Personalized Hotlink Assignment using Social Networks

Christos Makris, Konstantinos Siaterlis and Pantelis Vikatos

Computer Engineering & Informatics Department, University of Patras, Patras, Greece

Keywords:

Hotlink Assignment, Personalization, Social Networks.

Abstract:

In this paper, we introduce a novel methodology for personalized website reconstruction. We combine context

and popularity of the web pages and the information of user’s interest from social media. We present an

efﬁcient automatic web restructure placing suitable hotlinks between nodes of the generated website’s graph

using information of social media contrary to previous studies. In addition, our methodology includes an

innovative personalization scheme using a topic modeling approach to texts of users of social media to create

a graph of categories. We evaluate our approach counting user’s feedback about the ordering and relevance of

links to a website.

1 INTRODUCTION

It is difﬁcult to cover the needs of all users when the

number of pages and categories in a dynamic website

increase. Many websites contain hundreds or thou-

sands of different categories, creating difﬁculties to

ﬁnd the page that the user wants degrading the quality

of the website. Also, the amount of information that

is provided by various agents grows rapidly according

to informational needs. The reconstruction of a web-

site based on the ﬂuctuation of web page’s popularity

constitutes a signiﬁcant factor to enhance websites.

The goal of reconstruction is to reach popular pages

in fewer steps starting from the homepage, improving

the browsing experience. A well-cited methodology

is the use of additional links i.e. hotlinks, that con-

nect popular web pages with its descendants reducing

the distance from the home page. A ﬁrst approach has

been introduced in (Perkowitz and Etzioni, 2000) and

presents the idea of a modiﬁcation of the link structure

of the website, minimizing the steps from homepage

to popular pages using hotlinks. However, the unilat-

eral use of popularity as a factor of browsing enhance-

ment might not be efﬁcient due to the fact that differ-

ent users have different needs based on their prefer-

ences. Therefore the introduction of personalization

in the website reconstruction is necessary in order to

provide more targeted information. The extraction of

user’s interest and needs occurs in an explicit or im-

plicit manner. The main drawback of an explicit col-

lection of user feedback that it is not supported by all

users. In many studies, the implicit discovery of pref-

erences is fed by the browsing history of a user e.g.

clicks. An alternative approach is the use of the avail-

able information from social media in which a user

participates. Our study paper examines a method-

ology of website’s reconstruction by the concept of

hotlink assignment from a new point of view. An

algorithm is presented to ameliorate the accessibility

of non-popular pages which however are highly con-

ceptually relevant and recent trend according to so-

cial media, by adding extra links to them from highly

popular pages, resulting to fewer hops from the home-

page. Our methodology uses a local metric to rec-

ognize the accessibility of each web page separately

and the target is its minimization. Our study also de-

scribes an innovative personalization scheme discov-

ering a user’s interest through social media.

The main points of our contribution can be sum-

marized in the following sentences:

• We propose a holistic procedure of a website’s re-

construction.

• A generic scheme describes personalization

through social media.

• We formulate a personalized hotlink assignment

algorithm.

• We evaluate our methodology through users’ rel-

evance feedback.

The rest of the paper is structured as follows. Sec-

tion 2 overviews related work, we motivate our re-

search from current challenges and related studies. In

Section 3, we provide an overview of the methodol-

ogy describing each independent task. It is noted we

present the algorithm of personalized hotlink assign-

Makris, C., Siaterlis, K. and Vikatos, P.

Personalized Hotlink Assignment using Social Networks.

DOI: 10.5220/0006296800710079

In Proceedings of the 13th International Conference on Web Information Systems and Technologies (WEBIST 2017), pages 71-79

ISBN: 978-989-758-246-2

ment in Section 3.4.2. Section 4 provides an overview

of the implementation of the system for modules and

sub-modules respectively and presents a reference to

our experimental results. Finally, in Section 5, we dis-

cuss the strengths and limitations of our approach and

we conclude the paper with an outlook to future work.

2 RELATED WORK

Enhancing web browsing experience has gained the

interest of researchers. The concept of assigning

hotlinks to websites has been suggested by Perkowitz

and Etzoni in (Perkowitz and Etzioni, 2000) where a

site is transformed using shortcutting in order to be

browsed efﬁciently by users. The clairvoyant user

model (Bose et al., 2000; Czyzowicz et al., 2001;

Kranakis et al., 2001) and the greedy user model (Ger-

stel et al., 2003; Jacobs, 2010; Jacobs, 2011; Matichin

and Peleg, 2007; Pessoa et al., 2004a; Pessoa et al.,

2004b) constitute the main methodologies in which

the presence of the hotlinks is known only for the

present node and the whole site respectively. Based

on the clairvoyant model to study (Bose et al., 2000)

has presented the problem of assigning hotlinks prov-

ing that solving the problem to a directed acyclic

graph (DAG) is NP-hard and introducing the upper

and lower bounds on the expected number to reach

leaves from the root (homepage) of a complete bi-

nary tree. Considering the website as a tree, study

(Czyzowicz et al., 2001) shows an O(n

) algorithm

for assigning a hotlink which outperforms greedy ap-

proaches. Studies (Gerstel et al., 2003; Pessoa et al.,

2004a) are focused in the greedy model with running

time exponential in the depth of the tree and thus

polynomial for trees of logarithmic depth and an im-

plementation of this algorithm able to discover opti-

mal solutions for trees as presented in (Pessoa et al.,

2004b). An approach of the natural greedy strategy

achieves at least half of the gain of an optimal solu-

tion. An algorithm of 2-approximation in terms of

the gain has been presented in (Matichin and Peleg,

2007). An improvement to hotlink assignment is pre-

sented in (Dou

ıeb and Langerman, 2005) in which a

linear-time algorithm where dynamic operations such

as node insertion, deletion and weight reassignment

are available. A common feature of all these studies is

that they do not combine the popularity of pages with

information on social media contrary to our study.

Another difference of our approach is that we han-

dle website as a directed acyclic graph contrary to the

assumption that the underlying model is a tree. Also,

we provide a unique personalized reconstructed site

for each user based on hotlink assignment.

Also, the current scientiﬁc interests focus on per-

sonalization schemes for search engines, web pages

and information systems. Personalization methodolo-

gies can be summarized in three main categories in

the way that the necessary information is collected i.e.

Explicit, implicit and hybrid. A plethora of web pages

and information systems use explicit personalization

and infer user interest based on predeﬁned categories

that the user should select. For example, the well-

known search engine Google asks users to create a

proﬁle by selecting categories of interests. Contrary

to this technique study (Kelly and Teevan, 2003) ex-

amines the improvement to search accuracy through

personalization using implicit feedback information.

Another proposed method (Matthijs and Radlinski,

2011) collects web usage data e.g. page session, URL

and duration of visit; to discover the preferences of a

user. Also, study (Peng et al., 2012) proposes a struc-

ture of categories (tree) with reference to Google di-

rectory. The category tree is updated through visits of

websites and shows the degree of interest. Also, there

are studies that combine implicit feedback from user

and information of social media interactions. Study

(Carmel et al., 2009) describes the creation of a proﬁle

for each user using social networks for improvement

in web search. Also, study (Zhou et al., 2012) gathers

information from social media and implicitly person-

alizes the search results via query expansion. Further-

more, hybrid personalization schemes combine im-

plicit and explicit methods as study (Noll and Meinel,

2007) presents. Our methodology differs from the

previous ones by using a topic modeling algorithm

to social media text in order to create a graph where

each node constitutes a category of interest and each

weighted link the correlation between categories. The

importance of nodes in terms of ingoing and outgoing

links describes users’ preferences.

3 MODEL OVERVIEW

A summary of independent tasks that our methodol-

ogy consists of is given below:

• Generation of website’s graph. The website is

modeled as a directed graph G(V,E) where nodes

and edges are pages and links of the website

respectively. Each node has a popularity at-

tribute and each edge includes a context-similarity

weight.

• Extraction of website’s categories. Pages include

semantic information for search engine optimiza-

tion purposes. We extract the category of each

web page and create a list of categories for the

whole website.

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

Figure 1: System Architecture.

• Extraction of user’s interest. Our personalization

scheme produces a list of categories derived from

user’s texts in social media. We gather raw texts

that the user posts on social media, then a topic

modeling algorithm produces the topics that are

introduced to a personalized category graph and

the Pagerank algorithm creates a ranked list of cat-

egories indicating user’s top preferences.

• Creation of Hotlinks. We use shortcuts for a node

to one of its descendants. The goal is the enhance-

ment of the browsing experience. The distance

between the home page and popular pages is re-

duced.

• Reconstruction of website. A variation of Pager-

ank algorithm prioritizes pages in terms of pop-

ularity, context-similarity and personal interest.

The existence of new links favors the reconstruc-

tion of the website.

In the following subsection tasks and modules of our

model are described in detail and Figure 1 presents

the system architecture.

3.1 Generation of Website’s Graph

We model the website as a graph G(V, E). Our pro-

cedure uses crawling which is handled by a dedicated

crawler that is developed for the task and that allows

sampling pages and links in a manner that network

properties are preserved and can be used in our mod-

eling procedure. The crawling is oriented to discov-

ering new nodes in a Breadth-First search (BFS) ap-

proach. We introduce an attribute calculating the pop-

ularity of each page/node and taking account a page’s

clicks as well as ranking in the social media trend list.

In addition, we use a term based text similarity ap-

proach to compare all web page pairs. The value of

similarity in each pair is stored in NxN matrix where

N is the number of different web pages.

3.2 Extraction of Categories

In our methodology, we declare that each page be-

longs to one category. For instance, the bbc.com uses

categories such as news, sports, TV, music to structure

the information. Initially, the task ﬁnds the category

of each page using 2 different approaches:

1. Our crawler isolates the semantic information of

the page in order to extract the category of the

page e.g. RDF/XML

2. Topic modeling is performed at the text a web

page contains for the extraction of the topics.

From those topics a category is assigned.

We gather the categories and create a list which is

used in our personalization scheme as it is described

below.

3.3 Personalization Scheme

In this section, our personalization approach is de-

scribed. There are three main approaches to extract a

user’s interests for personalization purposes i.e. Ex-

plicit, implicit & hybrid. Our scheme belongs to

the implicit category using information through so-

cial media in order to determine the user’s preferences

as Figure 2 presents. Our personalization scheme

has been inspired from study (Makris et al., 2008)

in which a graph of categories is used depicting the

current user’s preferences in search engine results in

a query. We adopt the method of using a graph, but

we differentiate extracting topics from the social net-

work. Our methodology extracts this type of infor-

mation via social media and more speciﬁcally using

user’s texts e.g. posts, tweets. The rest of this section

is dedicated to explaining in depth the methodology

of our scheme.

First of all, the system can operate as intended given

the fact that:

Personalized Hotlink Assignment using Social Networks

Figure 2: Personalization Scheme.

• The user is active on social networks

• The user willingly connects his/her social network

account with the system

A time window is deﬁned and during this period texts

are collected constituting the current user’s interest.

The next step is the creation of a supertext for each

user which is reﬁned by a preprocessing module in-

cluding stop-words removal, tokenization and stem-

ming. The outcome of the preprocessing step is a

vector of words that is introduced in a Topic mod-

eling algorithm. Latent Dirichlet Allocation (LDA)

(Krestel et al., 2009) is used for the extraction of top-

ics. The extracted topics are semantically compared

with ﬁxed categories. This step is necessary to con-

nect the user’s interests with the categories from an

information system; e.g. search machine, website. To

compare a topic with a category, Wu&Palmer metric

(Equation 1) calculates semantic similarity using the

depths of two synsets in the WordNet taxonomies (Wu

and Palmer, 1994; Pedersen et al., 2004), along with

the depth of the LCS (Least Common Subsumer).

wup(s1,s2) =

2 ∗ depth(LCS)

depth(s1) + depth(s2)

(1)

Each word in the extracted topics is compared to cat-

egories and the category with the maximum similar-

ity is stored in a graph H if it exceeds a threshold

T . Graph H called the category graph, is a complete

graph containing nodes, which are the categories and

weighted links between categories whose weights are

formulated by the Equation 2,

w(e(u,v)) =

∑

(LDAr(t

)[LDAr(t

) − LDAr(t

)])

−1

(2)

where t

is the topic k that exceeds threshold T and

belongs to the category u. LDAr(t

) and LDAr(t

)

constitute the ranking of topic k and l by using LDA

procedure respectively. All topic combinations of the

nodes u,v are used.

For instance, to calculate the weight of the edge

e(1,2) in Figure 2, the ranking of topics Cat 1 and

Cat 2 is extracted and the weight is calculated in the

following manner: w(e(1,2)) =

1∗(2−1))

1∗(4−1)

3∗(2−3)

3∗(4−3)

= 1 +

−

= 4/3.

The weights in links are updated when a new sample

of texts is mined for a speciﬁc user. The proposed per-

sonalization scheme is generic and can be applied to

search engines in order to improve the ranking results.

3.4 Creation of Hotlinks

3.4.1 Hotlink Assignment

The concept of hotlink assignment (Perkowitz and

Etzioni, 2000; Czyzowicz et al., 2001; Bose et al.,

2000; Pessoa et al., 2004a) constitutes a methodology

of websites’ reconstruction, concerning the popular-

ity of the web pages. The goal is the enhancement of

browsing experience, reducing the distance between

the homepage and popular nodes by adding hotlinks

(shortcuts from a node to one of its descendants). Ac-

cording to previous studies, a website can be modeled

as a tree T = (V, E) which V is the set of web pages

and E is the set of links. Each leaf-web page con-

tains a weight representing the popularity of the web

page. We declare that T

is the tree formulated by an

assignment A of hotlinks. The expected number of

steps needed from the homepage to reach a web page

on a leaf is calculated by Equation 3

E[T

, pop] =

∑

is a lea f

(i)pop(i) (3)

where d

(i) is the distance of the leaf i from the root

in T

, and pop = {pop

: i is a lea f } is the proba-

bility distribution which derives from the distribution

of popularity weights on the leaves in initial tree T.

The minimization of this equation is the scope of the

hotlink assignment algorithm.

Study (Antoniou et al., 2010) proposes an innova-

tive methodology which not only uses the frequency

of accessibility of popular pages, but also introduces

the context-similarity of websites in the decision of

adding extra links from non-popular pages. We use

the same methodology as it is described in (Antoniou

et al., 2010), however, our approach combines infor-

mation given by social media in order to recalculate

the value of popularity in each node. In our methodol-

ogy a website is modelled as a directed acyclic graph

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

(DAG), G(V, E) where |V | = n, i.e. website has n

pages, and each edge has a weight w

(i, j) ∈ [0, 1],

which declares the content similarity between pages i

and j. Also each node has a popularity weight pop

In related works (Perkowitz and Etzioni, 2000; Czy-

zowicz et al., 2001; Bose et al., 2000; Pessoa et al.,

2004a; Antoniou et al., 2010), pop

is calculated by

the distribution of clicks in order to split web pages in

POP and NonPOP sets which include the popular and

non-popular web pages respectively. Contrary to this

concept, we consider web pages that have many vis-

its in the past may also be obsolete and outdated and

thus we introduce the information of trending topics

through social media.

Let T T = {tt

,...,tt

} be a set of Top-N trend topics

from social media, ranked from 1 to n. Also C

is the

category of the i

web page. The popularity of the

web page is formed by social media current trends as

the following equation shows:

rankFact = (1 + log

(N − rank(tt

)))

simFact = (1 + wup(tt

))

pop

= rankFact ∗ simFact ∗ clicks

(4)

where wup(tt

) is the maximum Wu&Palmer sim-

ilarity (Wu and Palmer, 1994; Pedersen et al., 2004)

between each trend and the category i, clicks

is the

s web page’s clicks and the rank(tt

) is the rank of

the topic with the maximum Wu&Palmer similarity.

The popularity of a website is increased if and only if

the category has the same or nearby semantic with the

trend and the trend is highly ranked. The lowest value

of pop

equals the number of visits to the web page

(clicks

The algorithm of hotlink assignment discovers all

paths between a random page (source) of the graph

and a page from the POP set (target). We count the

popularity of a path as it follows:

Path pop

i, j

∑

pop

#in edges

(5)

Figure 3 describes the calculation of the path’s popu-

larity. Candidate hotlinks are assigned between Non-

POP nodes, that exist in the path with the maximum

Path pop, and the target node. There are two criteria

that deﬁne the ﬁnal selection of hotlinks. Firstly, we

examine if the distance between the target and source

is reduced. Then, we check the semantic similarity

between target and NonPOP page. The hotlink with

the most semantic similarity is created. The Non-

POP page that is linked with a hotlink is removed

from the NonPOP set and the procedure is contin-

ued until NonPOP set is empty. Then the remain-

ing POP set is split as follows |POP| = |NonPOP| =

|Web − pages|/2. The algorithm is described in detail

in Section 3.4.2.

Figure 3: For example, path pop(A, D) =

∑

pop

node

#in edges

12/3 + 6/2 + 2/1 + 4/1 = 13. Let (A, B,C,D) be the path

with the maximum popularity. Let B be a node from Non-

POP set, creating an edge(red line) the path is reduced by

one.

3.4.2 The Algorithm

In this section, the personalized hotlink assignment is

described in the Algorithm 1. The algorithm is initi-

ated with the input elements of the Table 1.

Table 1: Algorithm’s Variables.

Variable Notion

G(V,E) Website Structure as a graph

userID User’s social media id := j

WC Website’s Categories

Th Threshold

CM Context-Similarity Matrix

pr Preprocessed super text of user j

Categories of Personalization Graph

TR Social media trends

T P

List of topics of User j

(V,E) Personalization Graph

At lines 1 through 5, the algorithm performs the ini-

tialization step and an ordered list of interest is ex-

tracted. The Procedure 1, rank Cat Graph(), is used

to extract the ordered list of categories PR

of each

user j. Internally the procedure uses a preprocess-

ing phase, which includes tokenization, the removal

of stopwords, the extraction of tokens’ lemma, in or-

der to create a supertext of terms ST

pr that they will

be introduced in LDA algorithm. The PageRank algo-

rithm creates the ranking list PR

of categories in the

category graph PG which each node is the website’s

category WC. The speciﬁcation of nodes and edges’

weight has been described in Section 3.3. At lines 6

Personalized Hotlink Assignment using Social Networks

to 9, the popularity of each node is calculated using

Equation 4. At lines 10-16, the POP and NonPOP

lists are created based on the distribution of popular-

ity in the nodes of the graph G.

Procedure 1: rank Cat Graph().

1: input T h, WC,C

, ST

2: output PR

3: PG

(V,E) = {}

4: T P

= LDA(ST

pr)

5: for each t ∈ T P

6: maxSim, cat := ﬁndMax(wup(t,WC))

7: if maxSim > T h then

8: C

(cat) = C

(cat) ∪ LDAr(t)

9: end if

10: end for

11: for each l ∈ C

12: for each m ∈ C

13: l,m := t

topic on cat C

14: if l 6= m then

15: V

= V

∪ m ∪ l

16: E

= E

∪ e(m,l)

17: w(e(m,l)) = Equation 2

18: end if

19: end for

20: end for

21: return PageRank(PG

)

From line 17 to 32, the algorithm creates the hotlinks

as it is described in Section 3.4.1. A slightly adjusted

PageRank (Aux) is used to rank the nodes of the graph

H combining the incoming links with nodes’ context-

similarity nodes. The similarity between web pages

is called Simetric and is initially equal to

#webpages

and Aux is equal to 0. We use the following recursive

formula for web page i:

Aux(i) = Aux(i) +

∑

v∈In(i)

Simetric(v) ∗ sim(v,i)

#outgoing links o f v

(6)

where In(i) is the incoming web pages to page i,

Simetric(i) =

#webpages

+(1 − q) ∗ Aux(i) where q is a

dumping factor and sim(v,i) is the context-similarity

of web pages v and i.

The output of the algorithm is a ranking list of web

pages based on popularity, context-similarity and per-

sonalized information from social media.

3.5 Reconstruction of Website

Algorithm 1 provides a list of links in a descending

order. Our approach uses this information in order to

reconstruct the website.

Algorithm 1: Personalized Hotlinks.

1: input G(V,E), userID, WC, Th, T

i, j

, C

, T R,

2: output H(V,E), R

= {}

3: H = G

4: PR

= rank Cat Graph(T h, WC,C

, PG

(V,E),

pr)

5: G

new

= G

6: for each node ∈ G

new

7: maxSim, rank(tt

) =

ﬁndMax(wup(T R, node{

category

}))

8: w

node

= Equation 4

9: end for

10: while |PoP| < |G|/2 do

11: maxPOP, maxNode = ﬁndMaxWeight(G

new

)

12: G

new

= G

new

− {maxNode}

13: POP = POP ∪ maxNode

14: end while

15: NonPOP = NonPOP ∪V

new

16: P

POP

= probabilityDistr(POP)

17: while NonPOP 6= {} do

18: source = chooseRand(G,1/|G|)

19: target = chooseRand(POP,P

POP

)

20: Path = ﬁndMaxPopPath(G, source, target)

21: Path = Path − {Path ∩ POP}

22: for each node ∈ Path do

23: G

temp

= H

24: E

temp

= E

temp

∪ e(target,node)

25: if minPath(G

temp

, source, target) <

minPath(G, source, target) then

26: candNodes = candNodes ∪ node

27: end if

28: end for

29: y = f indMaxCSim(candNodes,target)

30: E

= E

∪ e(target,y)

31: end while

32: R = Aux(H)

We declare that this procedure does not remove

links or nodes from the website. The reconstructed

website is updated when a current time window of so-

cial media crawling occurs or new pages are added to

the website. We denote that changes of user’s inter-

est can be detected by continuous sampling of his/her

social media activity. Also, we introduce the informa-

tion of trending topics through social media in order

to ﬁlter web pages that, even if they had many visits

in the past might be obsolete. The restructure starts

from the homepage of the website. The context of the

homepage and the links to other pages remain on the

page. We retrieve from graph H the links of the page

and the new ones (hotlinks) that have been created.

We place the new links based on the ranking list. We

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

traverse the graph using the outgoing links in a

Breadth-First search (BFS) approach. The procedure

continues until the content of all pages and the new

links are placed on the reconstructed website.

4 EXPERIMENTAL RESULTS

4.1 Implementation

We conducted an experimental procedure using rele-

vance feedback from users. The scenario includes the

creation of a web interface in which users are regis-

tered via a Twitter account. If the Twitter account is

active our system gathers the necessary information

(tweets & tweets of user’s friends). Then, a ranking

list of links with title, description and image is pre-

sented. The ranking list is produced using the algo-

rithm of hotlink assignment as it is presented in (An-

toniou et al., 2010). Each user has the ability to read

the description, browse to the provided links and the

obligation to score each page in scale 0-3 according

to his/her interest, where 0 means not interested and

3 means very interested. After submission of scores,

a new list of ranked links is presented based on Al-

gorithm 1. The user acts in the same way as in the

previous stage.

We aggregated news and their links from the

BBC

and a graph was created based on this well-

known website. The popularity of each page was re-

trieved from the web-based company Alexa

which

provides popularity metrics for domains and sub-

domains.

We used Twitter to extract the necessary informa-

tion. Our server gathers the recent 100 tweets from

the tested user as well as tweets from the recently

mentioned user’s friends. The selected tweets initiate

the personalization procedure as described in Section

3.3. It is noted that the LDA algorithm is formulated

for 10 topics in user’s supertext. Based on tests that

we selected the average number of tweets and men-

tioned friends was 73 and 34 respectively and the av-

erage number of nodes in the category graph was 24

as Table 3 shows.

Our system was implemented in Python 2.7. We

collected data from users and the users’ friends us-

ing the Twitter API and we performed topic modeling

on the tweets using LDA

. We used NLTK

module

for preprocessing and content similarity measurement

http://www.bbc.com/news

http://www.alexa.com/

https://pypi.python.org/pypi/lda

http://www.nltk.org/

and networkx

module for graph handling. The web

interface for the evaluation phase was designed with

PHP/HTML and users’ relevance feedback data was

stored on a dedicated MySQL server for post analy-

sis.

4.2 Results

Providing a better insight into the quality of our exper-

imental data, Table 2 and Table 3 present the statistics

of the website’s DAG as well as the average tweets

per user and the number of friends we have sampled

respectively.

Table 2: Hotlink Assignment Graph Stats.

Properties Value

Nodes 1220

Edges 2226

Avg Degree 1.824

Longest Path 15

Table 3: Personalization Stats.

Properties Value

Avg #tweets 73.6

Avg # rel categories 24.3

Avg # of friends 34.2

In Figure 4 we compare the distances from home-

page to random nodes between the context-similarity

based hotlinks assignment as it is presented in study

(Antoniou et al., 2010) and our implementation. We

present the percentage improvement in the distance as

we gradually increase the size of the website’s DAG.

We can consider that our implementation outperforms

and improves the distance by an average of 2%. Fur-

thermore, Figure 5 presents the improvement of rank-

ing on the nodes between our implementation and the

algorithm of (Antoniou et al., 2010). We present the

number of improved nodes increasing gradually with

the size of the website’s DAG. We can see that the

ranking of nodes is improved as we raise the Aux

of nodes relevant to the trending topics from social

media. Our website consists of 1200 nodes and the

algorithm ranks 60 nodes better due to trending top-

ics. Furthermore, we evaluate users’ relevance feed-

back via the normalized Discounted Cumulative Gain

(nDCG) as Equation 7 shows

nDCG

DCG

iDCG

(7)

https://networkx.github.io

Personalized Hotlink Assignment using Social Networks

Figure 4: Difference % on improved distance.

Figure 5: Improvement on nodes’ ranking.

DCG

= rel

∑

i=1

rel

log

(i + 1)

(8)

where, rel

is user’s response in the range 0(not in-

terested) - 3(very interested) for the i

result. The

iDCG is the ideal ranking based on user’s preferences.

We present the nDCG values in Figure 6. We con-

ducted the experiment over 30 individuals

. Each

individual evaluates our methodology for 10 inde-

pendent times. We present the average nDCG value

for each individual and for both implementations in

Figure 6. We consider that our implementation ef-

ﬁciently targets the user’s interests in comparison to

study (Antoniou et al., 2010). More speciﬁcally 83%

of users’ responses declare the provided information

on our methodology depicts their preferences.

5 CONCLUSIONS

Our study deals with the problem of personalization

in hotlink assignment. The ﬁrst innovation in our

methodology is the use of information about social

media and, in particular, trend topics on Twitter in or-

der to recalculate the attribute of popularity in each

undergraduate students of Computer Engineering and

Informatics Department, University of Patras, Greece

Figure 6: nDCG - Hotlink vs Personalized Hotlink.

node of the graph. It is noted that previous works deal

with the popularity of a web page only with clicks

that pages receive from users during browsing. The

second contribution is the detailed description of a

personalization scheme that handles explicit informa-

tion from raw texts of social media and creates a rank-

ing list of categories describing users’ interests. The

scheme is generic and can be used in implementations

that need personalization such as search engines and

information systems. Also, a new algorithm of per-

sonalized hotlink assignment is described. According

to the experimental procedure, our methodology pro-

vides efﬁcient results in terms of distance between the

homepage and other pages, ranking of web pages and

relevance on users’ preferences.

The main points of our contribution based on the

experimental results can be summarized in the follow-

ing sentences:

• We reduce the distance between the homepage

and other popular pages in the graph in a way that

affects the browsing experience due to the fact that

users can reach in fewer steps his/her preferences.

• We differentiate the nodes ranking via relevance

of the trend topics on Twitter. The experiments

show that the increase of the number of nodes

is correlated to the number of nodes that are im-

proved in terms of ranking.

• We provide a ranking list of web pages that efﬁ-

ciently targets the user’s interests.

As future work, we are interested in examining

the comparison of our methodology in different social

networks such as Facebook and identify the parame-

ters that inﬂuence the results of our algorithm. Our

study can capture user’s preferences in a single lan-

guage, however. our plan is to extend our work for

multilingual personalization.

WEBIST 2017 - 13th International Conference on Web Information Systems and Technologies

REFERENCES

Antoniou, D., Garofalakis, J., Makris, C., Panagis, Y., and

Sakkopoulos, E. (2010). Context-similarity based

hotlinks assignment: Model, metrics and algorithm.

Data & Knowledge Engineering, 69(4):357–370.

Bharambe, I. and Makhijani, R. K. (2014). Design and

implementation of search engine using vector space

model for personalized search. International Journal

of Advanced Research in Computer Science and Soft-

ware Engineering, ISSN, 2277:1019–1023.

Bose, P., Czyzowicz, J., Gasieniec, L., Kranakis, E.,

Krizanc, D., Pelc, A., and Martin, M. V. (2000).

Strategies for hotlink assignments. In International

Symposium on Algorithms and Computation, pages

23–34. Springer.

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 Proceedings of the 18th

ACM conference on Information and knowledge man-

agement, pages 1227–1236. ACM.

Czyzowicz, J., Kranakis, E., Krizanc, D., Pelc, A., and Mar-

tin, M. V. (Jun. 25-28, 2001). Evaluation of hotlink

assignment heuristics for improving web access. In

Proceedings of the Second International Conference

on Internet Computing (IC’01), volume 2, pages 793–

799. Citeseer.

Dou

ıeb, K. and Langerman, S. (2005). Dynamic hotlinks.

In Workshop on Algorithms and Data Structures,

pages 182–194. Springer.

Gerstel, O., Kutten, S., Matichin, R., and Peleg, D. (2003).

Hotlink enhancement algorithms for web directories.

In International Symposium on Algorithms and Com-

putation, pages 68–77. Springer.

Jacobs, T. (2010). An experimental study of recent hotlink

assignment algorithms. Journal of Experimental Al-

gorithmics (JEA), 15:1–1.

Jacobs, T. (2011). Constant factor approximations for the

hotlink assignment problem. ACM Transactions on

Algorithms (TALG), 7(2):16.

Kelly, D. and Teevan, J. (2003). Implicit feedback for infer-

ring user preference: a bibliography. In ACM SIGIR

Forum, volume 37, pages 18–28. ACM.

Kranakis, E., Krizanc, D., and Shende, S. (2001). Approx-

imate hotlink assignment. In International Sympo-

sium on Algorithms and Computation, pages 756–767.

Springer.

Krestel, R., Fankhauser, P., and Nejdl, W. (2009). La-

tent dirichlet allocation for tag recommendation. In

Proceedings of the third ACM conference on Recom-

mender systems, pages 61–68. ACM.

Makris, C., Panagis, Y., Plegas, Y., and Sakkopoulos, E.

(2008). An integrated web system to facilitate per-

sonalized web searching algorithms. In Proceedings

of the 2008 ACM symposium on Applied computing,

pages 2397–2402. ACM.

Matichin, R. and Peleg, D. (2007). Approximation al-

gorithm for hotlink assignment in the greedy model.

Theoretical Computer Science, 383(1):102–110.

Matthijs, N. and Radlinski, F. (2011). Personalizing web

search using long term browsing history. In Proceed-

ings of the fourth ACM international conference on

Web search and data mining, pages 25–34. ACM.

Noll, M. G. and Meinel, C. (2007). Web search personal-

ization via social bookmarking and tagging. In The

semantic web, pages 367–380. Springer.

Pedersen, T., Patwardhan, S., and Michelizzi, J. (2004).

Wordnet:: Similarity: measuring the relatedness of

concepts. In Demonstration papers at HLT-NAACL

2004, pages 38–41. Association for Computational

Linguistics.

Peng, X., Niu, Z., Huang, S., and Zhao, Y. (2012). Personal-

ized web search using clickthrough data and web page

rating. Journal of Computers, 7(10):2578–2584.

Perkowitz, M. and Etzioni, O. (2000). Towards adaptive

web sites: Conceptual framework and case study. Ar-

tiﬁcial intelligence, 118(1):245–275.

Pessoa, A. A., Laber, E. S., and de Souza, C. (2004a).

Efﬁcient algorithms for the hotlink assignment prob-

lem: The worst case search. In International Sympo-

sium on Algorithms and Computation, pages 778–792.

Springer.

Pessoa, A. A., Laber, E. S., and de Souza, C. (2004b). Efﬁ-

cient implementation of hotlink assignment algorithm

for web sites. In In Proceedings of the Workshop on

Algorithm Engineering and Experiments, pages 79–

87.

Wu, Z. and Palmer, M. (1994). Verbs semantics and lexical

selection. In Proceedings of the 32nd annual meeting

on Association for Computational Linguistics, pages

133–138. Association for Computational Linguistics.

Zhou, D., Lawless, S., and Wade, V. (2012). Improving

search via personalized query expansion using social

media. Information retrieval, 15(3-4):218–242.

Personalized Hotlink Assignment using Social Networks