=Paper=
{{Paper
|id=None
|storemode=property
|title=Exploiting Social Networks in Recommendation: a Multi-Domain Comparison
|pdfUrl=https://ceur-ws.org/Vol-986/paper_10.pdf
|volume=Vol-986
|dblpUrl=https://dblp.org/rec/conf/dir/BelloginCCD13
}}
==Exploiting Social Networks in Recommendation: a Multi-Domain Comparison==
https://ceur-ws.org/Vol-986/paper_10.pdf
Exploiting Social Networks in Recommendation:
a Multi-Domain Comparison
Alejandro Bellogína,b , Iván Cantadora , Pablo Castellsa , Fernando Díeza
a
Information Retrieval Group, Department of Computer Science, Universidad Autónoma de Madrid
b
Information Access, Centrum Wiskunde & Informatica
{alejandro.bellogin, ivan.cantador, pablo.castells, fernando.diez }@uam.esa ,
alejandro.bellogin@cwi.nlb
ABSTRACT similar items to the user’s closest cluster, by using the co-
Recommender Systems aim at automatically finding the most sine similarity measure. Other works focus on graph-based
useful products or services for a particular user, providing techniques for finding the most relevant items for a partic-
a personalised list of items according to different input and ular user, inspired by algorithms from quite different areas,
attributes of users and items. State-of-the-art recommender successfully bringing them to social recommendation [6].
systems are usually based on ratings and implicit feedback
given by users about the items. Recently, due to the large In this paper, we compare the performance of social filtering
number of social systems appearing in the so called Web methods with standard collaborative filtering (CF) baselines
2.0, where friendship relations between people are explicit, using four different datasets on three domains (bookmarks,
social contexts exploitation has started to receive significant music, and movies). With this goal in mind, in the next
interest. In particular, social recommenders have started to section we present the methods evaluated in this paper, then,
be investigated that exploit social links between users in a in Section 3 we discuss the datasets used. After that, in
community to suggest interesting items. In this paper we Section 4 we present the results obtained.
compare a series of experiments developed in recent years
with different datasets where standard collaborative and so- 2. SOCIAL FILTERING RECOMMENDERS
cial filtering techniques were analysed. We show that social
Inspired by the approach presented in Liu & Lee [8], we anal-
filtering techniques achieve very high performance in the
yse a pure social recommender that incorporates social in-
three domains discussed (bookmarks, music, and movies),
formation into the user-based CF model, named as friends-
although they may have lower coverage than traditional col-
based (FB). Standard user-based CF typically computes
laborative filtering algorithms.
predictions by performing a weighted sum over a set of sim-
Categories and Subject Descriptors ilar
Pusers (usually called neighbours) as follows [1]: s(u, i) =
H.3.3 [Information Search and Retrieval]: Information C v∈N (u) sim(u, v)r(v, i), where r(v, i) denotes the rating
Filtering given by user v to item i, and sim(u, v) is the similarity be-
tween the two users. In this context, FB makes use of the
General Terms same formula as the user-based CF technique, but replaces
Algorithms, Experimentation, Performance
the set of nearest neighbours (N (u)) with the active user’s
Keywords (explicit) friends.
Recommender systems, Social Networks, Evaluation
In [3] we propose a social popularity recommender (Soc-
1. INTRODUCTION Pop), where the algorithm suggests those items that are
With the advent of the Social Web, a variety of new rec- more popular among the set of the active user’s friends. A
ommendation approaches have been proposed in the litera- third social recommender is evaluated where explicit dis-
ture [1]. Most of these approaches are based on the exploita- tances between users in the social P graph are integrated in
tion of social tagging information and explicit friendship re- the prediction formula: s(u, i) = v∈X(u,L) K −d(u,v) r(v, i).
lations between users (social filtering recommenders) [5, 8]. This approach was originally proposed in [5] and named
Commonly, algorithms dealing with social context attempt as personal-social (PerSoc), where the authors use the
to exploit the social connections of an active user. For exam- Breadth-First Search algorithm in order to build a social
ple, Shepitsen et al. [10] employs a personalisation algorithm tree for each user (denoted as X(u, L)), where L is the max-
for recommendation in folksonomies that relies on hierar- imum number of levels taken into consideration in the al-
chical tag clusters, which are used to recommend the most gorithm, and K is an attenuation coefficient of the social
network that determines the extent of the effect of distance
d(u, v) (we use Dijkstra’s algorithm, K = 2 and L = 6).
Besides these pure social recommenders, hybrid social rec-
ommenders are useful not only for exploiting the social con-
text of a user, but for providing higher coverage in extreme
situations (such as the social or rating cold start, where
no social context or ratings are available for a particular
DIR 2013, April 26, 2013, Delft, The Netherlands. user). In this paper we analyse the performance of a com-
bination between the friends-based method described above
and the classic user-based CF method, where all the ac-
tive user’s friends along with the set of most similar near- Table 1: Obtained performance values for different
est neighbours are used to produce recommendations. We datasets (reported metric is P@10). Best value for
name this method user-and-friends-based (UFB). Alter- each dataset in bold.
Method Last.fm Delicious CAMRa-S CAMRa-C
natively, more complex hybrid recommenders can be defined
UB 0.009 0.008 0.072 0.052
based on random walks [6] and linear combinations of the
MF 0.025 0.003 0.038 0.026
predictions from several recommenders [4], but we leave the
comparison of these methods across several domains as fu- FB 0.043 0.023 0.057 0.050
ture work (some initial insights can be found in [3]). SocPop 0.021 0.011 0.001 0.001
PerSoc 0.085 0.054 0.344 0.342
UFB 0.014 0.008 0.077 0.053
3. A MULTI-DOMAIN PERSPECTIVE
We report results using four different datasets on three do- a good alternative to rating-based methods; here, we ex-
mains. The first one was gathered from the social music tend such conclusion to other domains like bookmarks and
website Last.fm. As described in [2], we built our dataset movies. Merging this strategy with CF (UFB), nonetheless,
aiming to obtain a representative set of users, covering all does not improve the results obtained by the approaches sep-
music genres, and forming a dense social network. This arately except in the movie domain, where the CF algorithm
dataset contains 1.9K users, 17.6K artists (17.0K of them shows better performance than in the other contexts.
tagged), 186.5K tag assignments (98.6 per user), and 25.4K
friend relations (13.4 per user). Additionally, when considering alternative evaluation met-
rics, we found in [2] that social filtering methods have lower
The second dataset was obtained from Delicious, a social coverage and novelty than traditional CF and content-based
bookmarking site for Web pages. Also described in [2], we recommenders; however, their diversity is higher, as mea-
built this dataset with the same goal in mind as the one sured using α-nDCG. These negative aspects could be im-
stated for Last.fm dataset: to cover a broad range of doc- proved by building hybrid recommenders, where the per-
ument’s topics, and obtain a dense social network. In this formance accuracy is slightly degraded at the expenses of
case, the dataset contains 1.9K users, 69.2K bookmarked better coverage and novelty [3, 2].
Web pages, 437.6K tag assignments, and 15.3K friend re-
lations. On average, each user profile has 56.1 bookmarks, 5. ACKNOWLEDGMENTS
234.4 tag assignments, and 8.2 friends.
This work is supported by the Spanish Government
(TIN2011-28538-C02-01) and the Government of Madrid
The third dataset used was provided in the social track of (S2009TIC-1542).
the CAMRa Challenge [9]. This dataset was gathered by
the Filmtipset community, and contains social links between
users, movie ratings, movie comments, and other attributes 6. REFERENCES
[1] Adomavicius, G., and Tuzhilin, A. Toward the next generation
of users and movies. However, in such dataset every test of recommender systems: A survey of the state-of-the-art and
user has a social network, which is not a realistic scenario, possible extensions. IEEE Trans. Knowl. Data Eng. 17, 6
since in many social media applications such as Delicious or (2005), 734–749.
Last.fm the social network coverage is only partial. Because [2] Bellogı́n, A., Cantador, I., and Castells, P. A comparative
study of heterogeneous item recommendations in social
of this, we create a fourth dataset where we incorporate a systems. Inf. Sci. 221 (2013), 142–169.
number of users with no friends in the new test set used in [3] Bellogı́n, A., Cantador, I., Dı́ez, F., Castells, P., and
our experiments, more specifically, such number corresponds Chavarriaga, E. An empirical comparison of social,
to the number of test users contained in the original test collaborative filtering, and hybrid recommenders. ACM TIST
4, 1 (2013), 14.
set (439 users). We denote the former dataset as CAMRa- [4] Bellogı́n, A., Castells, P., and Cantador, I. Self-adjusting
Social (CAMRa-S) and the latter as CAMRa-Collaborative hybrid recommenders based on social network analysis. In
(CAMRa-C). SIGIR (2011), W.-Y. Ma, J.-Y. Nie, R. A. Baeza-Yates, T.-S.
Chua, and W. B. Croft, Eds., ACM, pp. 1147–1148.
[5] Ben-Shimon, D., Tsikinovsky, A., Rokach, L., Meisels, A.,
Shani, G., and Naamani, L. Recommender system from personal
4. PERFORMANCE COMPARISON social networks. In AWIC (2007), K. Wegrzyn-Wolska and P. S.
Table 1 shows the performance results of the four social fil- Szczepaniak, Eds., vol. 43 of Advances in Soft Computing,
tering recommenders presented before on the four datasets Springer, pp. 47–55.
[6] Konstas, I., Stathopoulos, V., and Jose, J. M. On social
already described. We also use a standard user-based CF networks and collaborative recommendation. In SIGIR (2009),
method with 15 neighbours and Pearson’s similarity [1] (UB) J. Allan, J. A. Aslam, M. Sanderson, C. Zhai, and J. Zobel,
and a matrix factorisation approach in which the rating ma- Eds., ACM, pp. 195–202.
trix is factorised into 50 dimensions [7] (MF) as baselines. [7] Koren, Y., Bell, R. M., and Volinsky, C. Matrix factorization
techniques for recommender systems. IEEE Computer 42, 8
(2009), 30–37.
We observe that the best performing approach is the Per- [8] Liu, F., and Lee, H. J. Use of social network information to
Soc strategy, which adapts the well-known CF formula by enhance collaborative filtering performance. Expert Syst. Appl.
37, 7 (2010), 4772–4778.
weighting the similarity between the user’s and her neigh- [9] Said, A., Berkovsky, S., and Luca, E. W. D. Introduction to
bours’ rating-based profiles with the users’ distances in the special section on camra2010: Movie recommendation in
social graph. These results thus provide empiric evidence context. ACM TIST 4, 1 (2013), 13.
that combining CF and social networking information pro- [10] Shepitsen, A., Gemmell, J., Mobasher, B., and Burke, R. D.
Personalized recommendation in social tagging systems using
duces better recommendations than CF alone. Very inter- hierarchical clustering. In RecSys (2008), P. Pu, D. G. Bridge,
estingly, the FB strategy, which recommends items liked by B. Mobasher, and F. Ricci, Eds., ACM, pp. 259–266.
explicit friends, obtains acceptable precision values. As con-
cluded by Konstas and colleagues [6] for Last.fm, recommen-
dations generated from the users’ social networks represent