
This paper presented an approach that allows the
annotation of points of interest (POI) on videos on
the Web. Through this, users can mark their most
interesting points on the videos. This information
can thus be used in conjunction with the user profile
and interests to provide recommendations.
Moreover, this paper showed how POI can be used
to enhance video recommender systems and how to
create social networks of users with common interest
points and use it to perform recommendations. In
this approach, the user participates more
interactively and actively through collaboration.
POIENHANCEDVIDEORECOMMENDERSYSTEMUSINGCOLLABORATIONANDSOCIALNETWORKS
721
Furthermore, this approach can be applied in
different scenarios, and not only on the Web, but in
video on demand (VOD) services and on devices
such as Personal Video Recorder (PVR), on TV and
in mobile devices.
An important point is the POI marking, it is
influenced by the system design, and especially, by
the design of interface and of interaction.
For future works a prototype and the POI utility
function will be developed to validate the
propositions made. YouTube will be used as a video
source through its integration API. If the
propositions are positively confirmed, the approach
will be extended to other types of media (audio, text,
picture and TV program). In the case of audio and
TV programs the same approach used to mark
interest points on video can be used in audio and TV
programs. In the case of text the approach to mark
interest point will be made similarly as it is done
with a highlight text pen. In the case of one picture
the approach to mark interest points will be made by
the delimitation of interest regions on the picture.
Other approaches to mark interest points in video
will be tested as mark the beginning and the ending
of interest point in the video stream, that is indicated
to environments where the user can rewind the video
or review its, as on Distance Education.
ACKNOWLEDGEMENTS
This work is partially supported by CNPq (Brazilian
Council for Scientific and Technological
Development), and CAPES.
REFERENCES
Burke, R. (2007) Hybrid web recommender systems. The
Adaptive Web, Lecture Notes In Computer Science
4321, 377-408. Retrieved from ACM Digital Library.
Chakoo, N.; Gupta, R.; Hiremath, J. (2008) Towards
Better Content Visibility in Video Recommender
Systems, Frontier of Computer Science and
Technology, 2008. FCST '08. Japan-China Joint
Workshop on, 181-185, doi: 10.1109/FCST.2008.36
Golbeck, J.; Hendler, J. (2006) FilmTrust: movie
recommendations using trust in web-based social
networks, Consumer Communications and Networking
Conference, 2006, 282- 286, doi:
10.1109/CCNC.2006.1593032
Google INC. (2012) YouTube Statistics. Retrieved
January, 30, 20121, from http://www.youtube.com/
static?template=+press_statistics&hl=en
Hung, C.; Huang, Y.; Hsu, J. Y.; Wu, D. K. (2008) Tag-
based User Profiling for Social Media
Recommendation, Workshop on Intelligent Techniques
for Web Personalization and Recommender Systems;
the 23rd AAAI Conference on Artificial Intelligence,
USA, Jul, 2008. Retrieved from AAAI Digital Library.
Lekakos, G.; Caravelas, P. (2008) A hybrid approach for
movie recommendation. Multimedia Tools Appl. 36,
55-70, doi:10.1007/s11042-006-0082-7
Nathan, M.; Harrison, G.; Yarosh, S.; Terveen, L.; Stead,
L.; Amento, B. (2008) CollaboraTV: making
television viewing social again. 1st International
Conference on Designing Interactive User
Experiences for TV and video, 2008. UXTV '08. ACM,
85-94, doi:10.1145/1453805.1453824
Naudet, Y.; Mignon, S.; Lecaque, L.; Hazotte, C.; Groues,
V. (2008) Ontology-Based Matchmaking Approach
for Context-Aware Recommendations. International
Conference on Automated Solutions. IEEE Computer
Society, 218-223, doi: 10.1109/AXMEDIS.2008.38.
Nguyen, N. T.; Rakowski, M.; Rusin, M.; Sobecki, J.;
Jain, L. C. (2007) Hybrid filtering methods applied in
web-based movie recommendation system. 11th
International Conference, KES 2007 and XVII Italian
workshop on neural networks conference on
Knowledge-based intelligent information and
engineering systems: Part I (KES'07/WIRN'07), 206-
213, doi: 10.1007/978-3-540-74819-9_26.
Sarwar, B.; Konstan, J.; Borchers, A.; Herlocker, J.;
Miller, B.; Riedl, J. (2008) Using Filtering Agents to
Improve Prediction Quality in the GroupLens
Research Collaborative Filtering System. Conference
on Computer Supported Cooperative Work -
CSCW’98, 1–10, doi: 10.1145/289444.289509.
Qin, S., Menezes, R.; Silaghi, M. A (2010) Recommender
System for Youtube Based on its Network of
Reviewers, Social Computing (SocialCom), 2010
IEEE Second International Conference. 323-328, doi:
10.1109/SocialCom.2010.53
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
722
