=Paper=
{{Paper
|id=None
|storemode=property
|title=Insights on Social Recommender Systems
|pdfUrl=https://ceur-ws.org/Vol-910/paper8.pdf
|volume=Vol-910
|dblpUrl=https://dblp.org/rec/conf/recsys/NetoN12
}}
==Insights on Social Recommender Systems==
https://ceur-ws.org/Vol-910/paper8.pdf
Insights on Social Recommender System
Wolney L. de Mello Neto Ann Nowé
Vrije Universiteit Brussel Vrije Universiteit Brussel
CoMo Lab CoMo Lab
Brussels, Belgium Brussels, Belgium
wdemello@vub.ac.be ann.nowe@vub.ac.be
ABSTRACT tive Filtering; D.2.8 [Software Engineering]: Metrics—
Recommender Systems (RS) algorithms are growing more complexity measures, performance measures
and more complex to follow requirements from real-world
applications. Nevertheless, the slight improvement they of- Keywords
ten bring may not compensate the considerable increase in Recommender System, Collaborative Filtering, Social Rec-
algorithmic complexity and decrease in computational per- ommenders, Cold-Start Problem, Evaluation Metrics
formance. Contrarily, context aspects such as social aware-
ness are still not much explored. In view of that, this paper 1. INTRODUCTION
proposes insights on how to possibly achieve more efficient
Our generation faces several tough challenges within the
and accurate predictions for recommendations by exploring
current peta-, exa- or even zettabyte information era. Every
multiple dimensions of a RS architecture. A framework is
day we deal with huge amounts of information whose ma-
designed, comprised of a Facebook application called My-
nipulation and storage struggles even on high-end computer
PopCorn and some scenarios of user neighborhood RSs are
technologies. Shifting from the point of view of computer
proposed. The first one investigates how to recommend
capacity to an average single person, the problem gets even
movies based on a narrowed subset of collaborative data,
worse due to human being limitations. Online services are
extracted from the social connections of the active user.
examples of big data resources with increasing importance in
Secondly, connections between users enable a solution for
our lives. About two years ago, Google’s search engine used
the cold-start problem. Preferences from social connections
to process approximately half of the entire written works of
are aggregated, producing a temporary profile of the new
mankind per day [6]. Nowadays, it is impossible to avoid
user. Finally, a third dimension is explored regarding evalu-
such reality while working, studying, and entertaining your-
ation metrics. Results from traditional evaluation by offline
self. Perhaps this information overload comes with high cost,
cross-validation are compared to measuring prediction ac-
nevertheless, high benefit as well.
curacy of online feedback data. These insights propose how
Movie domain is a great context where information over-
community-based RS designs might take advantage of so-
load is a high potential pain point to be explored. Moreover,
cial context features. Results show that all three proposed
Netflix movie streaming service is a good motivation for this
solutions perform better assuming some conditions. Social
work due to two main reasons. Firstly, figures disclosed in
neighborhoods can often provide representative data for col-
[1] mention 75% of their sales come from recommendations.
laborative filtering user-neighborhood techniques, improving
Secondly, [1] reveals the decision of not implementing com-
a lot the RS performance in terms of computational com-
mercially the algorithm with around 10% improvement in
plexity metric without compromising prediction accuracy.
prediction accuracy, winner of US$ 1 million prize[8]. Tak-
Assuming a user has a dense social network, the cold-start
ing these facts into account, what would be the most poten-
problem can be easily tackled. Finally, rating prediction ac-
tial path to explore within the field of RSs? Is accuracy the
curacy performs better when evaluated online than by offline
most important metric to take into account? What about
cross-validation.
computational complexity and transparency? What about
online instead of offline evaluation methods?
Categories and Subject Descriptors Rather than building upon complex RS methods, this pa-
H.4 [Information Systems Applications]: Miscellaneous; per investigates a social framework for developing state-of-
H.3.3 [Information Search and Retrieval]: Collabora- the-art RS. Aiming at current main challenges, this paper
proposes contributions on how to tackle some of its most rel-
evant issues based on possibilities enabled by social context
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are information. The three explored RS challenges are: (i) per-
not made or distributed for profit or commercial advantage and that copies formance issues related to scalability of recommender sys-
bear this notice and the full citation on the first page. To copy otherwise, to tems; (ii) lack of knowledge about new users, known as cold
republish, to post on servers or to redistribute to lists, requires prior specific start problem; and (iii) definition of good evaluation meth-
permission and/or a fee. ods.
Copyright is held by the author/owner(s). Workshop on Recommendation Some insights are discussed based on how social-graph
Utility Evaluation: Beyond RMSE (RUE 2012), held in conjunction with
ACM RecSys 2012. September 9, 2012, Dublin, Ireland. data enable a good implementation of a user neighborhood
. RS algorithm, focusing not only on prediction accuracy but
33
also on other metrics such as scalability, computational com- social connections and building a virtual profile based on ag-
plexity and transparency. These insights lead to 3 hypothe- gregation methods, originally proposed for group RSs. [13]
ses listed below: describes 10 aggregation methods and empirically concludes
that social-based think is the best basis for generating an
i. A user’s social neighborhood is sufficiently representa- artificial preference profile. The author claims that Least
tive to provide efficient, in the sense of computational Misery, Average and Average without Misery are the most
complexity, and effective recommendations, in terms human-like reasoning techniques, achieving very good re-
of prediction accuracy; sults.
Transparency Users eventually question themselves about
ii. Social neighborhood connections can derive assump- the reasoning behind a recommendation. They are more in-
tions about new users taste, avoiding the cold-start clined to accept and evaluate better once they understand
problem; how an item has been suggested to him or her. Neverthe-
less, it is not always possible to provide such a transpar-
iii. Online evaluation of transparent recommendations should ent explanation. [9] presents a survey on content-based RS
be a valid metric within social RSs. and compares them to CF techniques also in terms of trans-
parency. The authors claim CF techniques are a black box,
2. RELATED WORK and it is indeed the truth for most cases. In the case of
user-neighborhood RSs, although RSs could tell to the ac-
In the introduction of the latest survey in RS field, [15]
tive user about people with close taste that influenced the
highlights current challenges for RSs. Some of them are
recommendation, privacy issues may not allow such trans-
investigated hereby, such as follows:
parency. In view of this challenge, this paper counteracts
Scalability In real-world applications, the number of in-
the affirmation made by the previously cited survey. It is
stances might often steeply increase in multiple dimensions
possible to give explanation on user-based collaborative fil-
such as number of users, items and, in turn, user-item pref-
tering technique once one assumes not having privacy issues,
erence signals. Despite being a good scenario for some RS
a tractable scenario within social networks, where connec-
algorithms to achieve better accuracy, bigger datasets may
tions previously agree on sharing some information. Besides
lead to a great increase in computational complexity.
this proposal, some solutions to tackle CF limitations related
[7] proposes an evaluation of top-N recommendation algo-
to transparency are proposed in [4].
rithms. Item-based RS is proposed as an alternative for non-
Evaluation One of the main modules of a RS design,
scalable user-based recommenders, since it performs bet-
evaluation strategy is a critical and subjective aspect to be
ter when there are many more users than items. Some
shaped throughout the whole process of building and main-
other item-based RSs avoiding scalability problems within
taining a RS. Even though most papers adopt accuracy as
memory-based CF algorithms are compared in [16].
the most important metric, one should consider many other
Regarding model-based CF techniques, [17] follows a rea-
evaluation criteria, as presented in [5]. Computational com-
soning that is similar to the solution presented in Section
plexity is one metric highlighted in the insight presented in
4.1, since both look for a narrowed neighborhood which
Section 4.1. Transparency is enabled by social context, as
does not to compromise general performance. Whereas the
discussed in Section 3.1.3. Besides exploring metrics, this
cited papers are based on clustering techniques, our heuris-
paper also focus on questioning methods (see Section 4.3).
tic consists of narrowing the database to a subset of user
Offline and online methods should be compared while mea-
social-graph connections. Although scalability is an intrin-
suring rating prediction accuracy.
sic disadvantage to user-based RS, the proposition of a local
neighborhood might overcome this drawback. User-based
RS is adopted since it enables some features related to the 2.1 Social Recommenders
social RSs, such as transparent explanations for each recom- In view of all issues previously listed and the fact some
mendation; state-of-the-art architectures might not be that attractive
Data Sparsity It is among the main bottlenecks for RSs. for commercial purposes, this paper dives into a RS de-
The lack of information is a big problem, especially during sign that is gaining special attention: Social RSs. Also
first interactions of a new user. This scenario is defined as called community-based recommenders, the basic architec-
the cold start or new user problem, which is traditionally ture embeds context data into either collaborative filtering
solved by requiring initial user information before any rec- or content-based algorithms, improving the RS performance.
ommendation is given. Nevertheless, this interaction is time According to [15], community-based paradigm is still a hot
consuming, since the user has to look for a couple of items topic and it is not possible to find a consensus about whether
to rate. To improve that, [14] has compared 6 techniques social recommenders have better performance. [19] presents
to generate this first list of items, aiming to maximize the a broad survey on social recommenders. One could see so-
percentage of rated items out of all items presented to a new cial data in two ways: (i) unweighted social graph; (ii) or a
user. more complex weighted social-graph. The former has been
Besides requiring this first interaction with the RS, one selected for this paper experiments based on empirical con-
could think of a temporary user profile in order to enable clusions made by [2] while comparing CF and Social Filter-
initial recommendations. [11] explores trust networks and ing. Similarities between friends were in average higher than
propose the incorporation of preferences from trusted users. the same correlation measurement between non-connected
Nevertheless, the new user still has to explicitly provide in- users. Moreover, both weighted and basic social RSs per-
formation about who are his/her trusted users. Our work formed the same or better than pure collaborative filtering
retrieves implicit information from social networks, regard- RSs for the referred case.
less trust measurements. The method consists of retrieving Further than looking at social connections, the latter is
34
Table 1: MyPopCorn and GroupLens datasets.
Users Ratings Movies
MyPopCorn 129 14k 3k
GroupLens 72k 10M 10k
a trust-based RS that focuses on weighted relationships. A
clear comparison between social RS and trust-based RS is
defined in [10]. Moreover, [3] highlights the possibility of ex-
plaining recommendations based on social connections and
the fact active users rate better the RS in case of existing
Figure 1: Social Graph representation of MyPop-
such transparency. Finally, the social RS described hereby
Corn database.
profits from an unweighted social graph.
3. FRAMEWORK 3.1.2 Rating Strategy
As claimed in [15, pg 15], the context in which a RS is In MyPopCorn, the user can choose a rating from 1 to
developed and its expected features determine the optimal 5 ‘stars’. Asymmetric labels were defined for each of the 5
algorithm to be adopted. Parameters such as movie do- stars to achieve a more homogeneous judgment, namely Bad,
main, social community context, rating strategy and sparse Regular, Good, Great and Masterpiece. Test users reported
data were definitely crucial to come up with the final ar- good feedback on the proposed rating strategy claiming this
chitecture described hereby. A Facebook application called discrete labeled design is certainly more intelligible, where
MyPopCorn 1 , the RS front-end, and a social based imple- users can have a hint of what each rating value may rep-
mentation of user neighborhood CF algorithm compose the resent. While following such design, this research aims at
current framework, to be presented in the two following sec- reducing subjectivity that is intrinsic to rating process, the
tions. core interaction responsible for obtaining the main input of
a Collaborative Filtering RS. This strategy also prevents the
3.1 MyPopCorn, a Facebook app as Front-End necessity of the RS to normalize user ratings.
The idea of building this movie recommender system and
making it available on a social network is due to the fact
3.1.3 Recommendation Strategy
social graph enables proposed recommendation experiments Recommendations are generated from two implementa-
based on social neighborhoods. Moreover, the capability of tions of user neighborhood recommenders, such as follows:
recommending to an active user and receiving an online feed- • Provided by a traditional user-based RS. The neigh-
back on rating prediction accuracy on recommended items borhood calculated among all users in the database;
is decisive to benchmark the implemented algorithms.
MyPopCorn is a web movie recommender system. Some • Provided by a social-graph user-based RS. A social
of its interfaces are composed as follows: neighborhood is based on the set of active user friends,
First screen presents a brief description of the main fea- to be described in more details in the next section.
tures before the user joins the application. After that, an A shuffled list of recommendations generated by both RS
active user can check statistics about top users and friends; implementations is presented to the user. Movie description
MyTaste is where a user can rate movies. Recommendation- and a continuous predicted value is presented. Therefore,
wise, this is one of the main interactions with the user, in recommendations are seen as a regression and not a clas-
which RS collects data; sification problem within this framework. Finally, at the
My Friends’ Taste presents a list of friends and their bottom of the frame one can see the explanation about each
respective number of ratings. The more ratings each friend recommendation(see Figure 2). In the first example on light
has, the bigger his or her basket gets. blue background, a message informs the recommendation
was “Based on all MyPopCorn database“. Alternatively, the
3.1.1 Social-Graph Data second message informs that is was “Based on friends with
The first collaborative data with ratings over movies were closest taste“, followed by the list of users Friend X and
taken from GroupLens 10M dataset. From that point, the Friend Y.
database was increased with ratings from users of MyPop-
Corn. Information about users, friendships are also made
persistent into the same database. The dataset used for the
experiments is summarized in Table 1.
In a very short timeframe, the application was accepted
by a good number of users. Almost 130 active users have
been exploring the application during 2 months time. Figure
1 illustrates all users who contributed for the experiments
carried out into this paper. The more movies a user rates,
the bigger the node is represented in the social graph. The
average degree of connections in this graph was 10.543.
1 Figure 2: Recommendation strategy in MyPopCorn.
http://mypopcorn.info/
35
This system is designed to give the most transparent rec- accuracy from social neighborhood recommendations will be
ommendations possible. In view of that, the reasoning be- as much precise as in the standard method.
hind the RS can be better understood by presenting the For the proposed experiment methods, standard neighbor-
real number as predicted rating value. Furthermore, ex- hood RS performs around 70k calculations, the number of all
plaining the recommendation with a list of users will trans- users in the merged dataset. In the case of social neighbor-
form a formerly impersonal recommendation into a social hood, the number of comparisons is relative to the degree of
passive interaction between friends. Due to privacy issues, each node (user) in the social graph, which varies from 0 to
presenting this list is only possible for the social neighbor- 49 for MyPopCorn dataset with an average degree of 10.543.
hood approach, where content sharing among users is agreed Concerning average runtime, whereas prediction process for
in advance. one rating takes around 950.55 ms for standard neighbor-
hood, after narrowing the search space to the set of social
3.2 Movie RS Back-End connections, it takes in average 69.975 ms, 92.63% lower.
The final architecture of the social-graph recommender Regarding accuracy, Figure 3 presents prediction accuracy
was developed on top of the user-based RS implementation error for this new neighborhood compared to the standard
provided in Mahout2 . User neighborhood CF paradigm has implementation. Both implementations were compared by
close reasoning to social user behavior, being the most rel- varying the size of the neighborhood k while experimenting
evant criterion that influenced this design choice. In pos- two values of threshold t=1 and t=2. This threshold defines
session of information about users taste, this user-centered the minimum number users in the neighborhood that rated
method focus on comparing similarity among users. Fur- a same candidate item. When t=2, the items rated by only
thermore, friendship data will be essential to enable modifi- one user in the neighborhood are not taken into account.
cations on the original algorithm. Insights on how to profit
from social context information in different dimensions will
be addressed below.
4. INSIGHTS ON RS CHALLENGES
As the title suggests, solutions to the current RS chal-
lenges listed in Related Work are described in this section.
Each of the following implemented scenarios tackle three
main challenges previously mentioned, namely computational
complexity issues of scalable user-neighborhood RSs; sparse
data about new users, known as cold start problem; and
definition of optimal evaluation methods for transparent and
non-transparent recommendations.
4.1 Social Neighborhood
The idea of narrowing the dataset to a subset of users
aims to tackle scalability constraints and increase real-time
performance, two issues that are intrinsic to user-based RS
[7]. Assuming that calculating an active user’s neighborhood
(comprised of k similar users) among his or her social con-
nections might be representative enough, good recommenda- Figure 3: Standard and Social Neighborhoods pre-
tions could be achieved without the necessity of comparing diction accuracy (RMSE ).
a user preference vector with all other users in the database.
This hypothesis is based on a related work comparing the
correlation between users similarity and the binary fact of The minimum RMSE = 0.8385664 was obtained by Stan-
being or not being friends[2]. It was observed that similar- dard neighborhood (k =3,t=2). Besides that, Social (k =2,t=2)
ities between friends are in average higher than the same achieved RMSE = 1.018598. Surprisingly, rating prediction
correlation measurement between non-connected users. accuracy also improved. Except for values of k neighbors
Experiments were performed in order to investigate the equal to 2 and 3, Social Neighborhood outperforms, in av-
three insights proposed above. A standard user-based neigh- erage, the standard method, confirming the first hypothesis
borhood RS setup is incrementally modified from the current for this scenario. Besides that, the value of threshold t=2
insight until the third one. This scenario focus on predict- performs better. The fact of accepting only items rated by
ing ratings contained in a training set comprised of 5% of all at least two users might have increased the confidence on
14.367 ratings provided by MyPopCorn users. The reason preference data, achieving better accuracy results. On the
for not adding any rating from GroupLens into the training contrary, hypothesis 2 was surprisingly refuted. Instead of
set of the standard neighborhood is allow a fair comparison performing almost the same as in the original approach, So-
between both neighborhoods. By applying two strategies, cial Neighborhood can significantly outperform prediction
namely Standard full neighborhood and hereby proposed accuracy for k > 3. While increasing the value of k, such so-
Social one, some hypotheses are tested: (i) Real-time rec- cial neighborhood enables a more accurate predictions and,
ommendation performance will become much more efficient probably, reaching higher serendipity.
while adopting social neighborhood; (ii) Rating prediction Remark: This approach is not available for people with
no or few friends, suffering from the cold start problem, to
2
Apache Mahout machine learning library be solved next.
36
4.2 Social Aggregation for Cold-start Problem ing no social connections. Prediction accuracy error was
One of the main issues related to RS, the cold-start prob- RMSE = 1.37461.
lem or new-user problem prohibit some active users to re- Compared to the accuracy evaluated in the experiments
ceive recommendations. In the dataset used for all experi- of previous sections (RMSE = 1.173435 for k =4, t=1),
ments, 21 users out of 129 have rated less than 10 movies, this proposed solution to the cold-start problem has de-
while others more than a thousand. These users with few creased performance in around 20%, considering the RMSE
ratings are almost unable to receive any recommendation. = 1.37461. In view of that, the proposed solution is con-
Instead of adopting the classic approaches such as content- sidered to be a good alternative for social RSs. Besides
based or presenting a list to be rated as from the first user not compromising the prediction accuracy significantly, this
interaction, this paper proposes a solution based on social- method should be considered in terms of how efficient the RS
graph information. It is based strategy from group RS based can deal with new users that are not interested in providing
on aggregating user profiles. One could see this problem many ratings as from the first interaction. Despite not be-
following the quote “Tell me who your friends are and I will ing an objective metric, the ability of solving the cold-start
tell you who you are“. This reasoning is also motivated by should be incorporated into RS evaluation.
the work carried out in [2], where social filtering is explored
and conclusions reinforce the suggested heuristic. Likewise, 4.3 Comparison of Evaluation Methods
[?] developed a probabilistic RS and achieved good results While the first insight focuses on the two objective evalua-
in experiments where active users were recommended items tion metrics, namely prediction accuracy and computational
based on the preferences of his or her social connections. complexity, this insight focuses on transparency, a subjective
On the contrary, the idea presented in this paper follows metric, and evaluation methods. The most popular evalua-
the same reasoning of absorbing social context data into tion metric throughout RS state-of-the-art, prediction accu-
the system to solve the cold-start problem, nevertheless, by racy benchmark is often based on offline cross-validation and
different means (based on group RS) and in a different RS error calculation over Root Mean Squared Error - RMSE. In
implementation technique (user neighborhood RS). view of that, this third and last section compares offline and
Among some aggregation techniques mentioned in the Re- online methods of calculating estimation accuracy together
lated Work, Average without Misery is adopted, since it finds with more transparent recommendations based on social ex-
a balance between the Least Misery and Average. It pre- planation. One hypothesis is that this online method might
serves the main advantages of both aggregation strategies make offline approach suboptimal for the context of social
originally applied to group RS and now reflected in the ag- recommenders. Instead of cross-validation, one should con-
gregated virtual profile to be considered by our single-user sider the social factor involved within online evaluation. Due
RS. It follows the human-like reasoning in which a group of to the strategy of recommending a list of movies whose pre-
people tend to select items that please, in average, most per- dicted ratings might not be always high and to make it more
sons involved. Moreover, it excludes items once rated below transparent, the predicted value is presented to the active
a defined threshold, as described by [13]. The same author user. Assuming that not many people tend to converge with
proposed such aggregation for solving the cold-start problem the RS prediction, this strategy will not bias the compari-
in [12], although in a different RS paradigm. Experiments son. Actually, we believe there are people who also try to
were run in order to test the following hypothesis: (i) Rec- diverge from what has been predicted.
ommendation accuracy for aggregated virtual social profile The current experiment intends to test the effect of ex-
performs not much worse than cross-validation of real rat- plained recommendations, as previously described in [18],
ings. Hence, it would be a feasible solution to the cold-start but now in the context of social RSs, as defined in the fol-
problem. lowing hypothesis: (i) Assuming social RSs where recom-
The social neighborhood method was adopted with pa- mendations based on social connections are explained, rat-
rameters k =4 and t=1, so that the most number of pre- ing estimation accuracy achieve better results if evaluated
dictions are enabled. The idea here is to investigate how online, instead of offline.
many active users had the cold-start problem, meaning their Besides RMSE, metrics such as novelty or serendipity were
neighborhoods were empty. While repeating the experi- taken into account while choosing higher values of k other
ments from last section in 5% of MyPopCorn ratings dataset, than the ones that reached minimum accuracy, shown in
around 103 users were in the testset. Nevertheless, RS could Figure 3. Although the same number of recommendations
not estimate any rating for 13 users due to empty neighbor- with standard and social neighborhood were generated, ac-
hood issue. 6 users had no social connections, what can tive users gave more feedback on the social ones. 119 online
not be solved by the method proposed here. The remaining feedbacks were provided, as presented in Table 2 in compar-
7 users had their ratings predicted with accuracy error of ison with the traditional offline method.
RMSE = 1.69588. As Table 2 shows, Standard Neighborhood method achieved
One should raise the question that this is not much data, a prediction accuracy of 1.0646 and Social Neighborhood RS
referring to the tiny set of 7 users. In view of that, an- setup achieved better rating prediction accuracy of RMSE
other experiment has been run on 50% of ratings in MyPop- = 0.9952. Both of them presented an improvement when
Corn dataset. Ratings of 44 users experiencing the cold- evaluated online other than offline. The decrease in RMSE
start problem were hidden iteratively in order to be pre- was of 14.16% and 6.64%.
dicted by the RS. Foreach of the 44 users, the RS generated Hypothesis was confirmed by the numbers shown in Table
a virtual profile based on aggregating all ratings from their 2. Surprisingly, online evaluation accuracy with Standard
friends, including those removed in order to artificially cause Neighborhood improved better (14.16%) than 6.64% gain
the cold-start problem. Only 8 new users(18%) could not be achieved by Social Neighborhood strategy. Finally, results
helped by this method of aggregation due to the fact of hav- have shown that, in average, RSs tend to present better
37
[4] J. Herlocker and J. Konstan. Explaining collaborative
Table 2: Online evaluation of social and standard filtering recommendations. of the 2000 ACM
neighborhood. conference on, pages 241–250, 2000.
Std. N. Social N.
[5] J. Herlocker, J. Konstan, L. Terveen, and J. Riedl.
Setup k =8, t=2 k =4, t=2
Evaluating collaborative filtering recommender
Offline systems. ACM Transactions on Information Systems
RMSE 1.240429 1.066049 (TOIS), 22(1):5–53, 2004.
Online
[6] D. Infographic. Visualizing the petabyte age, 2010.
RMSE 1.064686 0.995211
[7] G. Karypis. Evaluation of item-based top-n
Improvement 14.16% 6.64%
recommendation algorithms. In Proceedings of the
tenth international conference on Information and
knowledge management, pages 247–254. ACM, 2001.
accuracy results in online evaluations than offline for both [8] Y. Koren, R. Bell, and C. Volinsky. Matrix
explained and non-explained recommendations. factorization techniques for recommender systems.
Computer, 42(8):30–37, 2009.
5. CONCLUSIONS [9] P. Lops, M. Gemmis, and G. Semeraro. Content-based
recommender systems: State of the art and trends.
This paper first discussed the computational requirements
Recommender Systems Handbook, pages 73–105, 2011.
intrinsic to user neighborhood RS, by nature a non-scalable
algorithm. Based on the two most important evaluation [10] H. Ma, D. Zhou, C. Liu, M. Lyu, and I. King.
metrics, state space reduction enabled a decrease of 92.63% Recommender systems with social regularization. In
in computational complexity, while not compromising accu- Proceedings of the fourth ACM international
racy. Instead, the latter also improved. conference on Web search and data mining, pages
Social graph was essential to enable a solution to the cold- 287–296. ACM, 2011.
start problem. Tested with success in group RS, Average [11] P. Massa and P. Avesani. Trust-aware collaborative
without Misery enabled creation of virtual profiles based on filtering for recommender systems. On the Move to
active users network. Results confirmed the proposed hy- Meaningful Internet Systems 2004: CoopIS, DOA, and
pothesis, indicating this solution as a good alternative to ODBASE, pages 492–508, 2004.
this issue while presenting a decrease on prediction accu- [12] J. Masthoff. Modeling the multiple people that are
racy of only 20% by cross-validation. me. User Modeling 2003, pages 146–146, 2003.
Another important achievement was caused by transpar- [13] J. Masthoff. Group modeling: Selecting a sequence of
ent recommendations. Results from the third insight turn television items to suit a group of viewers. User
prediction accuracy by cross-validation an even more ques- Modeling and User-Adapted Interaction, 14(1):37–85,
tionable benchmark method. Both neighborhood formation 2004.
methods presented a considerable improvement of 6.64% [14] A. Rashid, I. Albert, D. Cosley, S. Lam, S. McNee,
and 14.12%. While choosing online evaluation methods, one J. Konstan, and J. Riedl. Getting to know you:
could have better conclusions about the RS quality. learning new user preferences in recommender systems.
In Proceedings of the 7th international conference on
Intelligent user interfaces, pages 127–134. ACM, 2002.
6. ACKNOWLEDGMENTS [15] F. Ricci, L. Rokach, and B. Shapira. Introduction to
This research is part of a master studies sponsored by recommender systems handbook. Recommender
Monesia: MObility Network Europe-Southamerica: an Insti- Systems Handbook, pages 1–35, 2011.
tutional Approach, an Erasmus Mundus External Coopera- [16] B. Sarwar, G. Karypis, J. Konstan, and J. Reidl.
tion Window. Item-based collaborative filtering recommendation
Thanks Lucas Carvalho, researcher at Federal University algorithms. In Proceedings of the 10th international
of Sergipe - Brazil, for cooperating on the development of conference on World Wide Web, pages 285–295. ACM,
the Facebook application named MyPopCorn. 2001.
[17] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl.
7. REFERENCES Recommender systems for large-scale e-commerce:
Scalable neighborhood formation using clustering. In
[1] X. Amatriain and J. Basilico. Netflix Proceedings of the Fifth International Conference on
recommendations: Beyond the 5 stars, 2012. Computer and Information Technology, pages
[2] G. Groh and C. Ehmig. Recommendations in taste 158–167, 2002.
related domains: collaborative filtering vs. social [18] N. Tintarev and J. Masthoff. Designing and evaluating
filtering. In Proceedings of the 2007 international explanations for recommender systems. Recommender
ACM conference on Supporting group work, pages Systems Handbook, pages 479–510, 2011.
127–136. Citeseer, 2007. [19] P. Victor, M. Cock, and C. Cornelis. Trust and
[3] I. Guy, N. Zwerdling, D. Carmel, I. Ronen, E. Uziel, recommendations. Recommender Systems Handbook,
S. Yogev, and S. Ofek-Koifman. Personalized pages 645–675, 2011.
recommendation of social software items based on
social relations. In Proceedings of the third ACM
conference on Recommender systems, pages 53–60.
ACM, 2009.
38