EXPLORING MULTI-FACTOR TAGGING ACTIVITY
FOR PERSONALIZED SEARCH
Frederico Durao, Ricardo Lage, Peter Dolog
IWIS — Intelligent Web and Information Systems, Aalborg University, Computer Science Department
Selma Lagerl¨ofs Vej 300, DK-9220 Aalborg-East, Denmark
Nilay Cos¸kun
Istanbul Technical University, Science and Technology Institution Building, 34469 Maslak Istanbul, Turkey
Keywords:
Personalized, Search, Tagging, User learning, User preferences.
Abstract:
Coping with ambiguous queries has long been an important part in the research of Web Information Systems
and Retrieval, but still remains to be a challenging task. Personalized search has recently got significant
attention to address this challenge in the web search community, based on the premise that a user’s general
preference may help the search engine disambiguate the true intention of a query. However, studies have shown
that users are reluctant to provide any explicit input on their personal preference. In this paper, we study how a
search engine can learn a user’s preference automatically based on a user’s tagging activity and how it can use
the user preference to personalize search results. Our experiments show that users’ preferences can be learned
from a multi-factor tagging data and personalized search based on user preference yields significant precision
improvements over the existing ranking mechanisms in the literature.
1 INTRODUCTION
Users often have different intentions when using
search engines (Xu et al., 2008). According to (Qiu
and Cho, 2006), queries provided by users to search
engines are under-specified and, thus, poorly translate
fully the meaning users have in mind. As a conse-
quence, according to (J
¨
aschke et al., 2007), only 20%
to 45% of the results match a user’s intentions. One
approach to increase user satisfaction is to personalize
search results.
Recently, strategies intended to disambiguate the
true intention of a query began to collect and analyze
user preferences in order to personalize the search re-
trievals. The user preferences have been modeled in
many ways in the literature (Biancalana, 2009; Xu
et al., 2008; Qiu and Cho, 2006), including analy-
sis of explicit data such as user profile and preference
models, or the implicit collection of data such as user
click history, visited pages log, or tagging activity. All
of these are indicators of user preferences utilized by
search engines to decide which items in the collection
of search results are more or less relevant for a partic-
ular individual.
In this work, we analyze user’s tagging activ-
ity to learn the user preferences and personalize the
search results. For this purpose, we consider tags
because they represent some sort of affinity between
user and resource. By tagging, users label resources
freely and subjectively, based on their sense of value.
Tags, therefore, become a potential source for learn-
ing user’s interests (Durao and Dolog, 2009).
In our approach, we build our user model incor-
porating various tag indicators of user’s preference,
i.e., each indicator relates to a factor for personalizing
searches. We therefore formalize the term factor as
an indicator of user’s preference denoted by a partic-
ular set o tags. For instance, a factor Z may represent
the set of tags assigned to the most visited pages of a
user and a factor Y may represent the set of tags as-
signed to the pages marked as favorites by the same
user. Our belief is that a multi-factor approach can
produce a more accurate user model, and thereby fa-
cilitate the search for what is more suitable for a given
user. The contributions of this paper are:
• We provide a personalized component to investi-
gate the problem of learning a user’s preference
based on his tagging activity. We propose a sim-
212
Durao F., Lage R., Dolog P. and Co¸skun N..
EXPLORING MULTI-FACTOR TAGGING ACTIVITY FOR PERSONALIZED SEARCH.
DOI: 10.5220/0003300502120220
In Proceedings of the 7th International Conference on Web Information Systems and Technologies (WEBIST-2011), pages 212-220
ISBN: 978-989-8425-51-5
Copyright
c
 2011 SCITEPRESS (Science and Technology Publications, Lda.)

ple yet flexible model to succinctly represent user
preferences based on multiple factors.
• Based on the formal user model, we develop a
method to automatically estimate the user’s im-
plicit preferences based on his and the commu-
nity’s tagging activity. We provide theoretical and
experimental justification for our method.
• Finally, we conduct an experimental and compar-
ative analysis to evaluate how much the search
quality improves through this personalization.
Our experiments indicate reasonable improve-
ment in search quality — we observed about
61.6% of precision improvement over traditional
text-based information retrieval and 6,13% of
precision gain over the best method compared —
demonstrating the potential of our approach in
personalizing search.
The rest of this paper is organized as follows. We
first provide an overview of personalized search and
social tagging activity in section 2, on which our work
is mainly based. In section 3 we describe our mod-
els and methods. Experimental evaluation, results and
discussion are presented in section 4. In section 5, we
review the related work and finally conclude the paper
in section 6.
2 BACKGROUND
In this section we briefly revise traditional text-
based information retrieval concepts followed by an
overview of collaborative tagging systems.
2.1 Information Retrieval and
Personalized Search
Traditionally, search engines have been the primary
tool for information retrieval. Usually the search
score is computed in terms of term frequency and
inverse document frequency (i.e. tf-idf) (Baeza-
Yates and Ribeiro-Neto, 1999). The tf-idf weighting
scheme assigns to term t a weight in document d given
by:
tf-idf
t,d
= tf
t,d
× idf
t
(1)
In other words, tf-idf
t,d
assigns to term t a weight
in document d that is: i)highest when t occurs many
times within a small number of documents (thus lend-
ing high discriminating power to those documents);
ii) lower when the term occurs fewer times in a doc-
ument, or occurs in many documents (thus offering
a less pronounced relevance signal); iii)lowest when
the term occurs in virtually all documents.
Eventually the search score of a document d is the
sum, over all query terms q, of the tf-idf weights given
for each term in d.
Score(q, d) =
X
t∈q
tf-idf
t,d
(2)
In order to achieve personalization, a further step
consists of selecting which documents d are relevant
or not for a given user. Technically, the traditional
scoring function must be adapted/integrated with a
personalization model capable of modeling the users’
preferences. This is usually done by adding new terms
to the query and re-weighting the original query terms
based on the user profile.
2.2 Collaborative Social Tagging
Collaborative tagging systems have become increas-
ingly popular for sharing and organizing Web re-
sources, leading to a huge amount of user generated
metadata. Tags in social bookmarking systems such
as Del.ici.ous
1
are usually assigned to organize and
share resources on the Web so that users can be re-
minded later on. Invariably, tags represent some sort
of affinity between user and a resource on the Web. By
tagging, users label resources on the Internet freely
and subjectively, based on their sense of values. Tags
then become a potential source for learning user’s in-
terests.
2.2.1 Tagging Notation
Formally, tagging systems can be represented as hy-
pergraphs where the set of vertices is partitioned into
sets: U = {u
1
, ..., u
k
}, R = {r
1
, ..., r
m
}, and T = {t
1
, ..., t
n
},
where U, R, and T correspond to users, resources, and
tags. A tag annotation, i.e. a resource tagged by a
user, is an element of set A, where: A ⊆ U × R × T .
The final hypergraph formed by a tagging system
is defined as G with: G = hV, Ei with vertices V =
U ∪ R ∪ T , and edges E = {{u, r,t} | (u, r,t) ∈ A}.
Particularly to understand the interests of a single
user, our models concentrate on the tags and resources
that are associated with this particular user, i.e. in a
personalized part of the hypergraph G. We then define
the set of interests of a user as I
u
= (T
u
, R
u
, A
u
), where
A
u
is the set of tag annotations of the user: A
u
= {(t, r) |
(u, t, r) ∈ A}, T
u
is the user’s set of tags: T
u
= {t | (t, r) ∈
A
u
}, R
u
is the set of resources: R
u
= {r | (t, r) ∈ A
u
}.
The introduction on information retrieval as well
as on tagging notations will serve as basis for describ-
ing our personalized search model in the following
sections.
1
http://delicious.com
EXPLORING MULTI-FACTOR TAGGING ACTIVITY FOR PERSONALIZED SEARCH
213

3 MULTI-FACTOR TAG-BASED
PERSONALIZED SEARCH
We now discuss how we personalize search results.
In Section 3.1 we first describe our multi-factor rep-
resentation of user preferences. Then in Section 3.2,
we describe how to use this preference information in
ranking search results.
3.1 User Preference Representation
In this work, tags from different factors such as tags
assigned to the user’s bookmarks and his own tags
serve as our learn units of user preferences. Further,
we understand that some tags are preferred over an-
other, meaning that the frequency of usage of a given
tag can denote its affinity with the tagger. In this
sense, we define the tag preference set, for each factor,
as a tuple (tag, tagFreq(tag)), where tagFreq(tag) is
a function that measures the user’s degree of interest
in that tag. Formally, we define this set for a particular
user u and factor f ∈ F (let F be the set of all possible
factors) as:
T
f
= {(t, tagFreq(t)) | t ∈ T
f
}, (3)
where tagFreq(t) =
n
t
|T
f
|
, n
t
is the number of oc-
currences of the tag t ∈ T
f
and |T
f
| is the amount of
tags in a given factor f . The set T
f
is normalized
such that
P
|T
f
|
i=1
tagFreq(i) = 1. To illustrate the user
tagging preference representation, suppose a user has
only two tags in one particular factor: “semantic web”
and “data mining”, and the first has been utilized three
times while the second has been utilized only once.
This means the user has been interested in “semantic
web” three times as much as he has been interested
in “data mining”. Then, the tag preference set of the
user for that factor will be represented as {(“semantic
web”,0.75),(“data mining”,0.25)}.
The composition of our multi-factor tag-based
user model T
0
u
extends the traditional set of user tags
T
u
(see subsection 2.2.1), with a disjoint union of tag
sets:
T
0
u
=
G
f ∈F
T
f
, (4)
where T
f
is the set of tags assigned to each fac-
tor f ∈ F. Next section explains how the tag-based
multi-factor approach is applied to personalize search
results.
3.2 Tag-based Personalization
Approach
The tag-based personalization approach decides
which resource r ∈ R is relevant to each user u ∈ U
based on his preferences established in the multi-
factor tag-based user model. In the context of this
work, we re-rank the search results by measuring the
similarity of tags that denote user preference and the
tags assigned to the retrieved items. With this, we pro-
mote the items closer to user’s preferences at the first
positions in the collection of search results.
Technically, we calculate the cosine similarity
(Baeza-Yates and Ribeiro-Neto, 1999) between each
vector of tag frequencies
−→
T
f
⊂
−→
T
0
u
from T
f
⊂ T
0
u
and
the vector of tag frequencies
−→
T
r
from the retrieved re-
sources T
r
of a given user query q. Further, we weigh
each vector
−→
T
f
with a coefficient, α
f
that determines
its degree of importance over the other factors in the
model. We incorporate the coefficients because dif-
ferent users may rely on tag factors differently. As a
consequence, the importance of each factor may vary
accordingly.
The coefficient values for each factor are auto-
matically estimated using the Ordinary Least Square
(OLS) linear regression. The goal of using OLS is to
minimize the sum of squared distances between the
observed tag frequencies of each factor in the dataset
and the best coefficient values predicted by linear ap-
proximation (Williams, 2007).
Once the coefficient values are estimated, the Tag-
based Similarity Score (TSS) can be calculated as:
T S S (
−→
T
r
,
−→
T
0
u
) =
|F|
X
f =1
α
f
∗
−→
T
r
.
−→
T
f
|
−→
T
r
||
−→
T
f
|
, (5)
The T S S value is then utilized to weigh the or-
dinary search score and thereby promote the items
matching the user interests to higher positions in the
search ranking. The Personalized Search Score (PSS)
over a resource r triggered by a query q in the set of
resources R is then defined as follows:
PSS(q, r, u) =
X
t∈q
tf-idf
t,r
∗ T S S (
−→
T
r
,
−→
T
0
u
). (6)
In the following section, with the intuit of demon-
strate the how the personalized search takes place, we
simulate a user search since the user’s initial query
until the re-ranking of search results.
WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies
214

3.3 Personalized Search in Action
In this section we describe how the personalized rank-
ing is realized when a user u ∈ U submits a query q for
a search engine:
1. Assume that a user u, whose preferences are de-
noted by the multi-factor tag-based model T
0
u
,
submits a query q for a given search engine;
2. The search engine returns a set of resources S ⊆
R that matches the entry query q. Each re-
trieved resource s ∈ S is assigned with a set of
tags defined by T
s
= {t1, .., t
z
}. The retrieved
items in S are initially ranked respecting the tf-idf
(Baeza-Yates and Ribeiro-Neto, 1999) ordering
τ = [s
1
, s
2
, ..., s
k
], where the ordering relation is
defined by s
i
≥ s
j
⇔ tf-idf(q, s
i
) ≥ tf-idf(q, s
j
).
3. For each s ∈ S , we weigh the τ with T S S (
−→
T
s
,
−→
T
0
u
).
The outcome is a personalized ranking list of
pages represented by S
0
. The rank list will follow
a new ordering τ
0
= [s
1
, s
2
, ..., s
k
], where s
i
∈ S
0
and the ordering relation is defined by s
i
≥ s
j
⇔
PS S (q, s
i
, u) ≥ persS core(q, s
j
, u).
4. The personalized result set S
0
⊆ R is then returned
to the user.
4 EXPERIMENTAL EVALUATION
To evaluate our approach, we analyze and compare
how the different similarity measures re-rank the re-
turned result list. For the matter of comparison, the
best personalization approaches will rank the rele-
vant/preferable retrieved items to the higher positions
in the result list.
We utilize the precision metric to measure and
compare the performance of our personalization ap-
proach over other similarity measures. By doing
so, the research goal of this evaluation is to assess
whether the multi-factor tags improve the precision
of traditional text-based information retrieval mech-
anism in leading users to find the information they
are actually looking for. We intend to address (and
attenuate) the problem witnessed by searchers when
the search results do not meet their real needs and,
instead, present them with undesired information.
4.1 MovieLens Dataset
We have utilized the MovieLens dataset for evaluating
our approach. The data set contains 1,147,048 ratings
and 95,580 tags applied to 10,681 movies by 13,190
users. The dataset was chosen because it allowed us
to build multi-factor user model base on two distinct
factors corresponding to two different user activities:
tagging and rating. The first factor refers to T
u
as the
set of user’ tags and the second refers to V
u
, the set
of tags assigned by other users to the pages rated four
or five stars by user u. In a scale from one to five, we
consider movies rated four or five as strong indicators
of user’s preferences. The evaluated multi-factor user
model is defined as follows:
T
0
u
= γ.T
u
+ β.V
u
(7)
where γ and β are the coefficients used to weigh
the importance of each tag factor.
Because the MovieLens dataset only provides lit-
tle information about the movies such as title and cat-
egory, we crawled the synopses of the movies in order
to create our search space. The synopses were ex-
tracted from The Internet Movie Data Base (IMDb)
2
,
a movie repository that catalogs information about
movies, including information about casting, loca-
tions, dates, synopses, reviews, and fan sites on the
Web.
4.2 Methodology and Evaluation
Metrics
In order to evaluate whether the proposed approach
matches our research goal, we performed a quantita-
tive evaluation in terms of precision of the search re-
sults. Precision refers to the proportion of retrieved
items that are actually relevant to the user’s query
(Baeza-Yates and Ribeiro-Neto, 1999). We calculated
the precision as prec(q, u) =
|R
q,u
|∩|R
0
q,u
|
|R
q,u
|
, where |R
q,u
| is
the amount of retrieved items for a query q triggered
by an user u and |R
0
q,u
| is the amount of relevant pages
expected to be retrieved. This set was composed by
movies rated four or five stars besides the query q as
a term in the text content. In this way, we could dis-
tinguish what was relevant or not for each user. Fur-
ther, we statistically computed the most representative
number of users to be assessed. The sample size was
calculated with confidence level set to 95% and con-
fidence interval set to 2%.
4.2.1 Queries and Search Space
For each user, we ran the top n queries using the most
popular terms that appeared in all indexed pages. We
prioritized the most frequent terms along all docu-
ments and the most frequent ones in each document.
It is important to mention that we filtered out the stop
words prior to processing the list of indexed terms.
2
http://www.imdb.com/
EXPLORING MULTI-FACTOR TAGGING ACTIVITY FOR PERSONALIZED SEARCH
215

Table 1: Size of different datasets we utilized in this paper. Count is the number of movies the dataset contains. N. users is
the number of users that generated those entities. For example, the second and third columns of the second row indicate that
4,009 users assigned 95,580 tags to movies. Total is the number of times the users rated or tagged the movies. The last three
columns indicate the same numbers after the pruning is being applied.
Before Pruning After Pruning
Dataset Count N. Users Total Count N. Users Total
Movies rated 8,358 9,753 1,147,048 7,075 6,476 211,746
Movies tagged 7,601 4,009 95,580 4,029 3,871 88,181
This list was calculated as a mean of tf-df (Baeza-
Yates and Ribeiro-Neto, 1999). This decision was
taken to avoid that any particular group of individu-
als was favored over another.
Once we have selected the queries, we focused on
defining the most appropriate search space. Since our
goal was centered on the precision, we decided to fo-
cus the search only at the top 30 ranked items. This
focused observation was motivated by the fact that
users usually don’t go further to encounter what they
are looking for, instead they reformulate queries to
better convey the information they are actually seek-
ing (Baeza-Yates and Ribeiro-Neto, 1999).
4.2.2 Pruning
Although substantial data was available, we pruned
the dataset for our analyses of the personalized search.
We focused on a set of movies with a minimum
threshold of tags and ratings. We began pruning
the present dataset by selecting items that had been
tagged with at least 5 distinct tags. We wanted to fo-
cus on tag sets with high lexical variability and less
redundancy. This care was taken to prevent that tags
from distinct factors could overlap with tags from ex-
pected documents in the case of low tag variability.
We iteratively repeated this pruning until reach a sta-
ble set of items and tags. Further, since we wanted to
explore the precision of the personalized search, we
only included users that had distributed ratings higher
than or equal to four stars. Otherwise we would not
have any source of information on user’s preference
about the retrieved items. The complete overview of
the dataset size is shown in Table 1.
4.2.3 Similarity Measure Comparison
The core of the tag-based personalization approach
is the computation of the Tag-based Similarity Score
(TSS) (see in section 3.2). In order to compare our
approach with other similarity measures, we changed
the TSS algorithm using other similarity algorithms
without considering the coefficients assigned to the
factors involved. Besides the cosine similarity, the
other similarity measures utilized are the Matching
Coefficient, Dice, Jaccard and Euclidean Distance
(all refer to (Boninsegna and Rossi, 1994)).
• The Matching Coefficient approach is a simple
vector based approach which simply counts the
number of terms (tags in our case), (dimensions),
on which both vectors are non zero. So, for vector
set
−→
v and set
−→
w the matching coefficient is calcu-
lated as matching(
−→
v ,
−→
w) = |
−→
v ∩
−→
w|. This can be
seen as the vector based count of similar tags.
• The Dice Coefficient is a term based similarity
measure (0-1) whereby the similarity measure is
defined as twice the number of terms common
to compared entities divided by the total number
of tags assigned in both tested resources. The
Coefficient result of 1 indicates identical vectors
(e.g.
−→
v and
−→
w) as where a 0 equals orthogo-
nal vectors. The coefficient can be calculated as
dice(
−→
v ,
−→
w) =
2|
−→
v ∩
−→
w|
|
−→
v |+|
−→
w|
.
• The Jaccard Coefficient measures similarity be-
tween sample sets, and is defined as the size of
the intersection divided by the size of the union of
the sample sets. The coefficient can be calculated
as: jaccard(
−→
v ,
−→
w) =
|
−→
v ∩
−→
w|
|
−→
v ∪
−→
w|
.
• The Euclidean Distance approach works in vector
space similar to the matching coefficient and the
dice coefficient, however the similarity measure
is not judged from the angle as in cosine rule but
rather the direct euclidean distance between the
vector inputs. The standard Euclidean distance
formula between vectors
−→
v and
−→
w is defined as
follows: euclidean(
−→
v ,
−→
w) =
q
P
n
i=1
(
−→
v
i
−
−→
w
i
)
2
.
• The Cosine Similarity is utilized to measure the
similarity between two vectors of n dimensions by
finding the cosine of the angle between them. The
cosine similarity for two vector between vectors
is calculated as: cosS im(
−→
v ,
−→
w) =
−→
v ·
−→
w
|
−→
v ||
−→
w|
.
Unlike our proposed TSS algorithm, here we do
not consider the coefficients assigned to the tag fac-
tors. For the sake of differentiation, we call the cosine
similarity with coefficient applied to our TSS algo-
rithm as Cosine
Coe f
.
WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies
216

Table 2: Mean of Precision.
Cosine
Coe f
Cosine Jaccard Dice Euclidean Matching
Min. 0.723 0.616 0.576 0.575 0.600 0.510
1st Qu. 0.765 0.710 0.576 0.577 0.600 0.583
Median 0.776 0.761 0.634 0.633 0.634 0.583
Mean 0.796 0.755 0.694 0.694 0.667 0.641
3rd Qu. 0.841 0.845 0.841 0.842 0.756 0.766
Max. 0.912 0.909 0.912 0.913 0.886 0.865
Std.Dev 0.052 0.087 0.131 0.130 0.083 0.105
4.3 Evaluation Results
Figure 1 shows the mean of precision (Subfigure 1(a))
and the box plot displays the distribution differences
between the five similarity measures (Subfigure 1(b)).
The higher precision values correspond to better per-
formance. The precision of our approach is displayed
in the first bar (Cosine
Coe f
) followed by the com-
pared similarity measures with their respective stan-
dard errors. We also include the non-personalized
precision result for the sake of comparison. This
ranking was based on the traditional tf-idf weighting
scheme where the text-based search has no support
from personal tags. It is important to emphasize that
all compared similarity measures produce a personal-
ized score except the non-personalized search.
Table 2 summarizes the statistics from our sam-
ple of precision values regarding the evaluated met-
rics. It shows the five-number summary (sample mini-
mum, lower quartile, median, upper quartile and sam-
ple maximum), the mean value, and the standard devi-
ation of the samples. It is worth noting that the sample
minimum value of our approach has higher precision
than any other similarity measure compared. This be-
havior is also noticed when compared with the sam-
ple maximum value. Similarly, the standard deviation
of the sample of our approach is lower than the stan-
dard deviation of the other samples. This indicates
less dispersion within the precision values of our ap-
proach. Furthermore, differences between all pairs of
similarity measures are significant (p ≤ 0.01) except
for those between Jaccard and Dice similarity.
As results show, our approach (Cosine
Coe f
)
achieved the highest precision rates. In particular, it
achieved 61.6% of precision improvement over tradi-
tional text-based information retrieval (non-pers) and
6,13% of precision gain over cosine similarity. The
overall result also indicates that the cosine-based sim-
ilarity measures (regarding or not the coefficients)
perform better than the other approaches. As ex-
pected, all similarity measures applied to the per-
sonalized search outperformed the non-personalized
method.
4.4 Discussion and Limitations
Concerning the role of tags as means of personal-
izing traditional information retrieval, we evidenced
that tags indeed can be used as learning units of
user’s preference. As results showed, this hypoth-
esis was confirmed when the precision of our solu-
tion considerably outperformed the precision of non-
personalized search. The immediate benefit of per-
sonalized search is the reduction of undesired re-
trievals at the first positions and thereby not distract-
ing users with unsolicited information. Part of this
success relates to the best estimation of coefficients
applied to each factor considered in the search. The
adoption of coefficients helped us to efficiently deter-
mine which tag factors are more representative for an
individual over the others. Regarding the applicabil-
ity of the approach, we are quite positive that existing
tag-based systems can utilize and/or adapt our solu-
tion to their reality.
The main limitation of the current approach is the
lack of tags. In this work we are not addressing this
problem since the goal is to emphasize the benefits
with the multi-factor approach. A potential solution is
seen in the work of (J
¨
aschke et al., 2007) that studies
tag prediction in social bookmarking environments.
Performance is another issue that was not formally
evaluated in the current study but we empirically ob-
served that the personalization process is at least 10%
more expensive than the basic keyword-based search.
However this cost cannot be judged isolated since
the personalized outcome can compensate the invest-
ment. Further investigation is necessary and planned
for future works.
5 RELATED WORK
In recent years many researchers utilize query log
and click-through analysis for web search personal-
ization. In (Qiu and Cho, 2006), the authors combine
a topic-sensitive version of PageRank (Haveliwala,
2002) with the history of user clicks data for search
EXPLORING MULTI-FACTOR TAGGING ACTIVITY FOR PERSONALIZED SEARCH
217

(a) Mean of Precision (b) Box Plot
Figure 1: Overall result and statistics from the experiment. Note that each of the similarity measure are utilized in the TSS
algorithm to personalize the search.
result personalization. Joachims et al. (Joachims
et al., 2005) study clicks applicability as implicit rel-
evance judgments. They show that users’ clicks pro-
vide a reasonably accurate evidence of their prefer-
ences. These studies are relevant to our work in con-
text of predicting user preferences but our approach
takes into account user annotations which are called
tags as user feedback. Tan et al. (Tan et al., 2006) pro-
pose a language modeling approach for query history
mining. Their small-scale study demonstrates signif-
icant improvement of personalized web search with
a history-based language model over regular search.
The user modeling approach described in (Shen et al.,
2005) is based on a decision-theoretic framework to
convert implicit feedback into a user profile that is
used to re-rank search results.
While user models are usually targeted at search
personalization, they could also be applied for person-
alized information filtering, as was shown in (Yang
and Jeh, 2006) which analyzes click history for the
identification of regular users’ interests. (Teevan
et al., 2009) shows that combining implicit user pro-
files from several related users has a positive impact
on personalization effectiveness. Recently, new ap-
proaches for adaptive personalization focus on the
user task and the current activity context. There are
several approaches trying to predict applicability of
personalization while considering the current context
of the user’s task on query submission (Teevan et al.,
2008; Dou et al., 2007).
Collaborative Filtering (CF) (Bender et al., 2008)
has become a very common technique for provid-
ing personalized recommendations, suggesting items
based on the similarity between users’ preferences.
One drawback of traditional CF systems is the need
for explicit user feedback, usually in the form of rat-
ing a set of items, which might increase users’ entry
cost to the recommender system. Hence, leveraging
implicit user feedback (Au Yeung et al., 2008), such
as views, clicks, or queries, has become more popu-
lar in recent CF systems. In this work, we leverage
implicit tagging information, which can be viewed as
a variant of implicit user feedback. (Carmel et al.,
2009) proposes a solution which considers the struc-
ture of the user’s social network assuming that people
which share same social ties have similar interests.
In addition to social network, they have another ap-
proach similar to our study taking into account user
tags to understand user interest. Comparing these
two approaches they noted that tagging activity gives
efficient information about user preferences for ac-
tive taggers. In order to measure relevance between
user and search result, (Gemmell et al., 2008) con-
siders topic matching instead of tags which we used
in our approach. They propose a ranking algorithm
which ranks web pages by the term matching between
user interest and resource’s topic. In (Sieg et al.,
2007), there exists a different approach to personal-
izing search result by building models of user as on-
tological profiles. They derive implicit interest scores
to existing concepts in domain ontology.
Focused mostly on tags, (Xu et al., 2008) pro-
poses an algorithm for personalizing search and nav-
igation based on personalized tag clusters. Slightly
similar to our model, they measure the relevance be-
tween user and tag cluster and try to understand user
interest while we calculate the similarity between
user and tags. Tags are used in different manners to
find search personalized results answering user needs
since tags are chosen by users personally. Similar to
our study, in (Noll and Meinel, 2007), a tag based re-
ranking model is presented taking into account tags
from Del.icio.us. Likewise our model, they compare
tags of search results and users to calculate new scores
of search results.
6 CONCLUSIONS
AND FUTURE WORKS
This study introduces a multi-factor tag-based model
WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies
218

to personalize search results. We analyze user’s tag-
ging activity to learn users’ preferences and use this
information to personalize the search. We evaluated
our approach with other personalization methods and
as a result we realized significant improvement of pre-
cision. As a future work, we intend to analyze the se-
mantic relationship between tags in order to catch hid-
den similarities that are not undertaken by this model.
In addition, we aim at enhancing the model by consid-
ering the tag decay. The goal is to perform a temporal
analyzes and filter the results according to the actual
users’ preferences.
ACKNOWLEDGEMENTS
The research leading to these results is part of the
project “KiWi - Knowledge in a Wiki” and has re-
ceived funding from the European Community’s Sev-
enth Framework Programme (FP7/2007-2013) un-
der grant agreement No. 211932. This work has
been supported by FP7 ICT project M-Eco: Med-
ical Ecosystem Personalized Event-Based Surveil-
lance under grant number 247829. The authors
would like to acknowledge GroupLens research at
University of Minnesota for providing the MovieLens
dataset.
REFERENCES
Au Yeung, C. M., Gibbins, N., and Shadbolt, N. (2008).
A study of user profile generation from folksonomies.
In Proceedings of the Workshop on Social Web
and Knowledge Management (SWKM2008),Beijing,
China, 21-25 April, 2007, pages 1–8.
Baeza-Yates, R. and Ribeiro-Neto, B. (1999). Modern In-
formation Retrieval. Addison Wesley.
Bender, M., Crecelius, T., Kacimi, M., Michel, S., Neu-
mann, T., Parreira, J. X., Schenkel, R., and Weikum,
G. (2008). Exploiting social relations for query expan-
sion and result ranking. In ICDE Workshops, pages
501–506. IEEE Computer Society.
Biancalana, C. (2009). Social tagging for personalized web
search. In AI*IA 2009: Emergent Perspectives in Ar-
tificial Intelligence, pages 232–242.
Boninsegna, M. and Rossi, M. (1994). Similarity mea-
sures in computer vision. Pattern Recognition Letters,
15(12):1255 – 1260.
Carmel, D., Zwerdling, N., Guy, I., Ofek-Koifman, S.,
Har’el, N., Ronen, I., Uziel, E., Yogev, S., and Cher-
nov, S. (2009). Personalized social search based on the
user’s social network. In CIKM ’09: Proceeding of the
18th ACM conference on Information and knowledge
management, pages 1227–1236, New York, NY, USA.
ACM.
Dou, Z., Song, R., and Wen, J.-R. (2007). A large-scale
evaluation and analysis of personalized search strate-
gies. In WWW ’07: Proceedings of the 16th interna-
tional conference on World Wide Web, pages 581–590,
New York, NY, USA. ACM.
Durao, F. and Dolog, P. (2009). A personalized Tag-Based
recommendation in social web systems. In Proceed-
ings of International Workshop on Adaptation and
Personalization for Web 2.0 (AP-WEB 2.0 2009) at
UMAP2009, volume 485.
Gemmell, J., Shepitsen, A., Mobasher, M., and Burke, R.
(2008). Personalization in folksonomies based on tag
clustering. In Proceedings of the 6th Workshop on In-
telligent Techniques for Web Personalization and Rec-
ommender Systems.
Haveliwala, T. H. (2002). Topic-sensitive pagerank. In
WWW ’02: Proceedings of the 11th international con-
ference on World Wide Web, pages 517–526, New
York, NY, USA. ACM.
J
¨
aschke, R., Marinho, L., Hotho, A., Schmidt-Thieme, L.,
and Stumme, G. (2007). Tag recommendations in
folksonomies. pages 506–514.
Joachims, T., Granka, L., Pan, B., Hembrooke, H., and Gay,
G. (2005). Accurately interpreting clickthrough data
as implicit feedback. In SIGIR ’05: Proceedings of the
28th annual international ACM SIGIR conference on
Research and development in information retrieval,
pages 154–161, New York, NY, USA. ACM.
Noll, M. G. and Meinel, C. (2007). Web search person-
alization via social bookmarking and tagging. In
ISWC’07/ASWC’07: Proceedings of the 6th interna-
tional The semantic web and 2nd Asian conference
on Asian semantic web conference, pages 367–380,
Berlin, Heidelberg. Springer-Verlag.
Qiu, F. and Cho, J. (2006). Automatic identification of
user interest for personalized search. In WWW ’06:
Proceedings of the 15th international conference on
World Wide Web, pages 727–736, New York, NY,
USA. ACM.
Shen, X., Tan, B., and Zhai, C. (2005). Implicit user mod-
eling for personalized search. In CIKM ’05: Pro-
ceedings of the 14th ACM international conference on
Information and knowledge management, pages 824–
831, New York, NY, USA. ACM.
Sieg, A., Mobasher, B., and Burke, R. (2007). Web search
personalization with ontological user profiles. In
CIKM ’07: Proceedings of the sixteenth ACM con-
ference on Conference on information and knowledge
management, pages 525–534, New York, NY, USA.
ACM.
Tan, B., Shen, X., and Zhai, C. (2006). Mining long-term
search history to improve search accuracy. In KDD
’06: Proceedings of the 12th ACM SIGKDD interna-
tional conference on Knowledge discovery and data
mining, pages 718–723, New York, NY, USA. ACM.
Teevan, J., Dumais, S. T., and Liebling, D. J. (2008). To
personalize or not to personalize: modeling queries
with variation in user intent. In SIGIR ’08: Proceed-
ings of the 31st annual international ACM SIGIR con-
EXPLORING MULTI-FACTOR TAGGING ACTIVITY FOR PERSONALIZED SEARCH
219

ference on Research and development in information
retrieval, pages 163–170, New York, NY, USA. ACM.
Teevan, J., Morris, M. R., and Bush, S. (2009). Discovering
and using groups to improve personalized search. In
WSDM ’09: Proceedings of the Second ACM Interna-
tional Conference on Web Search and Data Mining,
pages 15–24, New York, NY, USA. ACM.
Williams, G. (2007). Linear Algebra With Applications.
Jones and Bartlett Publishers, Inc., USA.
Xu, S., Bao, S., Fei, B., Su, Z., and Yu, Y. (2008). Explor-
ing folksonomy for personalized search. In SIGIR ’08:
Proceedings of the 31st annual international ACM SI-
GIR conference on Research and development in in-
formation retrieval, pages 155–162, New York, NY,
USA. ACM.
Yang, B. and Jeh, G. (2006). Retroactive answering of
search queries. In WWW ’06: Proceedings of the 15th
international conference on World Wide Web, pages
457–466, New York, NY, USA. ACM.
WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies
220