ARE RECOMMENDER SYSTEMS REAL-TIME

IN MOBILE ENVIRONMENT?

Towards Instantaneous Recommenders

Armelle Brun and Anne Boyer

LORIA-Nancy University, 615, rue du jardin botanique, 54506 Vandœuvre les Nancy, France

Keywords:

Recommender systems, m-Commerce, Computation time, Real-time, Usage mining, Association rules.

Abstract:

Recommendation technologies have traditionally been used in domains such as e-commerce to recommend

resources to customers so as to help them to get the right resources at the right moment. The interest of model-

based collaborative ﬁltering, as sequential association rules, in recommender systems has highly increased

over the last few years. These models are usually presented as real-time recommenders. In the last few years,

the m-commerce domain has emerged, that displays recommendations on the mobile device instead of the

classical screen of the computer. In this paper user privacy preservation is an important objective and one way

to be compliant with this constraint is to store the recommender on the mobile-side. Though model-based

recommenders are real-time, many of them require a signiﬁcant time to generate recommendations to users

and may not be real-time anymore when implemented on a mobile device. Although some works focused on

the way to decrease the time required to compute recommendations, the computation complexity still remains

relatively high. We put forward a new incremental recommender to get instantaneous recommendations when

exploiting usage mining recommender systems in the framework of m-commerce.

1 INTRODUCTION

The democratization of the Internet and network tech-

nologies has resulted in an increase in the volume of

information easily accessible. However, this profu-

sion of information leads to unsatisﬁed users as they

cannot get easily the information they search. A crit-

ical issue of Web applications is thus the incorpora-

tion of mechanisms for delivering information that

ﬁts user attempts to increase their satisfaction. In e-

commerce or m-commerce environments, it is a way

to increase both customer ﬁdelity and the selling rate

of the (web) stores.

Recommender systems (RS) are such a mecha-

nism, they provide users with personalized recom-

mendations on products based on either the knowl-

edge about users or their past behavior.

To build a RS based on users’ past behavior, one

can either analyze the content of the previously ac-

cessed resources by the active user (the content-based

approach) or use the information about which re-

sources have been accessed by which users, called

collaborative ﬁltering (CF).

CF algorithms can be classiﬁed into memory-

based and model-based algorithms (Adomavicius and

Tuzhilin, 2005). Memory-based RS face a scalabil-

ity problem. Model-based RS cope with this scala-

bility problem due to their off-line computation part.

In recent years, the interest of data mining models in

the framework of RS (Bozdogan, 2004) has highly in-

creased. In this paper we are interested in data min-

ing models and speciﬁcally the well-known associa-

tion rules.

In model-based approaches the computation time

is however still a question: can recommendations be

really provided within a small time?

Model-based and particularly data mining-based

RS are often presented as real-time RS (Huang et al.,

2006). What is a real-time system ? A real-time sys-

tem “updates information at the same rate as it re-

ceives data”

. In RS, the data received is the set of

user consultations. New data is thus available each

time the active user consults a resource. Thus, a real-

time RS is able to compute recommendations within

two resource consultations of a given user.

RS are usually implemented on the server-side,

they may however be implemented on the client-side.

In mobile-commerce (m-commerce) for example, the

http://www.thefreedictionary.com

101

Brun A. and Boyer A.

ARE RECOMMENDER SYSTEMS REAL-TIME IN MOBILE ENVIRONMENT? - Towards Instantaneous Recommenders.

DOI: 10.5220/0002807801010106

In Proceedings of the 6th International Conference on Web Information Systems and Technology (WEBIST 2010), page

ISBN: 978-989-674-025-2

RS can be implemented on the mobile. We consider

in thus work that privacy preservation is an important

feature to make users use a RS. Implementing RS on

the client-side has the advantage of preserving user

privacy, we thus choose to implement the RS on the

client-side.

M-commerce can be deﬁned as follows “m-commerce

applications not only cover the electronic commerce

applications, but also include applications that can

be performed at any time and from anywhere by us-

ing mobile computing technology” (Hu et al., 2006).

M-commerce is a recent application domain that has

emerged in the last decade, due to improvement of

mobile terminal capability, increase of bandwidth,

wireless and mobile networks. The research in recom-

mendation algorithms for mobile commerce has thus

also been accelerated (Tarasewich, 2003).

As m-commerce and e-commerce have a lot in

common, we are interested in studying the exploita-

tion of the well-known RS from e-commerce in the

m-commerce domain. Nevertheless, mobiles have

several limitations as a low computation capacity, in-

put/ouput facilities and memory. Due to the computa-

tion limitation, one can ask if RS are real-time when

implemented at client-side.

To answer this question, we speciﬁcally focus on

the exact way data mining models compute recom-

mendations and put forward a new incremental RS to

get instantaneous recommendations.

Section 2 presents RS in m-commerce and de-

scribes our scenario of application. Section 3 is an

overview of the SAR-based RS. Section 4 is inter-

ested in works about the improvement of the com-

putation time in data mining RS. The next section

presents the way recommendations are computed in

SAR-based RS. Then, we suggest a new incremental

RS that reduces the computation time. The last sec-

tion concludes and presents perspectives.

2 RS IN M-COMMERCE AND

SCENARIO

2.1 RS in m-Commerce

Related to the success of RS in e-commerce, person-

alized RS are now emerging in m-commerce. The

application domains are wide and varied, as tourism

or restaurant recommendations (Wan, 2009). M-

commerce has however some differences and speci-

ﬁcities compared to e-commerce, such as mobility,

limited processing, transmission capability, size of

display, battery power, etc. (Zenebe et al., 2005).

In m-commerce, the RS can be implemented in

several ways. In (Wan, 2009), it is implemented on

the server side. The mobile registers information

about the users and transmits it to the server that

computes recommendations. These recommenda-

tions are then sent to the mobile that displays them.

This solution faces security (connection stability) and

privacy problems.

(Tveit, 2001) proposes a P2P RS to cope with the

scalability problem of RS. In that case, the number

of communications between mobiles and the com-

munication capability is a problem and the author

stresses the importance of the computation power of

mobiles. This approach is also sensitive to malicious

behaviors.

RS can also be implemented on the client side (Lee,

2004). Client-side RS face several drawbacks such as

computational capacity and memory size. However,

communication bandwidth, security and privacy are

no more a problem as there is no communication

between the server and the mobile.

2.2 Scenario

This paper focuses on RS in m-commerce in the con-

text of supermarkets. The RS observes the users be-

havior (the products the users have been interested in)

and suggests them relevant products (linked to previ-

ous interests, special offers, etc.).

Concretely, the products the users are interested in

are captured by the mobile and recommendations are

displayed on the screen of the mobile.

In our scenario, we propose to use the RFID tech-

nology. The mobile is the RFID reader. A RFID tag

is put on items and is detected when it is nearby the

reader. We use this technology to identify the items a

user puts in his caddy and those he takes in his hand.

2.2.1 Recommend to Users

Customers use a caddy that they ﬁll with the products

they want to buy. A mobile device provided by the

store is placed on each caddy. Customers thus do not

need to use their own mobile (they can even not have

any mobile).

In that context, our priorities are user privacy and

computation time of recommendations. If the RS is

implemented on the server side, the number of com-

munications between the mobiles and the server is

huge. Moreover, quality and security of communi-

cations between the server and the mobile may be a

drawback, as presented in the previous section, lead-

ing to a non preservation of the privacy. To avoid

these problems, we choose to implement the RS on

WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies

102

the mobile-side. The remaining problem is thus the

computation time.

Each time a user is interested in a product i.e.

when he puts it in his caddy or takes it in his hand,

this information is captured by the mobile and is ex-

ploited by the RS on the mobile to propose to the user

the products related to those he has been interested in.

The customers do not need to be identiﬁed and no in-

formation is explicitly asked to them, users’ privacy

is thus preserved. Moreover, the information auto-

matically collected about the user is only known by

the mobile, this information is not transmitted to any

other device or server when the user shops.

2.2.2 Train the RS

The mobile device registers all the items the users are

interested in. When the customer arrives at the cash

desk, the information in the mobile device is transmit-

ted to the server. If the user has no discount card, the

information transmitted is anonymous. The RS model

is then trained and built on the server side. The order

of purchase is important here to recommend items to

users that they will see next and not recommend them

items they have already seen.

The only information transmitted to the server is

the sequence of products the user has actually pur-

chased. Those he has only been interested in are not

sent to the server. Thus, the information stored on the

server is similar to those stored by classical systems

with discount cards.

As the user does not need to be identiﬁed, no pri-

vate information is stored, and users are more enthusi-

astic to use the system. Moreover, the model is stored

only on mobile devices that are owned by the store,

it cannot be picked up by other persons or stores and

users do not have to store a software on their own mo-

bile. Additionally, the transactions with the server are

limited, the system is thus more secured.

3 DATA MINING MODELS IN RS

Many models have been studied in RS (Adomavicius

and Tuzhilin, 2005), among which we can ﬁnd data

mining-based RS. Due to the success of these mod-

els, we are interested in their ability to be used in m-

commerce. We present here data mining RS.

Data mining refers to “discovering, extracting and

analyzing knowledge from a large amount of data”

(Han and Kamber, 2001). In the framework of e-

commerce or m-commerce, the data is made up of

the traces of usage (traces of navigation) and the data

mining algorithms are used to predict user’s prefer-

ence and future behavior.

Data mining RS proceed in two phases to perform

recommendations. In the off-line phase, regular us-

age patterns of users are extracted. The on-line phase

is devoted to the generation of recommendations. The

RS exploits both the active user’s session and the pat-

terns/relationship obtained from the off-line phase.

The active user’s session is matched against the pat-

terns of the model to compute the scores of the re-

sources. The recommendation process is thus made

on-line. The resources with the highest scores are rec-

ommended to the user.

Usage mining approaches have the advantage to

be used to compute recommendations without requir-

ing that the active user is identiﬁed, or to compute

recommendations to an unknown user, which is a very

common situation in practice. In e-commerce and m-

commerce, it is of high importance to provide users

with recommendations when they are not identiﬁed,

not only to recommend them resources as yet as they

arrive on the web site (to improve their satisfaction)

but also to attract potential customers.

Sequential Association Rules (SAR) (Srikant and

Agrawal, 1996) are are used to capture ordered rela-

tionships between resources within sequences of nav-

igation. A SAR is an expression of the form X ⇒ Y ,

where X (called the antecedent) and Y (the conse-

quent) are sequences of resources. X ⇒ Y is consid-

ered as a SAR if both its support and conﬁdence are

above two thresholds that have to be ﬁxed.

A SAR means that, when users consult the se-

quence of resources in X , they usually consult Y . SAR

are widely used techniques in RS to discover useful

links within or between user sessions of navigation.

The off-line part of SAR-based RS searches all the

SAR according to their support and conﬁdence. Most

of the algorithms used to mine SAR are incremental

as GSP (Srikant and Agrawal, 1996), an APriori-like

algorithm, where SAR made up of k +1 resources are

deduced from the SAR of k resources.

SAR-based RS exploit the sequence of resources

in the active session of the user, and compare it to

the antecedents of the SAR discovered off-line. A

rule that matches the active session (called a match-

ing rule) is a rule with an antecedent that is a subse-

quence of the active session (as

) of the active user a.

The resources (r

) that are consequence of the match-

ing rules are then candidates to be recommended.

To compute the score of each resource

S(r

|as

), the conﬁdence of each matching rule

(conﬁdence(X ⇒ r

)) is used. As several matching

rules may have the same consequence, a resource

may have several scores. Several policies are used in

ARE RECOMMENDER SYSTEMS REAL-TIME IN MOBILE ENVIRONMENT? - Towards Instantaneous

Recommenders

103

the literature to deal with this problem as max policy

or sum policy (Brun and Boyer, 2009).

4 COMPUTATION TIME IN RS

At the opposite of memory-based RS, model-based

RS provide recommendations within a small time. Al-

though this feature is an advantage, many works have

been dedicated to the reduction of the computation

time required to suggest recommendations to users

(Yan et al., 1996; Nakagawa and Mobasher, 2003;

Brun and Boyer, 2009).

To reduce the computation time, one can focus ei-

ther on the off-line phase, or on the on-line phase.

4.1 Off-line Focus

Clustering is one way to reduce computation time. In

the framework of WUM, (Yan et al., 1996) performs

a clustering of sessions and a model is constructed on

each cluster. In the on-line phase, the active user ses-

sion is classiﬁed to the previously constructed clus-

ters, and resources are then recommended. Cluster-

ing of SAR is also performed. This model is not only

lighter than a general model, but also more relevant

for the active user, compared to the generation of a

single set of SAR.

We can also focus on the way to store the model

so as to the information of the model is accessible in

a short time. Many works (Mobasher et al., 2002)

propose to store navigation patterns in a sufﬁx tree.

(Mobasher et al., 2001) presents a frequent item-

sets graph for all kth order markov models that is also

used in the framework of SAR.

4.2 On-line Focus

Some other works speciﬁcally focus on the on-line

phase to decrease the computation time.

A ﬁrst way to decrease the computation time lies

in the reduction of the size of the active session con-

sidered, by using a sliding window (Mobasher et al.,

2001). This conﬁguration assumes that only the last n

visited resources inﬂuence the choice of the next re-

source the user will consult (and thus inﬂuences the

recommendation). All the resources before the win-

dow are not considered, thus reducing the number of

resources to consider and the computation time. Ob-

viously, when mining SAR, a sliding window is also

used. (Nakagawa and Mobasher, 2003) stores sequen-

tial patterns in a frequent sequence trie. As a sliding

window is used, only frequent patterns of size equal

to the size of the sliding window are considered.

In (Brun and Boyer, 2009) we have proposed any-

time WUM RS to provide users with recommenda-

tions whenever they want, recommendations can thus

be generated even though small time is available.

Most of these works focus on the way to store the

model and access it to compute recommendations. To

our mind, the most critical part for computing recom-

mendations in these RS is the effective matching of

the active session against the patterns discovered off-

line, that can take time.

5 REDUCING COMPUTATION

TIME IN RS

We now focus on the exact way recommendations are

computed in data mining RS, and show that this com-

putation is complex. Data mining RS may thus take

time to generate recommendations under some condi-

tions. We then focus on a new way to compute predic-

tions, by incrementally constructing the recommenda-

tion list.

To our best knowledge, no work has presented the

exact way the active user session is matched against

the model. We present here how this matching can be

performed and show that this task is complex.

Let a be the active user and as

= (r

, r

, . . . , r

)

his active session. The RS has to compute the score of

each resource r

∈ R. To perform these computations,

the RS matches the active session against the model.

5.1 Classical Approach

In the framework of SAR, the model is made up of a

set of SAR. As shown in section 4, SAR are generally

stored in a tree to decrease both the space complexity

of the model and the access time to the rules.

Most of the works deal with non contiguous SAR

both off-line and on-line, i.e. there can be additional

resources between the resources of the SAR.

To compute the score of each resource in r

∈ R,

the system has to ﬁnd all the rules in the model that

match the active session as

To perform this matching, the set of sub-

sequences of the active session as

is generated. Let

SubSeq be the function mapping as

to the list of its

sub-sequences. Let ss = SubSeq(as

) be the result-

ing set of sub-sequences. Then, given this set, each

sub-sequence is searched in the SAR tree. Assum-

ing that reaching a son node in a tree, when placed

in a node has a cost of 1, (Nakagawa and Mobasher,

2003) shows that the search of a sequence ss

in a tree

is performed in O(|ss

|) where |ss

| is the length of

WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies

104

. As there are 2

possible sub-sequences of size be-

tween 1 and n in the active session of size n, matching

the active session against the set of SAR is made in 2

searches. As the cost of each search is dependent on

its size, the overall cost is equal to n· 2

n−1

. This value

is under-evaluated as updating the score of r

(as sev-

eral SAR can lead to the same resource) is costly.

In e-commerce or m-commerce, the active session

may be large. For example, in supermarkets, a session

of 30 products leads to more than 1 billion searches

in the tree and thus more than 16 billions operations

in the tree. The number of operations exponentially

increases with the size of the active session.

Obviously, when using a sliding window, this

number decreases, as the size of the window is lower.

However, we have shown that, in the frame of web

navigation (Bonnin et al., 2009), the most accurate

recommendations are obtained with a sliding window

of 10 elements (5k operations).

In conclusion, using SAR in the frame of m-

commerce when implementing the RS on the mobile

requires a large computation time, and will probably

not compute recommendations in real-time.

5.2 Towards Incremental RS

Let a customer be searching in the supermarket the

products he is interested in. The recommendations

made at a given time are computed on products that

the user has accessed previously (the active session of

the user). The user then chooses a product that can

be in the recommendation list, or not. The RS has

then to present to the user a new list of recommended

products, etc.

When a user is interested in a product, at the exact

moment he takes it in his hand, the RS has to suggest

him products related to his interests (and also related

to this last product). If the computation time is high,

no recommendation may be available at that moment

and the user may take another product in his hand be-

fore the recommendation is computed. Thus, the sys-

tem is one phase late. If the user wants a recommen-

dation, he will have to wait and may be disappointed.

However, as previously noticed, in the frame of SAR-

based RS, when implemented on mobiles, the time

required to compute recommendations may be high.

To maximize the user satisfaction, the RS should

not be real time but instantaneous. Indeed, in the case

of real-time RS, recommendations can be made just

at the moment before the user takes in his hand an-

other product. This recommendation is made too late,

it should be made instantaneously just after the user

takes a product in his hand.

Let us recall that the RS exploits the sequence of

the products the user has been interested in (by using

SAR here). At a given time t, the recommendations

proposed are not independent of the recommended

items from the preceding time t − 1. The history ex-

ploited by the RS at the time t is partially identical to

the one used at time t − 1. The only difference is the

last item the user took in his hand, all the preceding

items considered are similar. We propose to take ad-

vantage of the previously computed recommendations

to compute new recommendations. We thus propose

an incremental recommender. We choose to update

the score of each product from the previous time t − 1

by exploiting the new product of time t. The com-

plexity of the recommendation will thus be highly de-

creased (only the last resource consulted will be stud-

ied). The recommendations will be instantaneous.

Let us recall that at each step, only the resources

with the highest scores are recommended, but the

score of all the resources is computed.

In the case of max policy, if the conﬁdence of a

matching rule is above the score of the resource of

the consequence part, then its score is updated and

replaced by the new conﬁdence value. In the case of

the sum policy, whatever is the conﬁdence value of a

matching rule, the score of the consequence resource

is updated: the conﬁdence is added to the score.

The most commonly used mining algorithm (as

APriori algorithm), constructs the SAR set step by

step, by increasing the size of the antecedent: SAR

made up of k +1 resources are deduced from the SAR

of k resources. This feature allows us to store SAR in

a tree, as (Mobasher et al., 2002).

Based on this observation, at time t the set of SAR

that match the active session and contain the resource

are dependent of the set of matching rules from the

step t − 1. The search of the set of matching rules at

step t can then be oriented as the candidate matching

rules can be partially deduced from step t −1.

Let two rules m

t−1

and m

with the antecedent of

t−1

included in the antecedent of m

, and the ﬁrst

elements of the antecedent of m

are similar to the an-

tecedent of m

t−1

. The last element of m

is the only

difference between the two antecedent parts. If m

t−1

matches at time t − 1, then m

will possibly match at

time t if the last resource the user has consulted is the

last element of the antecedent of m

In the SAR tree, the two SAR m

t−1

and m

are

stored in the same branch and m

is a son node of

t−1

. We thus propose to memorize, at each time t −

1, the set of SAR in the tree that matches the active

session. Concretely, a pointer on the node of each

matching SAR is stored. At the following time t, these

pointers are used to search if the product actually seen

by the user is among the son nodes of each pointer.

ARE RECOMMENDER SYSTEMS REAL-TIME IN MOBILE ENVIRONMENT? - Towards Instantaneous

Recommenders

105

In cases some matching rules are found, the score

of each product is updated and pointers are updated

too. The computation complexity of each time will be

highly reduced in that case. Classically, with a session

of size n, there were 2

possible sub-sequences. Here,

the number of sub-sequences studied is thus equal to

the number of sub-sequences from the preceding ses-

sion n − 1. This value is thus at most 2

n−1

, the num-

ber of sub-sequences searched is thus divided by at

least 2. In terms of overall cost, the cost of each sub-

sequence is no more dependent on the length of the

sub-sequence but has a ﬁxed cost: 1. The resulting

cost is thus 2

n−1

, compared to n · 2

n−1

in classical ap-

proaches. The cost is thus divided by n.

For example, with a session of size 30, the computa-

tion time is thus divided by 30. The recommendations

will thus be more likely made instantaneously.

6 CONCLUSIONS

Recommender systems have been a great success in e-

commerce applications. M-commerce is a recent ap-

plication domain that has emerged in the last decade.

M-commerce and e-commerce share many properties

and we proposed to study the exploitation of well-

known RS in the framework of e-commerce in the m-

commerce domain. We speciﬁcally focused on data

mining RS: sequential association rules. This recom-

mendation model has a high time complexity, but due

to its implementation on a server, it runs fast enough

to be a real-time RS. After having presented the con-

text of application of our research, and our focus on

privacy preservation we chose to implement the RS

on the mobile side. We have then detailed the SAR

model and showed that the computation of recom-

mendations is time consuming. We have then pro-

posed an incremental RS. This RS leads to a smaller

complexity and allows to implement SAR-based RS

on mobiles.

This work will be pursued by a study of other recom-

mendation models to propose incremental version.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the next

generation of recommender systems: A survey of the

state-of-the-art. IEEE transactions on knowledge and

data engineering, 17(6):734–749.

Bonnin, G., Brun, A., and Boyer, A. (2009). A low-

order markov model integrating long-distance histo-

ries for collaborative recommender systems. In Proc.

of the ACM Int. Conf. on Intelligent User Interfaces

(IUI’09), pages 57–66, Sanibel Islands, USA.

Bozdogan, H. (2004). Statistical Data Mining and Knowl-

edge Discovery. Chapman & Hall/CRC.

Brun, A. and Boyer, A. (2009). Towards privacy compliant

and anytime recommender systems. In In Proceedings

of the E-Commerce and Web Technologies Conference

(EC-Web09), pages 276–287.

Han, J. and Kamber, M. (2001). Data Mining: Concepts

and Techniques. The M. Kaufmann Series in DMS.

Hu, W., Yeh, J., and Lee, S. (2006). Adaptive web

browsing using web mining technologies for internet-

enabled mobile handheld devices. In Proceedings of

the 16th Information Resources Management Associ-

ation (IRMA 2006) International Conference.

Huang, Y., Kuo, Y., Chen, J., and Jeng, Y. (2006). Np-

miner: A real time recommendation algorithm by us-

ing web usage mining. Knowledge-Based Systems,

19:272–286.

Lee, S. (2004). A mobile application of client side person-

alization based on wipi platform. In Proceedings of

the Computational and Information Science Confer-

ence (CIS04), pages 903–909.

Mobasher, B., Dai, H., Luo, T., and Nakagawa, M. (2001).

Effective personalization based on association rule

discovery from web usage data. In in proceedings of

the 3rd International Workshop on Web Information

and Data Management, pages 9–15.

Mobasher, B., Dai, H., Luo, T., and Nakagawa, M. (2002).

Using sequential and non-sequential patterns for pre-

dictive web usage mining tasks. In Proc. of the IEEE

Int. Conf. on Data Mining (ICDM’2002).

Nakagawa, M. and Mobasher, B. (2003). Impact of site

characteristics on recommendation models based on

association rules and sequential patterns. In Proceed-

ings of the IJCAI’03 Workshop on Intelligent Tech-

niques for Web Personalization.

Srikant, R. and Agrawal, R. (1996). Mining sequential

patterns: Generalizations and performance improve-

ments. In Proc. of the 5th Int. Conf. on Extending

Database Technology, pages 3–17.

Tarasewich, P. (2003). Designing mobile commerce appli-

cations. Communications of the ACM, 46(12):57–60.

Tveit, A. (2001). Peer-to-peer based recommendations for

mobile commerce. In International Workshop on Mo-

bile Commerce, Proceedings of the 1st international

workshop on Mobile commerce, pages 26–29.

Wan, Z. (2009). Personalized tourism information system in

mobile commerce. In IEEE Int. Conf. on Management

of e-Commerce and e-Government, pages 387–391.

Yan, T., Jacobsen, M., Garcia-Molina, H., and Umeschwar,

D. (1996). From user access patterns to dynamic hy-

pertext linking. In Fifth Int. World Wide Web Conf.

Zenebe, A., Ozok, A., and Norcio, A. (2005). Person-

alized recommender systems in e-commerce and m-

commerce: A comparative study. In Proc. of the 11th

Int. Conf. on Human-Computer Interaction.

WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies

106