A topology based approach twittersdlfkjsdlkfj

A topology-based approach for followees recommendation in Twitter

Marcelo G. Armentano, Daniela L. Godoy and Analía A. Amandi
Universidad Nacional del Centro de la Provincia de Buenos Aires
CONICET, Consejo Nacional de Investigaciones Científicas y Técnicas
Argentina
{marmenta,dgodoy,amandi}@exa.unicen.edu.ar

Abstract which can be regarded as friends or acquaintances, while
most of them behave either as information sources or infor-
Nowadays, more and more users keep up with mation seekers. Users behaving as information sources tend
news through information streams coming from
to collect a large amount of followers as they are actually
real-time micro-blogging activity offered by ser-
posting useful information or news. In turn, information
vices such as Twitter. In these sites, information is seekers follow several users to obtain the information they
shared via a followers/followees social network are looking for and rarely post any tweet themselves.
structure in which a follower will receive all the
Finding high quality sources among the expanding micro-
micro-blogs from the users he follows, named fol- blogging community using Twitter becomes essential for
lowees. Recent research efforts on understanding
information seekers in order to cope with information over-
micro-blogging as a novel form of communication
load. In this paper we present a topology-based followee
and news spreading medium have identified differ- recommendation algorithm aiming at identifying potentially
ent categories of users in Twitter: information
interesting users to follow in the Twitter network. This algo-
sources, information seekers and friends. Users act-
rithm explores the graph of connections starting at the target
ing as information sources are characterized for user (the user to whom we wish to recommend previously
having a larger number of followers than follo- unknown followees), selects a set of candidate users to rec-
wees, information seekers subscribe to this kind of
ommend and ranks them according to a scoring function that
users but rarely post tweets and, finally, friends are favors those users exhibiting the distinctive behavior of
users exhibiting reciprocal relationships. With in-
information sources.
formation seekers being an important portion of
Unlike other works that focus on ranking users according
registered users in the system, finding relevant and to their influence in the entire network [Weng et al., 2010;
reliable sources becomes essential. To address this
Yamaguchi et al., 2010], the algorithm we propose explores
problem, we propose a followee recommender sys-
the follower/following relationships of the user up to a cer-
tem based on an algorithm that explores the topol- tain level, so that more personalized factors are considered
ogy of followers/followees network of Twitter con- in the selection of candidates for recommendation, such as
sidering different factors that allow us to identify
the number of friends in common with the target user. Since
users as good information sources. Experimental only the topology of the social structure is used but not the
evaluation conducted with a group of users is re-
content of tweets, this algorithm also differs from works
ported, demonstrating the potential of the ap-
exploiting user-generated content in Twitter to filter infor-
proach. mation streams [Chen et al., 2010; Phelan et al., 2009; Es-
parza et al., 2010] or to extract topic-based preferences for
1 Introduction recommendation [Hannon et al., 2010].
Micro-blogging activity taking place in sites such as Twitter The rest of this paper is organized as follows. Section 2
is becoming every day more important as real-time informa- reviews related research in the area. Section 3 describes our
tion source and news spreading medium. In the follow- approach to the problem of followee recommendation in
ers/followees social structure defined in Twitter a follower Twitter. In Section 4 we present the experiments we per-
will receive all the micro-blogs from the users he follows, formed to validate our proposal and in Section 5 we present
known as followees, even though they do not necessarily and discuss the results obtained and in Section 6 we com-
follow him back. In turn, re-tweeting allows users to spread pared our results with a related approach. Finally, in Section
information beyond the followers of the user that post the 7, we discuss some aspects of our proposal and present our
tweet in the first place conclusions.
Studies conducted to understand Twitter usage [Java et
al., 2007; Krishnamurthy et al., 2008] revealed that few
users maintain reciprocal relationships with other users,

2 Related Work influence in Twitter using only its topology (see for example
[Kwak et al., 2010]) gives the worst results for followee
The problem of helping users to find and to connect with recommendation since people that are popular in Twitter
people on-line to take advantage of their friend relationships would not necessarily match a particular user interests (if a
has been studied in the context of traditional human social user follows accounts talking about technology, he/she
networks. For example, SONAR [Guy et al., 2009] recom- would not be interest in Ashton Kutcher, one of the most
mends related people in the context of enterprises by aggre- influential Twitter accounts according to Kwak et al.
gating information about relationships as reflected in differ- [2010])
ent sources within an organization, such as organizational Recommendation technologies applied to Twitter have
chart relationships, co-authorship of papers, patents, projects mainly focused on taking advantage of the massive amount
and others. Chen et al. [Chen et al., 2009] compared rela- of user-generated content as a novel source of preference
tionship-based and content-based algorithms in making and profiling information [Chen et al., 2010, Phelan et al.,
people recommendations, finding that the first ones are 2009, Esparza et al., 2010]. In contrast, we concentrate in
better at finding known contacts whereas the second ones recommending interesting people to follow. In this direc-
are stronger at discovering new friends. Weighted mini- tion, Sun et al. [2009] proposes a diffusion-based micro-
mum-message ratio (WMR) [Lo and Lin, 2006] is a graph- blogging recommendation framework which identifies a
based algorithm which generates a personalized list of small number of users playing the role of news reporters and
friends in social network build according to the observed recommends them to information seekers during emergency
interaction among members. Unlike these algorithms that events. Closest to our work are the algorithms for recom-
gathered social networks in enclosed domains, mainly start- mending followees in Twitter evaluated and compared using
ing from structured data (such as interactions, co-authorship a subset of users in [Hannon et al., 2010]. Multiple profiling
relations, etc.), we propose a people recommendation algo- strategies were considered according to how users are
rithms that take advantage of Twitter social structure popu- represented in a content-based approach (by their own
lated by massive, unstructured and user-generated content. tweets, by the tweets of their followees, by the tweets of
Understanding micro-blogging as a novel form of com- their followers, by the combination of the three), a collabor-
munication and news spreading medium has been one of the ative filtering approach (by the IDs of their followees, by
primary concerns of recent research efforts. Kwak et al. the IDs of their followers or a combination of the two) and
[2010] analyzed the topological characteristics of Twitter two hybrid algorithms. User profiles are indexed and rec-
and its power for information sharing, finding some diver- ommendations generated using a search engine, receiving a
gences between this follower/followees network and tradi- ranked-list of relevant Twitter users based on a target user
tional human social networks: follower distribution exhibit a profile or a specific set of query terms. Our work differs
non-power-law (users have more followers than predicted from this approach in that we do not require indexing pro-
by power-law), the degree of separation is shorter than ex- files from Twitter users; instead a topology-based algorithm
pected and there is a low reciprocity (most followers in explored the follower/followee network in order to find
Twitter do not follow their followers back). Other works candidate users to recommend.
addressed the problem of detecting influential users as a The main difference between existent work and our work
method of ranking people for recommendation. In the pre- is that the mentioned approaches for followee recommenda-
vious study it was found that ranking users by the number of tions, except for the approach presented in [Hannon et al.,
followers and by PageRank give similar results. However, 2010], were evaluated using datasets gathered from Twitter,
ranking users by the number of re-tweets indicates a gap with no assessment about the target user interest in the rec-
between influence inferred from the number of followers ommendations. In other words, the target user interest in a
and that inferred from the popularity of user tweets. Coinci- followee recommended that is not in the current list of the
dently, a comparison of in-degree, re-tweets and mentions target user’s followees cannot be assessed within these data-
as influence indicators carried out in [Cha et al., 2010] con- sets in order to determinate the correctness of the recom-
cluded that the first is more related to user popularity, whe- mendation. For this reason, the approach proposed in this
reas influence is gained only through a concentrated effort work was evaluated with a controlled experiment with real
in spawning re-tweets and mentions and can be hold over a users.
variety of topics. TwitterRank [Weng et al., 2010] tries to
find influential twitterers by taking into account the topical
similarity between users as well as the link structure, TU-
3 Followees Recommendations on Twitter
Rank [Yamaguchi et al., 2010] considers the social graph The algorithm we propose for recommending followees on
and the actual tweet flow and Garcia and Amatriain [2010] Twitter consists in two steps: (1) we explore the target us-
propose a method to weight popularity and activity of links er’s neighborhood in search of candidates and (2) we rank
for ranking users. candidates according to different weighting features. These
The influence rankings presented by studies on the com- steps are detailed in Sections 3.1 and 3.2, respectively.
plete Twittersphere have not direct utility for followee rec-
ommendation since people get connected for multiple rea- 3.1 Finding candidates
sons. We demonstrated with our experiments that indegree, The general idea of the algorithm we implemented is to
which has proven to be a good representation of a user’s suggest users that are in the neighborhood of the target user,

where the neighborhood of a user is determined from the
follower/followee relations in the social network.
In order to find candidate followees to recommend to a
target user U, we based our search algorithm on the follow-
ing hypothesis: The users followed by the followers of U’s
followees are possible candidates to recommend to U. In
other words, if a user F follows a user that is also followed
by U, then other people followed by F can be interesting to
U.
The rationale behind this hypothesis is that the target user
is an information seeker that has already identified some
interesting users acting as information sources, which are
his/her current followees. Other people that also follows
some of the users in this group (i.e. is subscribe to some of
the same information sources) have interests in common
with the target user and might have discover other relevant Figure 1: Scheme for finding candidate recommendations
information sources in the same topics, which are in turn
their followees. Since we seek for sources of information to recommend,
This scheme is outlined in Figure 1 and can be resumed we assume that this kind of users will have a lot of followers
in the following steps: and that they will follow few people. If user x has no follo-
1. Starting with the target user, we first obtain the list wees, then only the number of followers is considered with-
of users he/she follows, let’s call this list S. out changing the significance of the weighting feature.
S= Us We use this metric as a baseline for comparison with oth-
er metrics. Our aim is to demonstrate that metrics for rank-
s∈ followees (U )
ing popular users on Twitter are not good for ranking rec-
2. From each element in S we get its followers, let’s ommendations of users that a target user might be interest in
call the union of all these lists L following. In [Kwak et al., 2010] it has been shown that the
rankings of users that can be obtained by number of follow-
L = U followers ( s ) ers and by PageRank [Brin and Page, 1998] are very similar.
s∈S
We opted to use this factor as an estimator of the “impor-
3. Finally, from each element in L, we get its followees tance” of a given user because the number of followers is a
to obtain the list of possible candidates to recom- metric by far more easily to obtain that the user PageRank in
mend. Let’s call the union of all these lists T. a network with an order of almost 2 billion social relations.
The second feature explored corresponds to the number
T = U followees (l ) of occurrences of the candidate user in the final list of |R|
l∈L
candidates for recommendations, as shown in Equation 2.
4. Exclude from T those users that the target user al-
ready follows. Let’s call the resulting list R. | {i ∈ R / i = x} |
R =T −S wo ( x) = (2)
|R|
Each element in R is a possible user to recommend to the
target user. Notice that each element can appear more than The number of occurrences of a given user x in this final
once in R, depending on the number of times that each user list is an indicator of the amount of (indirect) neighbors that
appears in the followees or followers lists obtained at steps also have x as a (direct) connection itself.
2 and 3 above. The third feature we considered is the number of friends
in common between the target user U and the candidate
3.2 Weighting features recommendation x:
Once we find the list R of candidate recommendations for
the target user, we explored different features to give a score wc ( x ) = followees ( x ) ∩ followees (U ) (3)
to each unique user x∈R.
The first feature explored is the relation between the
number of followers a user has with respect to the number Finally, we considered two combinations of these fea-
of users the given user follows, as shown in Equation 1. tures: the average of the three features, and their product:

followers( x) 1
w f ( x) = (1) ws ( x) = [ wo ( x) + w f ( x) + wc ( x)] (4)
followees( x) 3

link could be used to read the tweets published by the rec-
w p ( x) = wo ( x ) * w f ( x ) * wc ( x ) (5) ommended user in the case that the information provided by
the application was not enough to determine the student’s
It is worth noticing that the selection of these weighting interest in the recommendation. The question we asked
features was not arbitrary. Our choice was based on a deep students to ask themselves to determine whether a recom-
analysis of previous studies about Twitter and particular mendation was relevant or not was “Would you have fol-
properties of this specific network that makes general link lowed this recommended user in the first place (when select-
prediction approaches unsuitable. All the studies about the ing which users to follow in the first part of the experiment),
properties of the Twitter network agree in that there is a if you had know this account?” For example, if a given
minimal overlap with the features available on other online student was interest in technology and he/she had not dis-
social networks (OSNs). covered the account @TechCrunch during his/her first se-
lection of followees, that would be an interest recommenda-
4 Experiment setting tion because @TechCrunch tweets about news on technolo-
gy.
To evaluate the proposed algorithms, we have carried out a
preliminary experiment using a group of 14 users. These
users, 8 males and 6 females, were in the last years of their
5 Results
course of studies and were students of a Recommender We first evaluated the performance of the proposed algo-
Systems related course dictated at our University as an elec- rithm in terms of their overall precision in followees rec-
tive course during 2010. The students selected for the expe- ommendation. Precision can be defined as the number of
riment were volunteers familiarized with Twitter. relevant recommendations over the number of recommenda-
During the first part of the course, we asked these users to tions presented to the user and it can be also computed at
create a Twitter account and to follow at least 20 Twitter different positions in the ranking. For example, P@5 (“pre-
users who publish information or news about a set of partic- cision at five”) is defined as the percentage of relevant rec-
ular subjects of their interest. The general interests ex- ommendations among the first five, averaged over all runs.
pressed by users ranged between diverse subjects such as Figure 2 shows the precision achieved by the algorithm,
technology, software, math, science, football, tennis, basket, averaged between all users, for each weighting feature at
religion, movies, journalists, government, music, cooking, four different positions of the ranking: P@1, P@5, P@10
shoes, TV programs and even other students in their faculty. and P@20. The results of considering each feature separate-
Some users only concentrated on one particular subject ly and the two aggregations functions are showed in this
while others distributed their followees among several top- figure.
ics.
Then, we used the user IDs of the user accounts created
by the students as seeds to crawl a sub-graph of the Twitter
network corresponding to three levels of both followee and
follower relations, centered on each seed. The resulting
dataset consisted on 1,443,111 Twitter users and 3,462,179
following relations already existing among them.
During the second part of the course, we provided these
users with a desktop tool that allowed them to login to Twit-
ter and ask for followees recommendations. Since the users
who participated in the experiment were students of a “re-
commender systems” course, all of them had knowledge
about concepts such as rankings and metrics. As part of a
not compulsory practical exercise of the course they were
motivated to discover which metric better ranked recom-
mendation results and to write a brief report about the re-
sults they obtained for their particular case. The desktop
application provided for this exercise allowed students to
select the weighting feature by which they liked to rank Figure 2: Average precision for each weighting feature
recommendations, with no predefined order.
In all cases, 20 recommendations were presented to the We can observe several interesting facts in the results
users. Then, we asked the students to explicitly evaluate presented in Figure 2. First, it results that wo(x), the weight-
whether the recommendations were relevant or not according feature considering the number of occurrences of a user
ing to the same topical criteria they have chosen to select in the list of recommendations as gathered by the algorithm
their followees as information sources in the first place. For proposed, generates better precision scores than any other
each recommendation in the resulting ranking the applica- weighting feature explored. For this weighting feature we
tion showed the user name, description, profile picture and a obtained a good recommendation in the first position of the
link to the home page of the corresponding account. This ranking for 93% of the users. For longer ranking lists, preci-

sion decrease from 0.73 for P@5 to 0.64 for P@20, which concentrate relevant recommendations in the top positions
we believe are all good results. of the ranking. On contrary, we can observe that for wo(x)
It is worth noticing that although we reported results up to and wp(x) relevant recommendations tend to be clustered
P@20, recommendations lists tend to be shorter (frequently towards the top of the ranking.
5) in order to help the user to focus on the most relevant Although precision measure gives a general idea of the
results. In these small lists the algorithm reached good le- overall performance of the presented weighting features, it
vels of precision, recommending mostly relevant users. is also very important to consider the position of relevant
The weighting feature considering the number of follow- recommendations in the ranking presented to the user. Since
ers, wf(x), got the worst precision scores, with values under it is known that users focus their attention on items at the
30%. This fact reveals that this metric, although widely used top of a list of recommendations [Joachims, 2005], if rele-
in other approaches as mentioned in Section 2, is only good vant recommendations appear at the top of the ranking using
at measuring a user’s general popularity in the entire Twitter one algorithm and at the bottom of the ranking using the
network, but popularity does not necessarily translate into other, the first algorithm will be perceived as better per-
relevance for a particular user. Celebrities and politicians, forming by users even though their general precision might
such as Barack Obama (@barackobama), Lady Gaga be similar.
(@ladygaga), Yoko Ono (@yokoono), and Tom Cruise Discounted cumulative gain (DCG) is a measure of effec-
(@tomcruise) were a common factor in the rankings of tiveness used to evaluate ranked lists of recommendations.
many users regardless their particular interests. Among DCG measures the usefulness, or gain, of a document based
other popular users suggested that in some cases met the on its position in the result list using a graded relevance
user’s interests were popular blogs and news media such as scale of documents in a list of recommendations. The gain is
Mundo Geek (@mundo_geek), C5N (@C5N), El Pais, accumulated from the top of the result list to the bottom
(@el_pais), Mundo Deportivo (@mundodeportivo), Red with the gain of each result discounted at lower ranks. The
Hat News (@redhatnews) and Fox Sport LA premise of DCG is that highly relevant documents appear-
(@foxsportslat). ing lower in a list should be penalized as the graded relev-
A similar situation occurs with wc(x), the weighting fea- ance value is reduced logarithmically proportional to the
ture considering the number of friends in common between position of the result. The DCG accumulated at a particular
the target user U and the candidate user to recommend to U. rank position k is defined as shown in Equation 6:
Although precision is better than wf(x) for every size of the
recommendation lists, this weighting feature does not reach k reli (6)
the performances obtained with wo(x). This result is ex- DCG @ k = rel1 + ∑i =2
pected since the fact that two users U and X share a friend Y log 2 i
does not necessarily means that X is a good information DCG is often normalized using an ideal DCG vector that
source. has value 1 at all ranks. Figure 3 shows the normalized
We also found that ws(x) tends to perform poorly. This DCG obtained for both algorithms at four different positions
score is affected by the term corresponding to the relation of the ranking: nDCG@1, nDCG@5, nDCG@10 and
between the number of followers and the number of follo- nDCG@20.
wees, which in most cases is higher than the other terms
involved. This factor highly affects the overall average
among the three weighting features, causing a decrease in
precision.
The second score which combines the three weighting
features, wp(x), seems to overcome this problem since in
this case each weighting feature is multiplied to obtain the
final score. Nevertheless, celebrities and very popular Twit-
ter user accounts also tend to appear at the top positions of
the ranking diminishing the general precision again. How-
ever, the factor corresponding to wo(x) also makes good
recommendations to appear interleaved with some popular
users on Twitter.
Another interesting issue observed in the results pre-
sented in Figure 2 is that for both wf(x) and wc(x) precision
tend to keep almost constant across different sizes in the list
of recommendations and even with a slightly increment as
the size of the recommendation set increases. This fact
seems to contradict the definition of precision in the infor- Figure 3: Normalized Discounted Cumulative Gain at rank k
mation retrieval sense which, by principle, should decrease for each weighting feature
as the number of recommendations increases. However, this
behavior occurs because all wf(x), ws(x) and wc(x) does not

nDCG@1 is equivalent to P@1 by definition. Then, we them by the number of occurrences of each candidate in the
can see that scoring users with wo(x) always positions rele- list generated by this method. Among the advantages of this
vant users above in the ranking than other weighting fea- method when compared with content-based alternatives is
tures, seconded by wp(x). that recommendations can be found quickly based on a
Success at rank k (S@k) is another metric commonly simple analysis of the network structure, without consider-
used for ranked lists of recommendations. The success at ing the content of the tweets posted by the candidate user.
rank k is defined as the probability of finding a good rec- Nevertheless, we also believe that combining the proposed
ommendation among the top k recommended users. In other method with an analysis of the content of the tweets posted
words, S@k is the percentage of runs in which there was at by a user in the list of candidates can improve the precision
least one relevant user among the first k recommended us- of a followee recommender system, at the expense of com-
ers. Figure 4 shows the results we obtained for this metric putational performance.
with values of k ranging from 1 to 10.

Figure 5: Mean Reciprocal Rank for each weighting feature
Figure 4: Success at rank k for each weighting feature
6 Comparison with related work
For S@k we can observe results equivalent to nDCG@k. From the related work, the approach that we find more
Again, scoring users with wo(x) always positions relevant similar to ours (and the only one, up to our knowledge that
users above in the ranking than the other weighting features. experimented with real users in a controlled experiment) is
The ranking according wp(x) allowed users to find a relevant Twittomender, proposed by Hannon et al. [2010]. Although
recommendation always at the most at position 4 in the the results presented in [Hannon et al., 2010] are not fully
ranking, while for ws(x) we obtain success 1 at position 6. comparable to the results presented in this article since dif-
With this metric we can confirm that wf(x) and wc(x) are not ferent datasets were used, in this section we present a com-
good weighting factors by their own. parison about the precision reported for Twittomender and
To study further the algorithm ability to rank followees the precision obtained with our approach.
for recommendation, we used Mean Reciprocal Rank Twittomender create different indexes for all users in the
(MRR), a metric that measures where in the ranking is the dataset generated from different sources of profile informa-
first relevant recommendation. If the first relevant recom- tion. Four of these indexes are content-based, modeling
mendation is at rank r, then the MRR is 1/r. This measure users by their own tweets, by the tweets of their followers,
averaged over all runs provides insight in the ability of the by the tweets of their followees and by a combination of the
system to recommend a relevant user to follow in Twitter at three. The three remaining strategies are topology-based and
the top of the ranking. Figure 5 plots the MMR measure for model users by the IDs of their followees, by the IDs of
both proposed algorithms. their followees and by a combination of both.
This metric gives us another view of which weighting The strategy used for ranking users in the online experi-
feature generates better ranking of recommendations. We ment presented in [Hannon et al., 2010] generates the seven
confirm that wo(x) always ranks users better than the other rankings according to the different approaches described
proposed weighting features, while ranking users by their above and then generate a single ranking by merging those
“popularity” does not generate good recommendations. seven rankings. When merging the rankings they use a scor-
The experiments presented make us believe that there is ing function that is based on the position of each user in the
reason to be optimistic about the potential for a followee recommendation lists. In this way users that are frequently
recommender for Twitter using the method described in present in high positions are preferred over users that are
Section 3.1 to obtain a list of candidates and simple ranking recommended less frequent or in lower positions.

Hannon et al. performed a live user trial with 34 users, vector is then compared with the vector of terms corres-
reporting a precision of about 38.2% for k=5 and 33.8% for ponding to each candidate recommendation and the similari-
k=10. Table 1 summarizes the comparison between Twitto- ty obtained is considered in the generation of the ranking.
mender and our system. Notice that precision values for The results reported in this article make us feel really
Twittomender system are approximate because they were enthusiastic about the potentials of Twitter for building
taken (and in some cases computed) from the graphics pre- recommender systems of sources of information.
sented in the article.
It is worth noticing that although the number of volun- Twittomender Our
teers who participated in Twittomender experiment is more “live-user” trial approach
than twice the number of volunteers who participated in our
#seeds 34 14
experiment, the number of Twitter users involved in our
#users 100,000 1,443,111
experiment is by far higher than the number of users in their
database. Furthermore, Twittomender can only recommend P@5 ~38.2% 72.9%
users that are previously indexed. When a user is registered P@10 ~33.8% 67.9%
into the system, all his/her followees and followers profiles P@20 ~26.9% 64.3%
along with his/her own profile are indexed. Our work differs
from this approach in that we do not require indexing pro- Table 1: Comparative table between Twittomender and our
files from Twitter users; instead a topology-based algorithm approach
explores three levels of the follower/followee network in
order to find candidate users to recommend. References
[Brin and Page, 1998] S. Brin and L. Page. The anatomy of
7 Discussion and Conclusion a large-scale hypertextual Web search engine. Computer
In this article we presented a simple but effective algorithm Networks and ISDN Systems. Volume 30, Issue 1-7, pag-
for recommending followees in the Twitter social network. es 107-117. 1998.
This algorithm first explores the target user neighborhood in [Cha et al., 2010] M. Cha, H. Haddadi, F. Benevenuto,
search of candidate recommendations and then sorts these and K. Gummadi. Measuring user influence in Twitter:
candidates according to different weighting features: the The million follower fallacy. In Proceedings of the 4th
relation between the number of followers and the number of International Conference on Weblogs and Social Media
followees, the number of occurrences of each candidate in (ICWSM’10), Washington DC, USA, 2010.
the final list, the number of friends in common, and two [Chen et al., 2009] J. Chen, W. Geyer, C. Dugan, M. Mul-
combinations of the three features. ler, and I. Guy. Make new friends, but keep the old: re-
We evaluated the proposed algorithm with real users and commending people on social networking sites. In Pro-
we obtained satisfactory results in finding good followee ceedings of the 27th International Conference on Human
recommendations. We found that considering just the over- Factors in Computing Systems, pages 201–210, 2009.
lapping users among the different lists of follower and fol-
lowees explored by our crawling method gives better results [Chen et al., 2010] J. Chen, R. Nairn, L. Nelson, M.
than the other features considered. As expected, the inde- Bernstein, and E. Chi. Short and tweet: experiments on
gree of a user is not a good feature for ranking followee recommending content from information streams. In
recommendations. Considering the number of followers for Proceedings of the 28th International Conference on
ranking users put celebrities and popular Twitter accounts at Human Factors in Computing Systems (CHI’10), pages
the top of the list, but these recommendations are not neces- 1185–1194, 2010.
sarily interesting for a particular user. However, there are [Esparza et al., 2010] S. Garcia Esparza, M. P. O’Mahony,
some interesting recommendations discovered by this fea- and B. Smyth. On the real-time web as a source of rec-
ture, such as top bloggers who write about a particular sub- ommendation knowledge. In Proceedings of the 4th
ject or news media accounts. ACM Conference on Recommender Systems (Rec-
Although the results reported seems promising, we are Sys’10), pages 305–308, 2010.
planning to repeat the experiment this year in order to in-
[Garcia and Amatriain, 2010] R. Garcia and X. Amatriain.
volve more users in the experiment and obtain more statis-
Weighted content based methods for recommending
tical support for the results reported. Moreover, we are very
connections in online social networks. In Workshop on
optimistic about the potential improvements that we can
Recommender Systems and the Social Web, pages 68–71,
obtain by extending the presented approach with content-
Barcelona, Spain, 2010.
based techniques. A natural extension of our approach in
which we are currently working on is a hybrid algorithm [Guy et al., 2009] I. Guy, I. Ronen, and E. Wilcox. Do you
that filters the candidate recommendations found with the know?: recommending people to invite into your social
topology-based method with a content-based analysis of the network. In Proceedings of the 13th International Con-
tweets posted by the users. In this new approach, a target ference on Intelligent User Interfaces (IUI’09), pages
user U is modeled with a vector of terms built from a con- 77–86, 2009.
tent analysis of the tweets posted by U’s followees. This

[Hannon et al., 2010] J. Hannon, M. Bennett, and B.
Smyth. Recommending Twitter users to follow using
content and collaborative filtering approaches. In Pro-
ceedings of the 4th ACM Conference on Recommender
Systems (RecSys’10), pages 199–206, 2010.
[Java et al., 2007] A. Java, X. Song, T. Finin, and B. Tseng.
Why we twitter: understanding micrblogging usage and
communities. In Proceedings of the 9th WebKDD and
1st SNA-KDD 2007 workshop on Web mining and social
network analysis (WebKDD/SNA-KDD '07). ACM, New
York, NY, USA, pages 56-65. 2007.
[Joachims et al., 2005] T. Joachims, L. Granka, B. Pan, H.
Hembrooke, and G. Gay. Accurately interpreting click-
through data as implicit feedback. In Proceedings of the
28th annual international ACM SIGIR conference on
Research and development in information retrieval (SI-
GIR '05). ACM, New York, NY, USA, 154-161. 2005.
[Krishnamurthy et al., 2008] B. Krishnamurthy, P. Gill, and
M. Arlitt. A few chirps about twitter. In Proceedings of
the first workshop on Online social networks (WOSP
'08). ACM, New York, NY, USA, pages 19-24. 2008.
[Kwak et al., 2010] H. Kwak, C. Lee, H. Park, and S.
Moon. What is Twitter, a social network or a news me-
dia? In Proceedings of the 19th International Conference
on World Wide Web (WWW’10), pages 591–600, 2010.
[Lo and Lin, 2006] S. Lo and C. Lin. WMR–A graph-based
algorithm for friend recommendation. In Proceedings of
the 2006 IEEE/WIC/ACM International Conference on
Web Intelligence (WI’06), pages 121–128, Washington,
DC, USA, 2006.
[Phelan et al., 2009] O. Phelan, K. McCarthy, and B.
Smyth. Using Twitter to recommend real-time topical
news. In Proceedings of the 3rd ACM Conference on
Recommender Systems (RecSys’09), pages 385–388,
2009.
[Sun et al., 2009] A. R. Sun, J. Cheng, and D. D. Zeng. A
novel recommendation framework for micro-blogging
based on information diffusion. In Proceedings of the
19th Workshop on Information Technologies and Sys-
tems, 2009.
[Weng et al., 2010] J. Weng, E-P. Lim, J. Jiang, and Q. He.
TwitterRank: finding topic-sensitive influential twitter-
ers. In Proceedings of the 3rd ACM International Confe-
rence on Web Search and Data Mining (WSDM’10),
pages 261–270, New York, NY, USA, 2010.
[Yamaguchi et al., 2010] Y. Yamaguchi, T. Takahashi, T.
Amagasa, and H. Kitagawa. TURank: Twitter user rank-
ing based on user-tweet graph analysis. In Web Informa-
tion Systems Engineering, volume 6488 of LNCS, pages
240–253, Hong Kong, China, 2010.

A topology based approach twittersdlfkjsdlkfj

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to A topology based approach twittersdlfkjsdlkfj (20)

A topology based approach twittersdlfkjsdlkfj