Mining for Social Serendipity
Alexandre Passant1, Ian Mulvany2, Peter Mika3, Nicolas Maisonneauve4,
Alexander Lser5, Ciro Cattuto6, Christian Bizer7, Christian Bauckhage8,
Harith Alani9
1 DERI, National University of Ireland, Galway - alexandre.passant@deri.org
2 Nature Publishing Group - ian@mulvany.net
3 Yahoo Research, Barcelona - pmika@yahoo-inc.com
4 Sony CSL, Paris - n.maisonneuve@gmail.com
5 TU Berlin - aloeser@cs.tu-berlin.de
6 ISI Foundation - ciro.cattuto@roma1.infn.it
7 FU Berlin - christian.bizer@fu-berlin.de
8 Deutsche Telekom Laboratories - christian.bauckhage@telekom.de
9 University of Southampton - ha@ecs.soton.ac.uk
Abstract. A common social problem at an event in which people do
not personally know all of the other participants is the natural tendency
for cliques to form and for discussions to mainly happen between people
who already know each other. This limits the possibility for people to
make interesting new acquaintances and acts as a retarding force in the
creation of new links in the social web. Encouraging users to socialize
with people they don't know by revealing to them hidden surprising links
could help to improve the diversity of interactions at an event. The goal
of this paper is to propose a method for detecting "surprising" relation-
ships between people attending an event. By "surprising" relationship
we mean those relationships that are not known a-priori, and that im-
ply shared information not directly related with the local context of the
event (location, interests, contacts) at which the meeting takes place.
To demonstrate and test our concept we used the Flickr community. We
focused on a community of users associated with a social event (a com-
puter science conference) and represented in Flickr by means of a photo
pool devoted to the event. We use Flickr metadata (tags) to mine for
user similarity not related to the context of the event, as represented
in the corresponding Flickr group. For example, we look for two group
members who have been in the same highly specic place (identied by
means of geo-tagged photos), but are not friends of each other and share
no other common interests or, social neighborhood.
Key words: serendipity, online activity, context, ubiquitous computing
1 Motivation
In the course of discussion at the workshop a theme which this group came
to was to question the purpose of the applications that we make. We were led
to ask whether we could devise an application, based on available information
Dagstuhl Seminar Proceedings 08391
Social Web Communities
http://drops.dagstuhl.de/opus/volltexte/2008/1791

2 Authors Suppressed Due to Excessive Length
from the social web, that could in some way increase the delight that a person
experienced in their life. In thinking about what causes delight we quickly focused
on the idea of surprise, and in particular serendipitous surprises. Wikipedia
denes serendipity as "the eect by which one accidentally discovers something
fortunate, especially while looking for something else entirely"10. Julius Comroe
described it as "looking in a haystack for a needle and discovering a farmers
daughter". A well known example is the discovery of Penicillin by Alexander
Fleming. He failed to disinfect cultures of bacteria when leaving for vacation, only
to nd them contaminated with Penicillium molds, which killed the bacteria. He
had previously done extensive research into antibacterial substances, and so this
discovery had a particular resonance for him. To be sure, surprise is not always
a good thing. An example is the invention of Gelignite by Alfred Nobel, when he
accidentally mixed collodium (gun cotton) with nitroglycerin. Though initially
a good accident the personal regret that he experienced through the misuse of
his invention led to the creation of the foundation that now bears his name.
In a social situation, for instance a topic-oriented meeting, intuitively it is not
surprising that there are common connections between people, and that people
who share those connections socialize together. Those connections can be purely
social graph connections (they know people in common), geo-historical connec-
tions (they have been to similar events in the past) or connections of interest
(they have published in the same sub-eld). Even if two attendees at a confer-
ence are unaware of their shared trips to similar conferences in the past, such a
connection is not really surprising as it may be inferred from the shared context
of the current meeting. Intuitively these connections are not really surprising.
It is more surprising if two people shared a hidden relationship which is clearly
not related to the topic of the shared context. Such a connection may be even
even more surprising if that relationship is not common to other members of the
shared group. If such relationships, or connections, in the extended social graph,
or the social hyper-graph, could be uncovered, or mined prior to, or even at an
event, they may form a sucient catalyst for people at the event to forge new
relationships or connections.
2 Approach
2.1 The Social Hyper-graph
As Web 2.0 sites and social networking sites expand the number of traces that
people leave in the web grows. Traces of social connections have been available
for a long time, the co-authorship network of academic literature is one much
studies example, however as threshold for what constitutes a signal decreases
the volume of information available increases correspondingly. The poster boy of
the Web 2.0 world in 2008 has been the microblogging service Twitter 11, that
lead to ubiquitous life streaming, especially as people can update their account
10 http://en.wikipedia.org/wiki/Serendipity - Accessed on 22nd Oct. 2008
11 http://www.twitter.com - Accessed on 22nd Oct. 2008

Mining for Social Serendipity 3
using their mobile phone, but no less important is the emerging web of things.
Application layers such as Yahoo! Research's Fire Eagle make possible the cre-
ation of social networks that are aware of the location of their members in real
time12. Sites such as Doppler 13 already use this kind of information to mediate
connections between people 14. 15. Of emerging interest will be the traces left
by devices embedded in the environment. Mobile phones in particular already
provide a platform for embedded sensors. The traces left by these devices are
leaving nascent connections between not just people, but between people, places
and things. The type of graph that best describes these relationships will proba-
bly be some kind of an n-partate or hyper graph, connecting on-line and o-line
worlds. There is considerable scope for investigating which of the relationships
that may emerge from this kind of data constitute important relationships, but
in this paper we try to outline an approach to uncovering surprising relationships
in any given network.
2.2 Surprise is an Attribute.
A measure of surprise can only be considered in relation to the item which is
described as surprising, e.g. an event, a TV show, a person. Surprise can modeled
through an event based approached using Baseian statistics. By comparing the
dierence between the prior and posterior probablitliy of an event happening a
measure for the surprisingnes of the event can be generated. Following directly
the description of surprise given in the Formal Bayesian Theory of Surprise16 we
can set a measure for the surprise S of a the event or observation of interest D
given the model space M of the observer. This is given as the relative entropy
of the Kulback-Leibler divergence of prior and posterior probability distribution
of the likelihood of the event happening. What we are doing is measuring how
far apart the prior and posterior distributions are. When they are far apart the
event is surprising and it's measure is given by:
S(D;M) = KL(P (M jD); P (M)) =
Z
M
P (M jD) log
P (M jD)
P (M)
dM:
In the context of social web communities the event in question will be the
creation of a link in some network. The key question in regard to this work is
guring out how to model the M that an observer, or user, will have of the
12 http://feblog.yahoo.net/2008/07/17/being-social-with-fire-eagle/ - Ac-
cessed on 11th Nov. 2008
13 http://www.slideshare.net/blackbeltjones/reboot90-travel-serendipity -
Accessed on 11th Nov. 2008
14 http://www.slideshare.net/blackbeltjones/reboot90-travel-serendipity -
Accessed on 11th Nov. 2008
15 http://www.zephoria.org/thoughts/archives/2008/09/21/i_will_be_joini.
html - Accessed on 11th Nov. 2008
16 http://ilab.usc.edu/surprise/ - Accessed on 14th Nov. 2008

4 Authors Suppressed Due to Excessive Length
explicit links that they have in their network. By uncovering a hidden link that
they may not be aware of, we change the structure of the network of explicit
links of the observer. Which of the many threads that connect us to each other
will bring about the greatest change in the way that the observer perceives the
links they already know about.
2.3 Serendipity is a Surprising Event.
As stated, our goal in this paper is to describe and propose a system which
increases the serendipidous happenings in a users life. We can expect that a
typical event at a conference is that people will meet each other. Can we provide
a system which uncovers the meetings that would be maximally surprising for
the people involved? The prior expectation for people at a meeting is that they
will have nothing in common outside of the shared topic of the meeting with
other members of the meeting. By suggesting connections with people who have
a strong, but unknown connection between each other we may broker surprising
meetings.
More formally, we dene a surprising relationship as a relationship between
two individuals (e.g. Alice and Bob) having the following properties:
{ Alice and Bob do not know each other from before, hence do not share a
common edge in their relationship network, whether it is online or in the
real-world;
{ The unexpected shared interest is not related to the context (time, location,
friends ...) they share, and neither is it common to other members of the
community;
Basically, our system nds unexpected answsers to the question "What topic
do I share with X?". The more the topic is unrelated to the context in which the
question is asked, the bigger the surprise is.
2.4 A Simple Model
We propose to create a representation of the model of a user, or a user's interests,
based on the tags that they leave in social web applications. Hence, we bridge
o-line and on-line world, by identifying relationships between people based on
their past footprints on the Web 1. Ideally one would wish to make a more rened
model through utilizing as much information as was available, and appropriate
for use, however for a rst pass it is instructive to try with the an easily available
data set. By taking a simple distance metric of the vectors created by a users
personomy we can nd users from a group who may be close or far from each
other in the context of a given interest. By looking for personomy information
that may be outside of the information common to a group, the tags of an event
from Flickr, for instance, we can look for connections that might be maximally
surprising thanks to what people tagged in the past. Key to this is that the
metric over which we search is not a simple additive space in which we try

Mining for Social Serendipity 5
to minimize the distance between all of the available sources of information,
but rather that we also try to nd at least some attributes over which we nd
a maximal distance, so that the eventual measure we come up with has the
potential to reveal surprising connections.
Fig. 1. Mining surprise thanks to online footprints from the past
3 Related Work
In this section we brie
y introduce to existing research from the area of recom-
mender systems, web mining and previous work on theories of surprise.
3.1 Measuring Serendipity in Recommender Systems
In [2], the authors propose metrics unexpectedness and unexpectedness r for
measuring the serendipity of recommendation lists produced by recommender
systems. In their work unexpectedness is a metric for a whole recommendation
list, while unexpectednesss r is the metric that takes into account the ranking
in the list.

6 Authors Suppressed Due to Excessive Length
3.2 Web Mining
The authors of [1] proposes a new method for answering relationship queries on
two entities e.g., the connections between dierent places or the commonalities
of people. Their method matches top Web pages retrieved for individual entities
and automatically identies the connecting terms. To eectively lter out the
large amount of noise in the Web pages without losing much useful information,
we do windowing around query keywords, compute term weights based on the
characteristics of the two Web page sets, and only use the top potential con-
necting terms to compute the similarity values of Web page pairs. Their work
is orthogonal, since they are trying to minimize the distance between the con-
cepts requested in the relationship query. However, we are trying to maximize
the distance in a subset of the data avaliable about the query with the aim of
maximizing the surprise the result. The authors of [3] extract the location and
event data from Flickr data by usage distribution of each tag. They analyze
mappings of location and time metadata associated with photos and their tags
and apply dierent distance metrics.
4 Metrics for Evaluating the Serendipity of
Recommendation Lists
4.1 Experiment
In order to evauluate or approach, we designed the following algorithm.
{ Extract the tag distribution of the event based on the photos in the group
(E) from Flickr;
{ For each user:
 Compute the dierence of the event distribution (E) and the individual
distributions (U1; U2::);
 Store this as non event related interest (UNE1; UNE2);
{ Extract the tag distribution of all members in the group (G);
{ For each user:
 Compute the dierence of (G) and the individual distributions;
 Store this as non group related interest (UNG1; UNG2);
{ For each pair of users:
 Compute the correlation of their non event related interests or at least
nd one strong shared interest that are not common in U
{ Another example is to have an augmented experience of a place by showing
to the user his 'connection', e.g a friend of him has visited this place recently.
to diversify the interactions

Mining for Social Serendipity 7
4.2 Results
We ran that algorithm on a dataset of more than 900 pictures and around 40
people from a Flickr group related to the ISWC conference series 17. We indeed
identied two related users by shared topics that were not related to the event
topics. Unfortunately, it turned out that those two people were colleagues from
the same institution, however our algorithm was quite basic and did not take
into account even simply available information, such as the two individuals in
question being connected in Flickr by friendship.
5 Extending the Surprise
Based on our denition of social serendipity and surprising relationship, we iden-
tied several factors that might extend the strength of the surprise. Indeed, some
relationships are stronger that others. For instance, identifying that two people
in an event are the only two to have visited the Eeil Tower might be relevant
from the conference context point of view, but not at all from a global overview,
as the Eeil Tower is quite a famous location. Far more surprising would be had
those two people been the only two at the conference to have visited the Eiel
Tower french restaurant in Burnley.
5.1 Tag Specicity
Obviously, some tags are most restrictive that others. While our current approach
takes only the context into account to identify related tags, we must also consider
a global approach to identify really interesting tags. To achieve this goal, we can
rely on dierent strategies:
{ Considering the current social networking website that we analysed, a less
time a tag has been used in the whole service, the most relevant it is. Re-
garding Flickr, we can add a strongest weight to the tags that have been
used only a few times;
{ We can also extend the approach to use background knowledge. Ranking by
inverse number of results in search engines might also re-enforce the value
of the tag specity;
{ Wikipedia might also be used to identify specic tags. Here, we can consider
the fact that a page exists for that tag or not, but also its volume and its
activity. A page edited by only a few people certainly identify a niche topic
while a page edited by 100s of people is obviously related to a well-known
topic. In case a Wikipedia page is found, but not in any language, it might
also be a factor to identify a niche topic, i.e. foreign people that are interested
in a partucular aspect of a country's culture which is not known abroad:
17 http://flickr.com/groups/iswc/ Accessed on 14th Nov. 2008

8 Authors Suppressed Due to Excessive Length
5.2 Semantic Web and Tags Relationships
Instead of considering tag identity, we may also deal with related tag, thanks to
similarity measures and semantic distance. Bridging the gap between the tags
and the meaning they represent, modeled thanks to Semantic Web principles
and knowledge bases like DBpedia18, or more generally, data from the Linking
Open Data project19, would provide a large base of interlinking concepts. Then,
we should compute some semantic distance measures to identify related tags,
e.g. a tag identifying a band and another one related to the name of its singer.
5.3 Interlinking Social Networks and Folksonomis
Our current approach relies on a single Social Media platform to nd unexpected
relationships between people. To go further, we should consider the dierent so-
cial websites a user is member of, and consequently its whole social activity. With
the growth of Web2.0, it is becoming increasingly common for users to maintain
a presence in more than one site. For example, one could be bookmarking pages
in del.icio.us, uploading images in Flickr, listening to music in Last.fm, blogging
in Technorati, etc. The nature of these pursuits naturally leads users to express
the relevant aspects of their interests, which are likely to be dierent across the
sites. If such multiple identities and distributed activities could be brought to-
gether independently of the Web 2.0 sites thanks to the previously introduced
models, far better understanding can be reached about what users are interested
in.
Many popular sites are racing to develop tools to allow their users to carry
their personal proles to other sites. Within days from each other in May 2008,
Google, MySpace, and Facebook announced new initiatives for increasing social
prole portability called Friend Connect20, Data Availability, and Connect21 re-
spectively. Yet, those projects remain eorts of private companies and might
not be able to be linked each other for economic reasons. On the other end,
eorts from the Semantic Web community such as FOAF22 or SIOC23 might be
considered to ease that process. They provide a common semantics and related
tools to interlink distributed social networks and related data that will lead to
distributed, open and standard-based social graph modeling. Various applica-
tions providing those data from existing services have been build, for instance
exporters for Flickr24 of Facebook25, as well as exporters for major blogging
platforms26.
18 http://dpedia.org Accessed on 14th Nov. 2008
19 http://linkeddata.org Accessed on 14th Nov. 2008
20 http://www.google.com/friendconnect/ Accessed on 14th Nov. 2008
21 http://developers.facebook.com/connect.php Accessed on 14th Nov. 2008
22 http://foaf-project.org Accessed on 14th Nov. 2008
23 http://sioc-project.org Accessed on 14th Nov. 2008
24 http://apassant.net/home/2007/12/flickrdf/ - Accessed on 14th Nov. 2008
25 http://www.dcs.shef.ac.uk/~mrowe/foafgenerator.html - Accessed on 14
th Nov.
2008
26 http://sioc-project.org/applications - Accessed on 14th Nov. 2008

Mining for Social Serendipity 9
Cross-folksonomy integration will become the focus of much research and
development in the near future. There is a strong push towards opening up social
networking to support portability of data across various sites. . Folksonomies can
be interlinked at multiple levels:
{ User Correlation: Since the same individual may hold accounts in multiple
social networking sites, two separate folksonomies may be joined at the user
level. Such a joining enables one to analyse and study the tagging practices
of individuals across dierent tagging platforms.
{ Tag Correlation: It is likely that many tags will appear in more than one folk-
sonomy. As a result, cross-folksonomy networks can be generated by joining
the common tags.
{ Resource Consolidation: Finally, the resources themselves may appear across
multiple folksonomies. By understanding how resources in dierent folk-
sonomies relate to each other, e.g. through shared tags or users, it would
be possible to provide a consolidated view of resources distributed over mul-
tiple folksonomies.
The SCOT project27aims to achieve this goal of tag portability by providing a
model - and related tools - for tagging actions using Semantic Web technologies.
5.4 From tags to interest
Tags serve various purposes, such as for resource organisation, promotions, shar-
ing with friends, with the public, etc. However, studies have shown that tags are
generally chosen to re
ect their user's interests. If we can distill users' Web
2.0 activities to produce an accurate image of their interests, then far better
identication of surprising links could be achieved.
Tag combinations Moving from a cloud of tags to a list of interests is not trivial.
Not all tags re
ect an interest. For example, when we examined some del.icio.us
users, we found one person who tagged a resource with \time", which is an over-
general tag, and hence not very helpful for learning about the user's interests.
However, when we looked at other tags for this user on comparable posts, it
became clear that the particular interest of the user was astronomy, in which
the concept of time plays a specic role.
Tag disambiguation Dynamic tag disambiguation is another challenging prob-
lem. Some methods have been suggested, which are based on clustering the whole
collection of tags or resources, but such techniques are more suitable for static
environments where the data do not change too often. In the world of tagging,
this is hardly the case. Hence we need less demanding methods to disambiguate
tags, to cope with the highly dynamic nature of folksonomies, even if those
methods could never be perfect
27 http://scot-project.org Accessed on 14th Nov. 2008

10 Authors Suppressed Due to Excessive Length
Tag cleaning Tags are free text, and users can tag resources with any terms they
wish to use. On the one hand, this total freedom simplies the process and thus
attracts users to contribute. It also avoids the problem of forcing users into using
terms they do not feel apply, as opposed to enforcing the use of a set of terms.
For these reasons, the lack of constraints seems essential. On the other hand, it
generates various vocabulary problems, where tags can be too personalised, made
of compound words, mix plural and singular terms, meaningless, synonymous,
etc. To make tags more comparable, it will be necessary to deploy a number
of tag cleaning processes to increase their compatibility, which in return should
increase the quality of interest identication.
5.5 Temporal Factor
Finally, we can extend our model by taking the temporal aspect into account.
While a tagging action is usually represented as a tripartite graph, most websites
store the timestamp of that association. By considering this additional dimen-
sion, we should be able to identify tags that user shared at a given time which
can help to identify people sharing a common interest at the same time, or, con-
sidering geolocation tag, people that have been in the same place at the same
time. This last example shows that our method can help to bridge the gap be-
tween real world and online activities. More specically, we identify real-world
relationships (based on geolocation), thanks to online activities, thus closing a
triangle combining real-world and online representation. Of course, this tempo-
ral factor could be combined with the previous semantic distance, identifying
facts that two people were at the same time in two small pubs in a suburb of
Tokyo two years ago. Considering simple non-geographic tags, that temporal as-
pect can also be used to identify early-adopters or experts in a given eld, using
services as google trends of web archives.
5.6 Geographical Factor
By looking for geotags we can get an idea of the spatial area that the person has
visited. By correlating an area by it's size and by it's 'fame', as perhaps derived
from Wikipedia, we can weight those geotags accordingly. Smaller less famous
locations being more likely to lead to a surprising relationship.
5.7 Applications
We can envisage many application but two immediate applications that present
themselves are in the realms of dating services and scientic collaboration. By
being able to provide an ice-breaking connection between two people we imagine
that such a system would enable people to make connections faster. By looking at
the co-authorship network such a system may be able to recommend interesting
cross discipline partnerships. By nding people who are close in one metric (a
technique say), but distant in another metric (a eld say) such a system could
aid researchers by providing paths out of the local minima that they work in.

Mining for Social Serendipity 11
6 Further Work
The implementation that we developed was clearly a toy implementation. It may
be of interest to take forward some of the ideas surrounding the surprisingness
of signals (astonishment value). A key issue is trying to develop more formally
the relationship between a personomy and the model that a user might have
about the world around them. When we begin to measure distances over a space
in which we try to maximize our result over some dimensions and minimize our
result over other dimensions questions about the concavity of the space may
arise, in addition to questions about nding an optimal search strategy.
7 Conclusion and Comments
By thinking in a somewhat unexpected direction about the information that
we can infer from the electronic traces and detritus that people leave in their
electronic wakes, we have found a novel and potentially interesting methodology
for uncovering surprising relationships, however it should be noted that not all
things that are surprising are by default a good thing. Discovering that someone
whom one had expected to be similar actually holds radically dierent views from
oneself can lead initially to surprise and then could lead to discomfort. We term
the dierence between good surprise and bad surprise the dierence between
the uncanny and the unexpected. A signicant challenge for any system that
implements maximal surprise is to distinguish between good surprise and bad
surprise.
References
1. Gang Luo, Chunqiang Tang, and Ying li Tian. Answering relationship queries on
the web. In Carey L. Williamson, Mary Ellen Zurko, Peter F. Patel-Schneider, and
Prashant J. Shenoy, editors, WWW, pages 561{570. ACM, 2007.
2. Tomoko Murakami, Koichiro Mori, and Ryohei Orihara. Metrics for evaluating
the serendipity of recommendation lists. In Ken Satoh, Akihiro Inokuchi, Katashi
Nagao, and Takahiro Kawamura, editors, JSAI, volume 4914 of Lecture Notes in
Computer Science, pages 40{46. Springer, 2007.
3. Tye Rattenbury, Nathaniel Good, and Mor Naaman. Towards automatic extraction
of event and place semantics from
ickr tags. In SIRIR '07: Proceedings of the
30th annual international ACM SIGIR conference on Research and development in
information retrieval, pages 103{110, New York, NY, USA, 2007. ACM Press.