=Paper=
{{Paper
|id=None
|storemode=property
|title=Exploiting Social Ties for Search and Recommendation in Online Social Networks - Challenges and Chances
|pdfUrl=https://ceur-ws.org/Vol-581/gvd2010_3_1.pdf
|volume=Vol-581
|dblpUrl=https://dblp.org/rec/conf/gvd/Bischoff10
}}
==Exploiting Social Ties for Search and Recommendation in Online Social Networks - Challenges and Chances==
https://ceur-ws.org/Vol-581/gvd2010_3_1.pdf
Exploiting Social Ties for Search and Recommendation in
Online Social Networks – Challenges and Chances
Kerstin Bischoff
L3S Research Center / Universität Hannover
Appelstrasse 4
30167 Hannover, Germany
bischoff@L3S.de
ABSTRACT They form new (indirect) connections by reading and adopt-
Online social networking is a huge trend. On Facebook or ing from other peoples’ blogs or tweets. Similarly, in col-
MySpace people connect with their friends or make new laborative tagging systems like Last.fm, Delicious or Flickr
friends. They form new (indirect) connections by reading people share bookmarks and tagged resources with friends
and adopting from other peoples’ blogs or tweets. Simi- or unknown, similar users. Reasons are manifold: staying in
larly, in tagging systems like Delicious or Flickr people share touch, socializing, finding answers/experts, share resources
tagged resources with friends or unknown, similar users. Of- and knowledge, etc. Offline social networks have long been
fline social networks have long been studied in sociology, studied under the perspective of sociology, epidemiology,
epidemiology, etc. However, the new online networks offer and even thermodynamics. However, online networks offer
new ways to revisit old theories as well as to find emerging new ways to revisit old theories as well as to find emerging
trends with respect to information diffusion and sharing in trends in information diffusion and sharing in the Web. For
the Web 2.0. The goal is to explore and exploit relational search and recommendation, the topic of tie strength is in-
ties in a way to enable the mining of useful knowledge and teresting. In ‘real’ social networks, strong (i.e. family, close
effective information propagation/diffusion. Assuming no or friends) and weak ties (loose acquaintances) have been found
only potential ties explicitly given, the focus is first on the to show different characteristics e.g. with respect to services
analysis of collaborative tagging and its potentials for user offered (see section 3). As McAfee [19] pointed out, different
profiling, recommendations and search. Some first related kinds of ties, if supported by the right technology, may of-
studies, approaches and ideas for future work address the fer different potential benefits for information exchange and
identification and exploitation of weak and strong ties in collaboration (adapted from [19]):
online networks.
• Strong ties: Collaboration in a closed group, e.g. co-
workers (BSCW, CVS, Wikis)
Categories and Subject Descriptors
H.3.1 [Information Storage and Retrieval]: Content • Weak ties: Innovation, Non-redundant information,
Analysis and Indexing; H.3.4 [Information Storage and Network bridging (Email, Social networking systems)
Retrieval]: Systems and Software; H.3.5 [Information
Storage and Retrieval]: On-line Information Services • Potential ties: Efficient search, Tie formation (Blogo-
sphere, Bulletin boards, Folksonomies)
General Terms • No/Absent ties: Collective Intelligence, statistical pat-
Experimentation, Human Factors, Algorithms terns (Folksonomies, Prediction Market, Question An-
swering)
Keywords The goal is to explore and exploit relational ties in a way to
social network analysis, social ties, collaborative tagging, enable the mining of useful knowledge and effective informa-
search, recommendation tion propagation/diffusion so that people are provided the
information they need. Focusing on absent, potential and
1. INTRODUCTION weak ties, we will present first research results, algorithms
With the advent of the Web 2.0 online social networking has and ideas for future work. To better understand ‘collec-
become a huge trend. On platforms like Facebook or MyS- tive intelligence’ expressed via tags, i.e. statistical patterns
pace people connect with their friends or make new friends. found in folksonomies, we start by describing an analysis
of different tagging systems and summarize some experi-
ments on how to exploit social annotations to enrich meta-
data for multimedia resources. Web 2.0 tools and environ-
ments like the personalized Internet radio and social music
network Last.fm 1 have made collaborative tagging so pop-
ular: any user can assign freely selected words, in the form
of keywords or category labels, to shared content – thereby
describing and organizing these resources. As a result, a
Copyright is held by the author/owner(s). 1
GvD Workshop’10, 25.-28.05.2010, Bad Helmstedt, Germany. http://last.fm
huge amount of manually created metadata describing all data, such as music, pictures or videos, the gain provided
kinds of resources as well as user interests is now available. by the newly available textual information is substantial,
Such semantically rich user generated annotations are espe- since with most prominent search engines on the Web, users
cially valuable for recommending, searching and browsing are currently still constrained to search for multimedia using
multimedia resources, such as music, where these metadata textual queries. A large amount of tags is also accurate and
enable retrieval on the newly available textual descriptions reliable; in the music domain for example 73.01% of the tags
represented by tags. However users’ motivations for tagging also occur in online music reviews written by experts.
resources, as well as the types of tags differ across systems.
These tags represent quite a few different aspects of the re- Regarding search, our studies showed that most of the tags
sources they describe [2] and more research is needed on how can be used for search and that in most cases tagging be-
these tags or subsets of them can be used effectively for user havior exhibits approximately the same characteristics as
profiling and search in social networks. Weak and strong ties searching behavior. However, some noteworthy differences
and the corresponding topics of information diffusion and so- have also been observed. Namely, for the music domain, the
cial search are considered next. Weak ties are often ‘bridges’ usage context (i.e. situation suitable for listening to a par-
connecting different communities, thus bringing new infor- ticular song – e.g. “pool party”) is very useful for search, yet
mation (e.g. job seeking). Strong ties offer mutual support, underrepresented in the tagging material. Similar, for pic-
trust, but likely share knowledge, preferences, values and tures and music opinions or qualities (e.g. characteristics,
friends. The section will cover the related issues of iden- moods) queries occur quite often, although people tend to
tifying tie strength or characterizing friendship relations in neglect this category for tagging. Clearly, supporting and
online networks. It introduces recent ideas to analyze and motivating tags within these categories could provide addi-
exploit such networks to improve common approaches to tional information valuable for search.
search and recommendation.
2.2 Knowledge mining from tags
2. TAGS FOR USER PROFILING, SEARCH Our analysis [2] showed some clear gap between the tagging
and the querying vocabulary for music as well as pictures.
AND RECOMMENDATIONS For pictures, a large portion of tags refer to location infor-
As they offer a promising way to estimate similarity between
mation. However, queries targeting images much more of-
resources, users and resources or between different users,
ten name subjective aspects, e.g. “scary”, “rage” or “funny”.
the usefulness and reliability of tags is important for many
For music, tags predominantly name the genre (i.e. type),
search and recommendation algorithms. In the first section
though when searching for music, the majority of queries
the focus will be on analyzing tag usage patterns and their
falls into these categories: 30% of the queries are theme-
implications for user profiling, search and recommendation.
related, 15% target mood information. In this section we
Then, we will present approaches exploiting tags to enrich
will shortly present an approach for detecting moods and
resources or user profiles with additional information – music
themes for songs [2, 3, 5] and emotions in photos [4]. It relies
moods and themes as well as picture moods.
on collaborative tagging and aims at bridging exactly this
gap identified. The methods proposed can be used in various
2.1 Usefulness of tags for profiling and search ways: as part of an application where the recommendations
Here we present results of an in-depth study [2] of tagging for are presented to the user for selection, for indexing and thus
different kinds of resources and systems – Web pages (Deli- enriching the metadata indexes to improve searchability, or
cious), music (Last.fm) and images (Flickr ). Tags serve var- to automatically create mood-based picture catalogs.
ious functions based on system features like resource type,
tagging rights, etc. [17], and not all these tags are equally 2.2.1 Datasets
useful for user profiling or for interpersonal retrieval. For For our experiments we gathered data from several sources
being able to improve tag based user profiling and search, (for details please refer to [3, 4]):
we first need to know how tags are used and which types of
annotations we can expect to find along with resources. [17] • From the Allmusic.com website we collected the labels
identifies organizational motivations for tagging, other more of 178 different moods and 73 themes together with
social motivations include opinion expression, attraction of music tracks that fall into these categories.
attention, and self-presentation.
• For 13,948 songs obtained from Allmusic.com, we could
Our analysis revealed the necessity and usefulness of a com- get user tags from the social music platform Last.fm.
mon tag classification scheme for different collections, which • To investigate whether lyrics can provide added value
allows the comparison of the types of tags used in different in the task of mood and theme recommendation, we
tagging environments. For example, the distributions of tag also obtained the lyrics for our tracks from lyricsdown-
types strongly depend on the resources they annotate: for load.com and lyricsmode.com.
Flickr and Delicious the most frequently used tags (50% of
the cases) refer to topic concepts (i.e. what the resource is • For the purpose of deriving mood labels for pictures,
about), while for Last.fm, type-related tags (e.g. genres) are we manually selected Flickr groups that correspond
the most prominent ones. Other interesting results of this to the emotion/mood labels in the hierarchy of hu-
analysis refer to the added value of tags to existing content: man emotions presented in [25]. The taxonomy com-
More than 50% of existing tags bring new information to the prises the six primary emotions “Love”, “Joy”, “Sur-
resources they annotate. For the music domain this is even prise”, “Anger”, “Sadness”, and “Fear”, each of which
the case for 98.5% of the tags. Especially for multimedia has more fine-grained secondary emotions. We found
corresponding Flickr groups for 17 out of the 25 sec- Themes like “Late Night” or “Summertime” strongly depend
ondary emotions, including the six primary emotion on each person’s individual associations.
labels. For all pictures in the identified groups we
downloaded all associated user tags from Flickr. Pictures. A first set of experiments aimed at recom-
mending mood labels corresponding to the primary human
2.2.2 Picture Mood Recommendation Algorithm emotions. In this case, the classes to be learned by the classi-
fiers consisted of the union of all data belonging to all under-
The assumption on which we base our recommendation is
lying secondary emotions (e.g. the Love class comprises all
that the existing tags attached to music songs or photos
data gathered from the Flickr groups for Affection, Lust and
can possibly provide information regarding the correspond-
Longing). For the experimental runs on secondary emotion
ing mood (or theme). The user given tags are used as in-
label recommendations, each secondary emotion class rep-
put features for training a classifier over all mood or theme
resented a class to be learned. We perform a 10-fold cross
classes. Here we also make use of the Weka2 implementation
validation and evaluate the performance of our method ac-
for the Naı̈ve Bayes Multinomial classifier, which produces
cording to standard IR metrics. All recommendations corre-
for all pictures in the test set probability distributions over
sponding to the primary human emotions achieved very high
all classes of moods. For music, training and evaluation is
quality, with a value close to 1 for H@3 and even a H@1
done using the Allmusic.com ground truth, created manu-
between 0.61 and 0.91 - they clearly outperform a baseline
ally by music experts. All distinct tags span the feature
random classifier. We also compute the overall performance
vector of a song, weighted by the frequency with which the
over all primary emotion classes, as averages weighted by
tag occurred for a song. Similarly, all pictures pertaining
the number of instances corresponding to each class. The
to a specific mood class (i.e. social group on Flickr ) rep-
results are very good, with a value of 0.97 for H@3 and 0.93
resent the positive training examples, while pictures taken
for M RR. For primary emotions, correlation between class
randomly from the rest of the classes build up the set of
size and performance is medium: Pearson’s r is 0.45 for H@3
negative examples. The number of positive and negative ex-
and 0.63 for H@1, RP , and M RR. Thus, when misclassify-
amples for a class was equally balanced. In order to ensure a
ing instances the classifier is biased to incorrectly assigning
fair classification of the data, here all tags related to a mood
one of the two dominant classes “Fear” or “Sadness”. The
or emotion were deleted using WordNet synset information.
overall weighted results for the secondary human emotion
label recommendations are almost identical with the case of
2.2.3 Experiments and Results primary emotions. Again, we find a classifier bias towards
Music. Different experimental runs were performed using popular classes, correlation between a priori probability of
either tags, lyrics, or both plus varying classes to be pre- a class and performance is smaller for secondary emotions
dicted. Since the 178 mood terms from Allmusic.com are (Pearson’s r is between 0.32 and 0.37 for the different eval-
hardly distinguishable for a non-expert, these labels were uation measures). Looking at the different mood classes
manually mapped to the hierarchy of basic human primary individually, six out of 17 achieve 0.88 or higher for H@1
and secondary emotions (see 2.2.1). This resulted in 22 sec- and 0.93 or more for H@3. For four moods, “Affection”,
ondary mood terms and six primary classes. For themes, “Zest”, “Irritation” and “Lust”, performance is considerably
the eleven classes were reduced to nine by applying a Word- lower with H@3 ranging from 0.18 to 0.52. The main reason
Net based clustering accounting for word overlap in synsets. is the relatively small number of pictures contained in each
The best performing methods are those using tags as input of these groups, which made learning more difficult.
features, while classification based on lyrics performs worst.
Combining tags and lyrics achieves good results, sometimes Compared to the music mood and theme recommendations,
even slightly outperforms tag-based method. While incorpo- inference of moods/emotions for picture resources is of higher
rating lyrics features led to good results for genre [3], they do quality. This is probably due to the data which was used as
not seem to be indicative of the mood of a song. For themes, ground truth: mood-related Flickr groups, manually created
there is a slight, yet rarely significant, effect. For the case by users. The ground truth gathered from AllMusic.com
of theme recommendations, given the original eleven themes had to be mapped to the hierarchy of human emotions to
we achieve a H@3 of 0.80 based on tags and lyrics. The best reduce the extremely high number of classes. This process
results, H@3 of 0.88, are achieved for the algorithm using potentially introduces some noise into the data. The results
a combination of tags and lyrics as features and applying a confirm once more the hypothesis on which we based our
WordNet synonymy based clustering on the theme classes. recommendation approach: existing tags can give good indi-
Compared to themes, mood recommendations do not per- cations regarding the corresponding moods of the pictures/
form as well when using many classes, achieving only a H@3 moods and themes of music songs.
of 0.64 for the manually clustered 22 secondary emotions.
Reducing the number of clusters to the 6 first level classes 3. SEARCH AND RECOMMENDATIONS
of basic human emotions boosts the performance consider-
ably and for the best method using tags and lyrics as input
IN SOCIAL NETWORKS
features we achieve a H@3 value of 0.89. Micro-evaluating From sociology as well as social network analysis as a dis-
results per specific class, shows that some classes are rela- cipline in its own right quite some descriptive statistics and
tively easy to recommend. Others may require special atten- generative models have become popular to characterize so-
tion or some level of disambiguation. In general, those class cial networks. Well-known is that weak ties often act as
labels that are harder to recommend appear more ambiguous bridges and thus hold potential for new information and the
with the corresponding annotations being mostly subjective. generation of creative ideas, job offers (e.g. [6, 10]). On the
other hand, strong ties offer support and trust (e.g. see [15])
2
http://www.cs.waikato.ac.nz/∼ml/weka/ and show tendencies for homophily and transitivity [20, 9].
This means they likely share knowledge, preferences, values 3.2 Approaches exploiting ties
and friends. While many ‘real-world’ studies had to cope Approaches for efficiently searching and propagating infor-
with design issues (amount of data, retrospective informant mation in online communities build strongly upon methods
accuracy, etc.), online networks offer huge and interesting developed in social networks analysis. Real networks like the
datasets to work with. With respect to exploiting online so- WWW, the Internet, spreading of diseases/epidemics, so-
cial connections for search and recommendation, two broad cial, biological and linguistic networks have so far been stud-
areas need to be studied: ied mostly with respect to structure. Graph theoretic mea-
sures like density, indegrees, outdegrees, centrality, diame-
• How can ties be modeled based on implicit and explicit ter, (structural) cohesion, etc. indicate the potentials and
indicators found in online social networks? How can bottlenecks of a network. As an example, one experimental
these ties be characterized? Do we find homophily and finding recently receiving a lot of attention is the ‘small-
heterophily as expected from offline social relations? world phenomenon’ (also ‘6 degrees of separation’)[14, 10].
How do these ties evolve over time? While early works investigated patterns of communication in
small, closed groups, recent work analyzed communication
• How useful are weak and strong ties in search and flows in huge social networks e.g. based on mobile phone calls
recommendation (information propagation)? How can [21], instant messaging [14], or the cascading propagation of
they be incorporated to show their potential benefit in information through the blogosphere [11]. In their studies,
information systems (e.g. diversity/non-redundancy)? [21] found that weak ties are crucial for the structural in-
Are there restrictions by contexts or domains? tegrity of the network. Strong ties, on the other hand, are
important for sustaining local groups/cliques. Concerning
We address related work, open issues for both topics in the information propagation both types of connections are not
following sections to then conclude with planed future work. sufficient, the first due to infrequent, rare contact, the lat-
ter due to being bound to their local groups. Epidemic or
3.1 Inferring and characterizing tie strength gossip-based algorithms adapt such established patterns to
In sociological theory, an impressive amount of work has enable efficient spread of information for distributed com-
been done regarding the measurement of tie strength. A lot puting or in Peer-2-Peer systems [7].
of reliable indicators have been identified, e.g. interaction
frequency, duration, intimacy (e.g. [18]), etc. ([8] for a quick Similarly, social search and recommendation algorithms try
overview). Kahanda and Neville [12] recently presented a to exploit the communication and interaction patterns found
machine learning approach to automatically identify strong in social networks as well as, for example, the trust and
friends. The authors formulated a link strength prediction similarity typical for strong ties. Referral Web is a first
task: For each friend pair (u,v) given their user profile at- approach to integrate social networks and Collaborative Fil-
tributes like age, gender, etc., their interactions (writing on tering (CF) [13]. Here, a social network was constructed
the friend’s wall, tagging a photo), and network informa- from cooccurring names in the WWW, for example, links
tion (e.g. number of mutual friends) a supervised learning on a home page or co-authorship. Queries that can be an-
method decides whether they are “top friends”. Evaluation swered based on this network have the form “which con-
on data from the public Purdue Facebook network, where nection do I have to XY” or “documents about databases by
users can nominate best friends within the “Top friends” ap- people close to XY”. [26] models real-world information flows
plication, showed that with an AUC of 87% best friends in order to give recommendations and rank users according
can be successfully distinguised from weak ties. Those best to influence based on the usage of certain communications
classification results were achieved on network-transactional paths. For this, diffusion rate between users is computed
features (i.e., moderate transactional activity like wall posts based on access time/order to the same documents. The
by interactions with other users).Thus, user interactions are automatic evaluation shows that standard CF algorithms
highly predictive, but it is also necessary to consider such can be outperformed in accuracy by up to 80%. Moreover,
transactional events in the context of user behavior within the underlying social network can be used to overcome the
the larger social network. Surprisingly, attribute similarity sparsity/missing data problem, for example, by applying fac-
features lead to classification results close to a random clas- tor analysis on the user-item-matrix enriched with explicit
sifier, indicating that the homophily assumption does not user connections [16]. For personalized recommendations of
hold for this Facebook network or the important attributes new posts concerning some news item, [24] extend their CF
are not available in Facebook. recommender system in a way that strong social network
ties (here: members of a thematic group) indicate a high
In a similar work, Gilbert and Karahalios [8] predict tie value of a post with respect to completeness and simplicity.
strength as a linear combination of 74 Facebook variables Weak ties, in contrast, imply diversity of opinions. From
(e.g. last comment, num friends, wall words). Besides com- ratings given to posts the system learns a user’s preference
parable results, a mapping of (sociological) dimensions to regarding completeness and diversity, to which recommen-
the different variables is provided enabling generalization of dations are adapted. [23] presents a framework for social
the approach. In [27], Yamamoto & Matsumura analyzed search and recommendation that integrates classical CF at-
optimal heterophily between senders and receivers in terms tributes for users and resources with an ontology and social
of blogging influence (tracked via re-occurring terms and connectivity (explicit friendship or ‘spiritual’, i.e. similar in-
links) and domain knowledge. They found that the ma- terests modeled via tags) within a scoring model. A small
jority of pairs favor small heterophily, in particular people evaluation study shows that ‘spiritual’ connections in par-
most often adopt topics or products when the sender is just ticular improve search results significantly, but not for all
slightly more influential. kinds of queries. Social query expansion by tags used by
friends, however, did not lead to improved performance. In Analyses and applications for enhancing multimedia ir.
a related work, [1] demonstrated that social search, imple- Journal of Web Semantics, Special Issue on “Bridging the
mented as search among all friends having used a query term Gap” - Data Mining and Social Network Analysis for
Integrating Semantic Web and Web 2.0, 2010.
as tag before, possibly combined with an authority score for
[5] K. Bischoff, C. S. Firan, R. Paiu, W. Nejdl, C. Laurier, and
users can yield the best precision for search in Flickr. Also M. Sordo. Music mood and theme classification - a hybrid
for efficiently searching inside collaborative tagging networks approach. In Proceedings ISMIR, 2009.
like Delicious incorporating social connections between users [6] R. S. Burt. Structural holes and good ideas. American
and between tags proved useful. A top-k algorithm com- Journal of Sociology, 110:349–399, 2004.
bined with dynamic tag expansion and dynamically extend- [7] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson,
ing search over socially connected users can answer queries S. Shenker, H. Sturgis, D. Swinehart, and D. Terry.
considerably faster than traditional approaches [22]. Epidemic algorithms for replicated database maintenance.
In Proceedings PODC ’87, 1987.
[8] E. Gilbert and K. Karahalios. Predicting tie strength with
4. CONCLUSIONS & FUTURE WORK social media. In Proceedings CHI, 2009.
Online social networks offer great data to analyze and exper- [9] M. Granovetter. The strength of weak ties. The American
iment with for enhancing user profiling, search and recom- Journal of Sociology, 78:1360–1380, 1973.
mendation. The concept of tie strength seems a promising [10] M. Granovetter. The strength of weak ties: a network
framework for identifying the diverse potentials different on- theory revisited. Sociological Theory, 1:201–233, 1983.
line social relations can bring. First, collaborative tagging [11] D. Gruhl, R. Guha, D. Liben-Nowell, and A. Tomkins.
Information diffusion through blogspace. In Proceedings
provides reliable, non-redundant and interpersonally valu- WWW, 2004.
able metadata, that can be used to enhance searchability [12] I. Kahanda and J. Neville. Using transactional information
of resources as well as estimate user-user or user-item sim- to predict link strength in online social networks. In
ilarities. For this, no explicit friendship relationships have Proceedings ICWSM, 2009.
to be given. The results of our comparative tagging anal- [13] H. Kautz, B. Selman, and M. Shah. Referral web:
ysis provide more insight into the use of different kinds of combining social networks and collaborative filtering.
tags for improving search. With respect to weak and strong Communications of the ACM, 40(3):63–65, 1997.
ties, more research is needed on how to model ties based [14] J. Leskovec and E. Horvitz. Planetary-scale views on a large
instant-messaging network. In Proceedings WWW, 2008.
on explicit or implicit (e.g. tags) indicators. We plan to
[15] D. Z. Levin and R. Cross. The strength of weak ties you can
conduct experiments on learning tie strength and exploit- trust: The mediating role of trust in effective knowledge
ing it for search and recommendation within other kinds of transfer. Management Science, 50(11):1477–1490, 2004.
social networks, e.g. the music platform Last.fm. For dif- [16] H. Ma, H. Yang, M. R. Lyu, and I. King. Sorec: social
ferent domains, system designs and available transactional recommendation using probabilistic matrix factorization. In
data, results may deviate from the previous studies. So far, Proceedings CIKM, 2008.
there are still no unambiguous results regarding homophily [17] C. Marlow, M. Naaman, D. Boyd, and M. Davis. Ht06,
in recent online networks. Characterizing the relationships tagging paper, taxonomy, flickr, academic article, to read.
In Proceedings HYPERTEXT, 2006.
people form online and studying how these relations (or their
[18] P. V. Marsden and K. E. Campbell. Measuring tie strength.
attributes like taste profile similarity) evolve over time are Social Forces, 63(2):482–501, 1984.
important to assess the value of ties for improving search and [19] A. McAfee. How to hit the enterprise 2.0 bullseye. online:
recommendations. Applying standard social network analy- http://andrewmcafee.org/2007/11/
sis procedures on the new large datasets will also shed ad- how to hit the enterprise 20 bullseye/, 2007.
ditional light on larger community structures around strong [20] M. McPherson, L. Smith-Lovin, and J. M. Cook. Birds of a
and weak ties in general. More importantly, few work has feather: Homophily in social networks. Annual Review of
been done so far on how to incorporate tie strength into in- Sociology, 27:415–444, 2001.
formation retrieval and recommendation. First experiments [21] J.-P. Onnela, J. Saramäki, J. Hyvönen, G. Szabó, D. Lazer,
K. Kaski, J. Kertész, and A.-L. Barabási. Structure and tie
will use tie strength within the similarity computation of strengths in mobile communication networks. Proc. Natl.
users (to users and items). Information diffusion within per- Acad. Sci., 104, 2008.
sonal networks will be studied and a model derived on which [22] R. Schenkel, T. Crecelius, M. Kacimi, S. Michel,
social recommendations can be build. T. Neumann, J. X. Parreira, and G. Weikum. Efficient
top-k querying over social-tagging networks. In Proceedings
SIGIR, 2008.
5. ACKNOWLEDGMENTS [23] R. Schenkel, T. Crecelius, M. Kacimi, T. Neumann, J. X.
This work was partially supported by the GLOCAL project Parreira, M. Spaniol, and G. Weikum. Social wisdom for
funded by the European Commission (Contract No. 248984). search and recommendation. IEEE Data Eng. Bull.,
31(2):40–49, 2008.
6. REFERENCES [24] A. Seth and J. Zhang. A social network based approach to
[1] M. Bender, T. Crecelius, M. Kacimi, S. Michel, personalized recommendation of participatory media
T. Neumann, J. X. Parreira, R. Schenkel, and G. Weikum. content. In Proceedings ICWSM, 2008.
Exploiting social relations for query expansion and result [25] P. Shaver, J. Schwartz, D. Kirson, and C. O’Connor.
ranking. In Proceedings ICDEW 2008, 2008. Emotion knowledge: Further exploration of a prototype
[2] K. Bischoff, C. S. Firan, W. Nejdl, and R. Paiu. Can all approach. Journal of Personality and Social Psychology,
tags be used for search? In Proceedings CIKM, 2008. 52(6):1061–1086, 1987.
[3] K. Bischoff, C. S. Firan, W. Nejdl, and R. Paiu. How do [26] X. Song, Y. Chi, K. Hino, and B. L. Tseng. Information
you feel about “dancing queen”?: deriving mood & theme flow modeling based on diffusion rate for prediction and
annotations from user tags. In Proceedings JCDL, 2009. ranking. In Proceedings WWW, 2007.
[4] K. Bischoff, C. S. Firan, W. Nejdl, and R. Paiu. Bridging [27] H. Yamamoto and N. Matsumura. Optimal heterophily for
the gap between tagging and querying vocabularies: word-of-mouth diffusion. In Proceedings ICWSM, 2009.