ANTI-FOLKSONOMICAL ITEM RECOMMENDATION
SYSTEM BASED ON SIMILARITIES BETWEEN ITEM CLUSTERS
IN SOCIAL BOOKMARKING
Akira Sasaki, Takamichi Miyata, Yoshinori Sakai
Department of Communications and Integrated System
Graduate School of Science and Engineering, Tokyo Institute of Technology
2-12-1-S3-67, Ookayama, Meguro-ku, Tokyo, Japan
Yasuhiro Inazumi
Graduate School of Science and Engineering, University of Toyama
3190, Gofuku, Toyama-shi, Toyama, Japan
Aki Kobayashi
Department of Computer Science and Communication Engineering, Kogakuin University
1-24-2 Nishishinjuku, Shinjuku-ku, Tokyo, Japan
Keywords:
Social bookmarking, Recommendation system, Collaborative filtering.
Abstract:
Web-based bookmark management services called social bookmarking has been in the spotlight recently. So-
cial bookmarking allows users to add several keywords called tags to items they bookmarked. Many previous
works on social bookmarking using actual words for tags, called folksonomy, have come out. However, essen-
tial information of tags is not represented in their tag names, but in the classification of items by tags. Based
on this assumption, we propose an anti-folksonomical recommendation system for calculating similarities be-
tween groups of items classified according to tags. In addition, we use hypothesis testing to improve these
similarities based on statistical reliability. The experimental results show that our proposed system provides
an appropriate recommendation result even if users tagged with different keywords.
1 INTRODUCTION
Recommendation systems have been widely re-
searched (Li and Zaiane, 2004) (Kazienko and
Kiewra, 2004) (Ishikawa et al., 2002) (Gunduz and
Ozsu, 2003). Most of them are based on ‘collabo-
rative filtering’, which is a method for predicting a
specific user’s preferences for new items using pref-
erences obtained from many other users (Goldberg
et al., 2003) (Resnick et al., 1994) (Sarwar et al.,
2001).
Generally speaking, collaborative filtering is de-
fined as a method for estimating preferences for items
that users have not yet found by comparing prefer-
ences of items that they have already browsed, not by
using item context. We use the term ‘collaborative
filtering’ in a narrow sense as a method for calculat-
ing preferences as the rating of each unknown item
against preferences of previously viewed items. The
algorithm is as follows:
1. Collect preferences of viewed items.
2. Calculate similarities between the focus user and
others based on their preferences of commonly
viewed items.
3. Calculate preference of each unknown item based
on similarities between the focus user and others
who have already viewed it.
As Sarwar et al. acutely pointed out, more items than
users leads to poor recommendation results because
of the sparsity of preference data (Sarwar et al., 2001).
As other recommendation systems, novel web
services called social bookmarking (SBM) have ap-
peared in recent years. SBM allows users to annotate
610
Miyata T., Sakai Y., Inazumi Y., Kobayashi A. and Sasaki A.
ANTI-FOLKSONOMICAL ITEM RECOMMENDATION SYSTEM BASED ON SIMILARITIES BETWEEN ITEM CLUSTERS IN SOCIAL BOOKMARKING.
DOI: 10.5220/0001824206050612
In Proceedings of the Fifth International Conference on Web Information Systems and Technologies (WEBIST 2009), page
ISBN: 978-989-8111-81-4
Copyright
c
2009 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
each item with one or more keywords called ‘tags’.
Niwa et al. investigated the recommendation sys-
tem by using tags created by SBM users (Niwa et al.,
2006). They aggregated similar tags to reduce word
redundancyand made tag clusters with keywords hav-
ing the same meaning.
However, a keyword may have various meanings
depending on the context. Golder et al. pointed out
these types of tags as being polysemous (Golder and
Huberman, 2005). For example, ‘apple’ has multiple
meanings: a sweet red fruit or the consumer electron-
ics company. If the recommendation system includes
these types of tags in only one tag cluster, it cannot
recommend items to users who use the tag in other
contexts. Due to this problem, a vocabulary-based
recommendation system may lead to inappropriate re-
sults.
Therefore, we did not focus on the vocabulary of
tags and instead propose tag-based collaborative fil-
tering. In addition, we define the novel similarity be-
tween item clusters based on hypothesis testing.
The rest of this paper is organized as follows. In
section 2, we introduce a related study of ours. Next,
we explain our recommendation system algorithm in
section 3. In section 4, we evaluate and discuss our
recommendation system, and we conclude in section
5.
2 RELATED STUDY
In this section, we introduce a related study of con-
ventional recommendation systems using an SBM
service and based on the co-occurrence of items.
2.1 Social Bookmarking (SBM)
2.1.1 What is SBM?
SBM services enable users to store, organize, search,
manage, and share bookmarks of web pages on the
Internet. SBM has a specific feature called ‘tags’.
Tags are keywords created by each SBM user for cat-
egorization of web pages. In SBM services, users
can share their bookmarks with other users and also
browse other users’ bookmarks. For example, users
can browse a list of SBM users who bookmarked the
same item as they did and also browse a list of items
that are tagged with the same keyword.
2.1.2 Folksonomy
‘Folksonomy’, from folks’ and ‘taxonomy’, is a
method of collaboratively managing tags to catego-
Rank of tag name by frequency
Frequency of tag name
Figure 1: Tag name frequencies in descending order.
rize itemsDFolksonomy describes the bottom-upclas-
sification systems that emerge from social tagging
1
.
Compared to conventional ‘taxonomy’, Shirky ar-
gues that folksonomy has an advantage in flexibility
for changing the consensus of how objects should be
classified because tagging systems do not use a con-
trolled vocabulary
2
.
2.2 Conventional Study on SBM
Over the past few years, a considerable number of
studies have been conducted on SBM (Golder and
Huberman, 2005) (Mika, 2005) (Yanbe et al., 2007)
(Hotho et al., 2006) (Jaschke et al., 2007) (Niwa et al.,
2006).
Niwa et al. proposed a web page recommenda-
tion system based on folksonomy (Niwa et al., 2006).
This system is used for calculating the ‘affinity level’,
which is a scalar value representing the relationship
between users and tag clusters (sets of tags for reduc-
ing word redundancy). In other words, it recommends
items based on the similarities in users’ tag records.
However, there is empirical evidence against this
idea of recommendation. A histogram of tags at-
tached to a particular item from an actual SBM ser-
vice
3
is shown in Fig. 1. We sorted the tags by
unique vocabulary in descending order. The tag
name frequencies clearly exhibit long-tail behavior.
These are two types of words: top-ranked words (a
few words used by many people) and bottom-ranked
words (many words used by few people). The im-
1
A. Mathes,
Folksonomies Cooperative Classification and Communi-
cation Through Shared Metadata,
http://www.adammathes.com/academic/computer-
mediated-communication/folksonomies.html
2
C. Shirky,
Ontology is Overrated: Categories, Links, and Tags,
http://www.shirky.com/writings/ontology overrated.html
3
delicious, http://delicious.com/
ANTI-FOLKSONOMICAL ITEM RECOMMENDATION SYSTEM BASED ON SIMILARITIES BETWEEN ITEM
CLUSTERS IN SOCIAL BOOKMARKING
611
Keyword space
Item space
Figure 2: Relationship between items and keywords.
portant point to note is that the latter situation causes
word redundancy.
Let us look more closely at the bottom-ranked
words. We can find unusual tags like ‘toread’ (sticky
note for future reference. Golder et al. also pointed
it out as ‘Task Organizing’(Golder and Huberman,
2005)), web/app’ (users’ own directory hierarchy),
and so on. That is, in an actual SBM service, most
users’ tags are not for others’ convenience but for
themselves to manage their own bookmarks for their
future use.
Let us now return to Niwa et al.s study. The rec-
ommendation system of using tag names is implicitly
based on the collaboration of SBM users. The system
then discards the bottom-ranked words, but retains a
large percentage of the total tags in the practical SBM
service.
We will discuss SBM tagging in detail. An exam-
ple of the relationship between items and keywords
is shown in Fig. 2. The word redundancy is clear.
Users use the words ‘robots’ and AI’ for tagging the
same concept of ‘machines that can think’. In ad-
dition, there is another word redundancy. Users use
the same word, ‘robot’, for tagging two different con-
cepts, ‘machines that can think’ and ‘automated sys-
tems’. A recommendation system using only the co-
occurrence of words would never recommend items
in these cases.
In contrast, our proposed system works even if
users tag with different keywords because it does not
pay attention to the vocabulary of the tags. Instead, it
is rather similar to systems based on co-occurrence of
items.
2.3 Conventional System based on
Co-occurrence of Items
Rucker and Polanco developed a system for recom-
mending items by calculating similarities between
folders (categories by user) of bookmarks (Rucker
and Polanco, 1997). Their system is similar to our
proposed system from the viewpoint of only using
sets of items. However, their system does not rank
each recommended item.
A comparison between our proposed system and
conventional systems discussed previously is shown
in Table 1. As can be seen, our system calculates
similarities between item clusters by using hypothe-
sis testing for finding similar ones. Furthermore, our
system calculates the recommendation rate; therefore,
it can rank each item.
3 PROPOSED SYSTEM
3.1 Recommendation based on SBM by
using “Item Cluster”
In this paper, we focus on item clusters’, which are
sets of items classified by the tags used by each user.
Each user has the same number of item clusters as
the number of tags he/she uses in the SBM service.
When a user issues a query by selecting a tag from
his/her tag records, the system searches for items to
recommend by focusing on the similarities between
query item cluster corresponding to the query and
recommender item clusters correspondingto other
tags in the scope of commonly bookmarked items.
3.2 Model of Item Cluster
We focused on a particular tag t
query
that is tagged by
user u
focused
. B
s
refers to all the items bookmarked by
u
focused
, and T
s
refers to all the items tagged t
query
by
u
focused
. All of the items A (bookmarked by all users
in the SBM service) can be classified into three sets
exclusively, as shown in Fig. 3.
1. Bookmarked by u
focused
, and tagged with t
query
(T
s
)
2. Bookmarked by u
focused
, but tagged without t
query
(B
s
T
s
)
3. Not bookmarked by u
focused
(B
s
)
We define a set of items tagged with a certain tag
name, like T
s
as an ‘item cluster’. Let us consider
two item clusters, ‘query item cluster T
s
’ and ‘recom-
mender item cluster T
o
’. T
s
is an item cluster tagged
with t
query
by u
focused
, and T
o
is tagged with t
j
by u
i
(note that u
i
is not u
focused
). We studied the concep-
tual similarity between T
s
and T
o
. Here, k is the num-
ber of items that are included in both T
s
and T
o
. m
s
is
the number of items in T
s
that u
i
tags with a different
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
612
Table 1: Comparison between proposed method and conventional methods.
Collaborative Conventional Conventional Proposed
filtering (Niwa et al., 2006) (Rucker and Polanco, 1997) method
Focused on user tag user and category user and tag
Co-occurrcence of item tag (records) item item
Rating yes yes no yes
Similarity co-occurence ratio co-occurrence ratio not defined likelihood ratio
query
query
focused
focused
focused
Figure 3: SBM modeling regarding relationship among
items, users, and tags.
tj
tquery
focused
n
k
Figure 4: Item recommendation by comparing item clus-
ters.
tag name from t
j
. m
o
is the number of items in T
o
that
u
focused
tags with a different tag name from t
query
. Let
m = m
s
+m
o
and n = m+k. The relationship between
k and n shows a conceptual similarity of the two item
clusters. n and k are shown in Fig. 4 and described as
n = |(B
s
B
o
) (T
s
T
o
)| (1)
k = |T
s
T
o
|. (2)
Next, we look at the expected similarity of the two
item clusters. Here, we assume that there are only two
relationships between item clusters similar view-
point and different viewpoint. If two users tag items
from similar viewpoints, the expected probability that
both users tag the same item is assumed to be p
1
. Oth-
erwise, the expected probability is assumed to be p
0
.
Here, p
1
> p
0
. p
1
and p
0
can be estimated by ob-
serving all item clusters to separate them into similar
and different viewpoints. Desirable items should be
recommended from similar item clusters. The con-
ceptual similarity between T
s
and T
o
is defined by
sim(T
s
, T
o
) = log
L(n,k, p
1
)
L(n,k, p
0
)
= klog
p
1
p
0
+ (n k)log
1p
1
1p
0
,
(3)
where
L(n, k, p) =
n
C
k
p
k
(1 p)
nk
. (4)
The log likelihood ratio of whether the similarity
of two clusters is likely to be p
1
or p
0
is shown in Eq.
3. Here we assume the relationship between k and n
follows a binomial distribution (Eq. 4) with param-
eters p
1
and p
o
for similar and different viewpoints,
respectively.
Finally, we define the recommendation rate of
each item by using similarities between the item clus-
ters. The system selects one item i as the candidate
to be recommended from B
s
(i B
s
). We define is
recommendation rate by calculating the sum of simi-
larities between the query item cluster and each rec-
ommender item cluster that contains i.
R(T
s
, i) =
T
o
T
all
χ(T
o
, i)sim(T
s
, T
o
) (5)
χ(T
o
, i) =
1 if i T
o
sim(T
s
, T
o
) 0
0 otherwise
(6)
where T
all
is the set of all item clusters.
3.3 Procedure of the Proposed System
The recommendation algorithm is as follows:
u
focused
issues a query by selecting tag name t
query
from his/her tag records. we define it as T
s
.
1. Calculate each T
o
of sim(T
s
, T
o
) as Eq. 3.
2. Calculate each χ(T
o
, i) as Eq. 6.
3. Calculate each i of R(T
s
, i) as Eq. 5.
4. Sort items according to recommendation rate and
recommend i whose R(T
s
, i) is top th
rec
.
Here, th
rec
is the number of items to be recommended.
Fig. 5 is an example of the procedure. T
a
, T
b
and
T
c
are item clusters. Now we consider a recommenda-
tion for T
a
. That is, we define T
a
as query item cluster
ANTI-FOLKSONOMICAL ITEM RECOMMENDATION SYSTEM BASED ON SIMILARITIES BETWEEN ITEM
CLUSTERS IN SOCIAL BOOKMARKING
613
2 3
5
5
0
1
4
R( , ) = 5
a
T
i
1st.
2nd.
(T , )= 1
T
b
i
(T , )= 0
T
c
i
R( , ) = 6
a
T
i
(T , )= 1
T
c
i
(T , )= 1
T
b
i
1
1
1
2
2
2
a
T
T
c
i
i
1
2
a
T
i
T
b
i
1
2
sim( , )= 5
T
b
a
T
T
1
sim( , )= 1
T
c
a
T
Figure 5: Procedure of the proposed system.
(T
s
) and the others as recommender item clusters (T
o
).
i
1
and i
2
are the items to be recommended. That is,
the user who makes T
a
has not bookmark these items.
First, the system calculate similarity between
query item cluster T
a
and each recommender item
clusters T
b
, T
c
. Next, it checks every χ. For ex-
ample, i
1
is included in T
b
so χ(T
b
, i
1
) = 1. On
the other hand, T
c
does not include i
1
so χ(T
c
, i
2
) =
0. Finally, it calculates each recommendation rate
R(T
a
, i
1
), R(T
a
, i
1
) based on sum of products which
are computed by multiplying similarity by χ. Thus,
R(T
a
, i
1
) = sim(T
a
, T
b
)χ(T
b
, i
1
) + sim(T
a
, T
c
)χ(T
c
, i
1
).
4 EXPERIMENTS
We performed three experiments using live data ob-
tained from del.icio.us, which is one of the most fa-
mous SBM service sites. In these experiments, we set
p
0
= 0.1, p
1
= 0.6.
4.1 Performance Evaluation
We randomly collected data of 1,000 SBM users in
August, 2006. They had bookmarked about 310,000
unique items (URLs) and had tagged items with about
260 keywords on average. Therefore, we collected
about 260,000 sets of item clusters. We used all of
these item clusters for the experiments.
4.1.1 Evaluation Method
We used the collected data for calculating similarities
and for evaluating our system. We masked tag infor-
mation, that is, we hid whether all items included in
B
s
were tagged or not and revealed them after a rec-
ommendation was made.
The evaluation method was as follows.
1. Select T
s
from the collected data. We defined the
items included in T
s
as correct class X.
Figure 6: Recall, precision, and F-measure for each number
of recommended items.
2. Calculate recommendation rate of each item cor-
responding to B
s
, and recommend items from top
to th
rec
-th. We defined these items as recom-
mended class R.
3. Count the number of items X, R, and R X, then
calculate the recall and precision.
Recall and precision were defined as follows.
recall =
R X
X
(7)
precision =
R X
R
(8)
In addition, we used F-measure, defined as follows.
F-measure =
2· recall· precision
recall+ precision
. (9)
In this evaluation, we focused on the top 100
query item clusters, which were ranked by the number
of items in T
s
. The biggest number of B
s
was 17,960,
and the smallest was 865. The average was 6,758.4.
The biggest number of T
s
in the query item clusters
was 1,991, and the smallest was 477. The average
was 729.28.
Note that we omitted isolated items that had been
tagged only by T
s
to evaluate net performance.
4.1.2 Experimental Results
The results of the evaluation are shown in Figs. 6 and
7. The averages of recall, precision, and F-measure
for each query cluster are shown in Fig. 6, and the re-
lationship between recall and precision for item clus-
ters 1 to 8 (Table 2) is shown in Fig. 7.
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
614
Table 2: Details of item clusters obtained from experimental results.
Cluster 1 Cluster 2 Cluster 3 Cluster 4 Cluster 5 Cluster 6 Cluster 7 Cluster 8
u
focused
user316 user87 user796 user878 user555 user313 user190 user51
|B
s
| 2006 1632 2242 10078 17965 4782 3925 2701
t
query
web randomlink music javascript art History Shopping funny
|T
s
| 486 178 52 424 448 209 139 147
Recall
Precision
Cluster 1
Cluster 5
Cluster 2
Cluster 3
Cluster 7
Cluster 8
Cluster 4
Cluster 6
Figure 7: Relationship between recall and precision for
each item cluster
4.1.3 Discussion
In Fig. 6, we can see the precision was 0.78 when
our proposed system recommended the top 100 items
ranked by their similarities. The precision decreased
to 0.67 for 200 items. The F-measure, which is an im-
portant measure for recommendation, peaked at about
the top 400. Judging from the above, we may say that
our proposed system is useful for at least the top 400
items, which is enough for a recommendation system.
The coefficient of correlation between the number
of items in the item clusters and the maximum of the
F-measure is -0.34. This fact shows the robustness of
our proposed system. Its robustness is also evident
in Fig. 7, especially in clusters 1 to 3 and clusters 6
to 8. These clusters gave good results in spite of the
various numbers of items.
In some cases, we found cases of low precision,
such as cluster 4, and cases of low recall, such as
cluster 5. We can say with fair certainty that one of
the reasons was a lack of data. We could only gather
data from 1,000 people, which is less than 1% of SBM
users.
We found another reason by looking at the data.
The name of the tag in cluster 4 was ‘javascript’;
however,most of the recommender item clusters were
‘programming’. The scope of the recommender item
clusters seemed to be broader than that of cluster 4.
However, lack of data makes the similarities between
cluster 4 and these recommender item clusters rela-
tively high. The precision decreased because cluster 4
frequently recommended items out of the focus. The
name of the tag in cluster 5 was ‘art’, and most of
the recommender item cluster was ‘webdesign’. The
scope of the recommender item cluster seemed to be
narrower than that of cluster 5. Therefore, cluster
5 was recommended in only a part of items recom-
mended and recall was decreased.
These problems were caused by the lack of data,
but there is further room for investigation. For ex-
ample, we will combine item clusters to create more
suitable recommender item clusters. However, a more
comprehensive study on creating data lies outside the
scope of this paper.
4.2 Comparison 1: Recommendation
based on Folksonomy
In this section, we compare the recommendation sys-
tems based on folksonomy with our proposed system
from the viewpoint of recall and precision. We show
the difference between the two methods in Table 3
(Comparison 1 vs Proposed).
4.2.1 Evaluation
The comparative recommendation system based on
folksonomy is as follows.
1. User inputs a tag name into the system as a query.
2. System recommends items with such a tag in de-
scending order.
4.2.2 Experimental Results
We show an example of the comparative results in
Fig. 8. This result corresponds to the item cluster
which t
query
is ajax, B is 2118 and T is 84. Our method
clearly outperformed the system based on folkson-
omy.
ANTI-FOLKSONOMICAL ITEM RECOMMENDATION SYSTEM BASED ON SIMILARITIES BETWEEN ITEM
CLUSTERS IN SOCIAL BOOKMARKING
615
Table 3: Comparative experiments
Comparison 1 Comparison 2 Proposed
Query tag user&tag user&tag
Co-occurrence not used category category
Similarity number of users co-occurrence ratio likelihood ratio
Precision
Recall
Proposed
Sum tag
Jaccard
Figure 8: Proposed system vs. system using tag names vs.
system using Jaccard coefficient.
4.2.3 Discussion
One can safely state that the recommendation system
based on item clusters can produce better results than
the recommendation system based on folksonomy.
One might also think that folksonomy would lead to
better results than our system when the query word
is used commonly. However, these results show that
this may not be true. Note that ‘ajax’, which refers to
javascript programming techniques, is a well-known
word among web programmers.
Let us look closely at the results to find why our
system is more appropriate than the comparative one.
In a query item cluster, the items tagged with ‘ajax’
show us high quality interfaces or programming tech-
niques. On the other hand, in a recommender item
cluster, the tagged items show us only an implemen-
tation of ajax. That is, these item clusters are based on
different opinions even though the tags are the same.
Moreover, recommendation system based on folkson-
omy cannot recommend items to users who use sin-
gular tag names such as ‘java/app’ or ‘***java***’.
Our proposed system, however, is not limited by the
tag name.
4.3 Comparison 2: Similarity by
Jaccard Coefficient
In this section, we compare the similarity by Jaccard
coefficient with that based on hypothesis testing. We
Figure 9: Numerical comparison between similarities based
on hypothesis testing and Jaccard coefficient.
show the difference between the two methods in Table
3 (Comparison 2 vs Proposed).
4.3.1 Evaluation
There are conventional systems for comparing the
similarity of sample sets, such as the Jaccard and co-
sine coefficients. The Jaccard coefficient is defined
as the two sample sets’ intersection divided by their
union. We can define the Jaccard coefficient for our
situation as follows.
sim
Jaccard
(s, o) =
k
n
, (10)
where n and k are the values in Eqs. (1) and (2).
Therefore, we assume a comparative system re-
placing the Jaccard coefficient with similarities based
on hypothesis testing. It can be said that the system
is a conventional simple collaborative filtering system
based on tags.
4.3.2 Results
We show an example of the comparative results in
Fig. 8. This result corresponds to the item cluster
previously described in 4.2.2. Our system is clearly
more appropriate than the system based on the Jac-
card coefficient.
4.3.3 Discussion
We explain why our similarity is better than the con-
ventional similarity based on the Jaccard coefficient.
WEBIST 2009 - 5th International Conference on Web Information Systems and Technologies
616
A comparison between hypothesis testing and the
Jaccard coefficient is shown in Fig. 9. Lines (1)-(a)
and (b) show the same similarities by the Jaccard co-
efficient. On the other hand, lines (2)-(a) and (b) show
the same similarities by hypothesis testing. Line (1)
and (2)-(a) shows the value of 0.6 and (1) and (2)-(b)
shows the value of 0.4.
Then, it is open to question to equate the case of
n = 4, k = 3 with the case of n = 20, k = 15. The
former case would arise more often than the latter.
Therefore, in the Jaccard coefficient, a small value
of n leads to worse results. In other words, we have
to avoid any accidental co-occurrence for a high-
precision and high-recall recommendation system.
Hypothesis testing shows a small value when n or
k is small and a large value when both n and k are
large. Then, Eq. 3 can calculate similarity except in
accidental co-occurrences.
5 CONCLUSIONS
We proposed a novel recommendation system using
SBM data. Several conventional systems using folk-
sonomy have focused on actual tag names. However,
we focused on item clusters, which are sets of items
tagged by each SBM user. We assumed SBM users’
behavior follows binomial distribution and used hy-
pothesis testing to calculate the similarities between
two item clusters. In addition, we evaluated our rec-
ommendation system. The results showed high re-
call and precision. We compared our proposed sys-
tem with the systems using actual tag names and
showed that our proposed system was more appropri-
ate. We also compared our proposed similarity calcu-
lation based on hypothesis testing with a conventional
similarity calculation and verified that our resultant
similarities were better than the conventional ones.
REFERENCES
Goldberg, D., Nichols, D., Oki, B. M., and Terry, D. (2003).
Using collaborative filtering to weave an information
tapestry. Communications of the ACM, 35(12):61–70.
Golder, S. and Huberman, B. A. (2005). The structure of
collaborative tagging systems. Journal of Information
Science.
Gunduz, S. and Ozsu, M. T. (2003). A user interest model
for web page navigation. Proceedings of International
Workshop on Data Mining for Actionable Knowledge.
Hotho, A., Jaschke, R., Schmitz, C., and Stumme, G.
(2006). Information retrieval in folksonomies: Search
and ranking. 3rd European Semantic Web Confer-
ence(ESWC2006), pages 411–426.
Ishikawa, H., Nakajima, T., Mizuhara, T., Yokoyama, S.,
Nakayama, J., Ohta, M., and Katayama, K. (2002).
An intelligent web recommendation system: A web
usage mining approach. International Symposium
on Methodologies for Intelligent Systems, pages 342–
350.
Jaschke, R., Marinho, L. B., Hotho, A., Schmidt-Thieme,
L., and Stumme, G. (2007). Tag recommendations
in folksonomies. Knowledge Discovery in Databases
(PKDD 2007), pages 506–514.
Kazienko, P. and Kiewra, M. (2004). Integration of rela-
tional databases and web site content for product and
page recommendation. International Database Engi-
neering and Applications Symposium (IDEAS’04).
Li, J. and Zaiane, O. R. (2004). Combining usage, con-
tent, and structure data to improve web site recom-
mendation. Proceedings of Web KDD-2004 workshop
on Web Mining and Web Usage.
Mika, P. (2005). Ontologies are us:a unified model of social
networks and semantics. 4th International Semantic
Web Conference(ISWC2005), pages 122–136.
Niwa, S., Doi, T., and Honiden, S. (2006). Web page rec-
ommender system based on folksonomy mining. Pro-
ceedings of the Third International Conference on In-
formation Technology: New Generations (ITNG’06).
Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and
Riedl, J. (1994). Grouplens: An open architecture for
collaborative filtering of netnews. Proceedings of the
1994 Computer Supported Cooperative Work Confer-
ence, pages 175–186.
Rucker, J. and Polanco, M. J. (1997). Siteseer: Personalized
navigation for the web. Communications of the ACM,
40(3):73–75.
Sarwar, B. M., Karypis, G., Konstan, J. A., and Riedl, J.
(2001). Item-based collaborative filtering recommen-
dation algorithms. Proceedings of the 10th Interna-
tional World Wide Web Conference (WWW10), pages
285–295.
Yanbe, Y., Jatowt, A., Nakamura, S., and Tanaka, K. (2007).
Can social bookmarking enhance search in the web?
Proceedings of the 7th ACM/IEEE-CS joint confer-
ence on Digital libraries.
ANTI-FOLKSONOMICAL ITEM RECOMMENDATION SYSTEM BASED ON SIMILARITIES BETWEEN ITEM
CLUSTERS IN SOCIAL BOOKMARKING
617