Future Generation Computer Systems 43–44 (2015) 99–109

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

A cloud based health insurance plan recommendation system:A user centered approach

Assad Abbas a, Kashif Bilal a,b, Limin Zhang a, Samee U. Khan a,∗

a North Dakota State University, USAb COMSATS Institute of Information Technology, Abbottabad, Pakistan

h i g h l i g h t s

• We present a cloud based health insurance plan recommendation system.• We propose a standard ontological representation for all the health insurance plans.• An algorithm to determine the similarities between the user requirements and plans is presented.• We propose a ranking technique based on the Multi-attribute Utility Theory (MAUT).

a r t i c l e i n f o

Article history:Received 21 January 2014Received in revised form15 July 2014Accepted 15 August 2014Available online 27 August 2014

Keywords:Cloud computingRecommendation systemMulti-attribute utility theoryPlan ranking

a b s t r a c t

The recent concept of ‘‘Health Insurance Marketplace’’ introduced to facilitate the purchase of healthinsurance by comparing different insurance plans in terms of price, coverage benefits, and qualitydesignates a key role to the health insurance providers. Currently, the web based tools available to searchfor health insurance plans are deficient in offering personalized recommendations based on the coveragebenefits and cost. Therefore, anticipating the users’ needs we propose a cloud based framework thatoffers personalized recommendations about the health insurance plans.We use theMulti-attribute UtilityTheory (MAUT) to help users compare different health insurance plans based on coverage and cost criteria,such as: (a) premium, (b) co-pay, (c) deductibles, (d) co-insurance, and (e) maximum benefit offered bya plan. To overcome the issues arising possibly due to the heterogeneous data formats and different planrepresentations across the providers, we present a standardized representation for the health insuranceplans. The plan information of each of the providers is retrieved using the Data as a Service (DaaS). Theframework is implemented as Software as a Service (SaaS) to offer customized recommendations byapplying a ranking technique for the identified plans according to the user specified criteria.

© 2014 Elsevier B.V. All rights reserved.

1. Introduction

The ever increasing use of information and communicationtechnologies has brought exponential growth in the volumes ofdigital data over the Internet. Consequently, the data within thesystems has started flooding at the rates that were never ob-served previously [1]. Moreover, besides huge data volumes, thetrends of multi-source data origination and management that areformidable in nature have engendered the concept of ‘‘Big Data’’

∗ Corresponding author. Tel.: +1 701 231 7615.E-mail addresses: assad.abbas@my.ndsu.edu (A. Abbas),

kashif.bilal@my.ndsu.edu (K. Bilal), limin.zhang@ndsu.edu (L. Zhang),samee.khan@gmail.com, samee.khan@ndsu.edu (S.U. Khan).

http://dx.doi.org/10.1016/j.future.2014.08.0100167-739X/© 2014 Elsevier B.V. All rights reserved.

[2]. Furthermore, traditional data management tools are limitedto handle such huge data volumes [1]. Therefore, the big datanecessitates the use of new and efficient techniques and technolo-gies to manage the data with multiple dimensions. Just like the bigdata prospects in e-commerce and other scientific and technolog-ical domains, the healthcare community is also witnessing hugehealthcare content being instigated from various points of care andWeb-based health communities [3,4]. Miller [1] categorizes themajor sources of health related big data into: (a) payer–providerbig data consisting of electronic health records, insurance records,pharmacy prescriptions, patient feedback, and responses, and (b)genomics-driven big data comprising of genotyping and sequenc-ing data. From the payers’ perspective, the datamatrixmay consistof hundreds of thousands of elements having various character-istics, such as demographics and medical treatment histories [1].

100 A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109

Consequently, the electronic medical information from dispersedlocations, such as clinics, hospitals, medical labs, financial organi-zations, and health insurance organizations has been integratedand the phenomenon has emerged e-health [5]. However, theexchange and integration of electronic medical information man-aged by several healthcare providers is costly and difficult to ad-minister that necessitates the use of cloud computing services inthe e-health domain [6]. The cloud computing paradigm exhibitstremendous potential to enhance the collaboration among varioushealthcare domains and to deal with the challenges, such as scala-bility, agility, cost effectiveness, and round the clock availability ofhealth related information [7].

The typical entities of a cloud based e-health system are: (a)healthcare providers, such as hospitals, clinics, pharmacies, labo-ratories, and (b) health insurance providers [8]. All of the afore-mentioned entities of the cloud based e-health systems are tightlyintegrated with each other to offer the convenient and on demandaccess over the electronic health data. For example, the relatedelectronic health records originated at hospitals and clinics aremade accessible to pharmacies and insurance companies for drugprescriptions and insurance claims processing, respectively. Cur-rently, the role of insurance providers in the contemporary healthcloud systems is only as a claim processing entity. However, withthe inception of the Patient Protection and Affordable Care Act(PPACA) commonly known as the Affordable Care Act, the healthinsurance providers may emerge as the major entity of the cloudbased e-health systems. The PPACA introduces the concepts of ‘‘In-suranceMarketplace and Health Insurance Exchanges’’ to facilitatethe individuals and small businesses to search the suitable healthinsurance plans [9,10]. More formally, the health insurance ex-change as defined by the US Department of Health andHuman Ser-vices helps the consumers and small businesses to buy insuranceplans by permitting easy comparisons of available plans based onthe price, coverage benefits, and quality [10]. Currently, there ex-ist various other Web based tools that are meant to search healthinsurance plans. However, such tools are deficient in providingrecommendations about the health insurance plans in accordancewith the multifaceted user requirements. The apparent reason forthe incompetence of the existing tools is their unawareness aboutthe diversified coverage requirements of the users. Moreover, thetools do not allow consumers to specify their coverage needs andinstead only acquire a few parameters, such as gender, age, and to-bacco use as input. Consequently, the users are returned with longlists of health insurance plans from different insurance providersirrespective of the fact that such recommendations may not sat-isfy the requirements of the users. Moreover, filtering such hugedata to find the desired information is an arduous task. Therefore,this is the high time for the development of health insurance planrecommendation systems with the capability to offer recommen-dations according to the diverse user coverage needs and financialconstraints. Obviously, such a task can be accomplished by com-paring the customer needs with the various health insurance plansto determine the most feasible plans.

In this regard, we focus on the aspect that has not been ad-dressed by the researchers in the near past. We argue that theexisting cloud based e-health services should be extended to of-fer knowledge based recommendations about health insuranceplans. Previously, a lot of research has been carried out on therecommendation systems to offer personalized recommendationsabout products, services, and locations. However, there is no rec-ommendation service that offers recommendations about healthinsurance plans based on the multifaceted requirements of theusers and consumers. Keeping in view the efficacy of deployingthe recommendation system for health insurance plans in thecontext of the PPACA, we leverage the use of cloud computingto offer recommendation services according to the user elicited re-quirements. Under the perspective of the PPACA, more and more

users will be looking for health plans being offered under the in-surance marketplace as well as by the private insurance providersin coming years. In addition, the health insurance providersare also expected to offer more plans considering the growth anddiversity in the user coverage and cost requirements. As a re-sult, the volumes of health data across the providers will intenselyincrease. Consequently, the demand for expensive InformationTechnology (IT) infrastructure will increase. Therefore, the cloudcomputing services seem quite practical to manage the hugedata volumes and to cut the costs [7]. The reason is that therequirements to purchase expensive infrastructure, such as thehigh performance computingmachines and storage are eliminatedwhen all the processing tasks are delegated to the cloud servicesproviders. The cloud computing paradigmenables the scalability orresizable compute capacity through the virtual machines [11,12].The services offered by the cloud computing are offered through anetwork while ensuring the Quality of Service (QoS) and are inex-pensive and on-demand [13]. The cloud users are charged for theuse of hardware and software resources [14–16].

We propose a cloud based requirements driven recommenda-tion framework for health insurance plans according to the tailoredrequirements of users. The rationale behind offering customizedinsurance plans is to effectively deal with the immense diversity ofthe health insurance coverage requirements among different cat-egories of users. For example, a user that belongs to a geographi-cal area where certain diseases are more common as compared toother regions may be more interested to have coverage for thosediseases. Likewise, individuals who interact with chemicals dur-ing their work hours are vulnerable to different diseases, such asskin problems and cancer. Consequently, such individualsmight beinterested in insurance plans that offer coverage for the aforesaidproblems.

We propose a user centered approach that offers a rich require-ment gathering interface to elicit user requirements for decisionmaking and insurance plan recommendation. The user centeredaspect of the proposed approach permits the users to specify re-quirements in terms of cost and coverage. As a result, the users areenabled to compare various health insurance plans based on thefulfillment of the criteria laid down by the users themselves. Weemployed an ontology based methodology to overcome the issuesof data heterogeneity across various health insurance providers.Each of the health insurance providers maintains a repository ofhealth insurance plans ontologies in an autonomous way with thefacility to add, remove, or update the ontology repositories. Consid-ering the large numbers of insurance plans by different providerswith heterogeneous data sources, we employ the concept of Dataas a Service (DaaS) [17]. The DaaS is an approach that is used toretrieve plans data from different providers for subsequent com-parisons with the user requirements. In the proposed framework,the users’ requirements are captured and transformed into the userontology. The plan ontologies maintained by each of the providersare retrieved based on the elicited user requirements using theDaaS. The ontologies retrieved are matched with the user require-ments and a similarity score is calculated. For true characteriza-tion of the effectiveness of the framework, we employ a rankingtechnique based on theMulti-attribute Utility Theory (MAUT). TheMAUT is an important analytical technique that aids in decisionanalysis by capturing the decision makers’ preferences based onmultiple independent objectives [18,19]. The proposed health in-surance plan recommendation system permits the users to specifythe preferred criteria or attributes, such as cost and coverage re-quirements over which the recommendation decisions should bebased. The preferred attributes are assignedweights based on theirrelative importance to the other attributes. We used the Rank Or-der Centroid Method (ROC) and the ratio method to test the ef-fectiveness of the plan ranking process. The experimental resultsdepict that the ROC method is more feasible in ranking the resultsas compared to the ratio method of weight assignment.

A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109 101

The salient contributions of the paper are:

• We present a user centered cloud based health insurancerecommendation framework to recommend a ranked list ofhealth insurance plans that best match with the user coveragerequirements and the indicated decision criteria or attributes.

• We propose a requirement gathering engine for user require-ments elicitation and for subsequent transformation into XMLschemas.

• A standard ontology representation/schema is proposed togive a standardized representation to all the plans so thatinformation about all the plans could be retrieved through theDaaS.

• A tree based matching algorithm is proposed to determine thestructural similarities between the users’ elicited requirementsand the insurance plans.

• A ranking strategy is proposed that ranks the health insuranceplans based on various user specified attributes or criteria interms of their relative importance using theMAUT. The decisionis based on the significance of attributes, such as: (a) premium,(b) co-pay, (c) deductibles, (d) co-insurance, (e) maximumbenefit, (f) providers network, and (g) fulfillment of essential,desirable, and optional requirements.

• The experiments are conducted on our locally administeredUbuntu cloud computing setup to determine the efficacy of theapproach. The interface to the cloud environment is providedby implementing the system as Software as a Service (SaaS).

The rest of the paper is organized as follows. Section 2 presentsthe preliminary concepts particular to the health insurancerecommendation and ontology. The architecture of the proposedplatform is presented in Section 3. The results and discussion arepresented in Section 4, whereas Section 5 discusses the relatedwork. Section 6 presents the conclusions and highlights the futureresearch work.

2. Preliminary concepts

Before a detailed discussion on the architecture of the proposedcloud based health insurance plan recommendation system, wepresent brief discussion on certain preliminary concepts. Thebackground and motivation of the proposed cloud based healthinsurance recommendation system are presented in Section 2.1.Section 2.2 presents discussion on the ontology for healthinsurance and the concept of DaaS.

2.1. Background and motivation

‘‘Big data as defined by a US congress report in August 2012 isa term used for describing large volumes of complex and variabledata with high velocities that entails sophisticated techniques tocapture, store, distribute, manage, and analyze the information[20]’’. Currently, the electronic health records coupled with theinnovative tools for big data analytics have opened new horizonsfor mining information to achieve highly effective outcomes[21,22]. The requirements, such as storage, processing, analysis,and continuous availability of enormous health data call forutilizing the emerging technologies, such as the cloud computing[23]. As already stated in Section 1 that currently there is huge in-flux and out-flux of health data in contemporary e-health systemsthat aremanaged by small andmedium sized health organizations.Moreover, the context of our proposed framework that emphasizeson shifting all the health data and the health insurance plans datain the e-health systems will significantly upraise the volumes ofhealth data. Furthermore, the PPACA also mandates the individualand families to have health insurance coverage. Therefore, it isneeded more than ever to offer the consumers such a mechanism

that helps them in selection of the best suited insurance plans interms of coverage and other aspects, such as the premium, co-pay,deductibles, co-insurance, the maximum benefit limit of the plan,and the providers’ network.

Currently, in the United States, the dataset about individu-als and family health insurance plans shortlisted as the qualifiedhealth plans under the insurance marketplace comprises of morethan 78,000 medical plans [24]. Similarly, for dental insurance,over 45,000 plans have also been identified in the insurance mar-ketplace [25]. The aforementioned numbers only depict the plansshortlisted as qualified health insurance plans. There could also beother plans that have not yet been certified under the insurancemarketplace. The above numbers are also expected to increase innear future when more and more consumers will start accessingthe insurance marketplace. Therefore, enormous increase in thehealth data in e-health systems is expected in near future. Conse-quently, the need for the development of sophisticated tools andtechniques for big data analytics in the healthcare domain has sig-nificantly increased. However, small and medium sized healthcareorganizations may face problems of resource scarcity in terms ofhardware, software, network services, and storage to manage suchhuge volumes of data and deliver round the clock access. Therefore,using the cloud computing services in the aforementioned scenariois quite pertinent because of the key benefits of the cloud, such asscalability [26] and elasticity [27] and pay per use model. Anotherkey benefit in embracing the cloud services is the significant re-duction in the infrastructure development and management cost.Therefore, the entities dealingwith the health related data can pro-cess the huge volumes of data with the sophisticated computingmachines at affordable prices.

2.2. Ontology for health insurance plans

Across the huge corpus of health insurance providers, all theproviders maintain their own datasets locally and possibly thedatasets may be heterogeneous in terms of terminology and struc-ture. The typical issues that may arise from the heterogeneousdata formats across different health insurance providers includethe integration and reconciliation of data originated frommultiplehealth insurance providers. Moreover, the heterogeneity besidesdata semantics is also immensely concerned about the structureand representation of data at the source locations. Consequently,a standardized representation is required to unify the distributeddata related tohealth insurance plans so that the information aboutall the providers and plans could be stored in a standard schema.Ontologies and the semantic web technologies offer the means topresent a standardized representation of distributeddata fromhet-erogeneous sources [28]. The semantic web is a particularly de-signed framework that promotes the development of mechanismsto share and utilize information from multiple resources in a dis-tributed architecture [29]. Ontology consists of vocabulary to de-scribe the particular view of a domain. As defined by Gruber [30],ontology is a specification of conceptualization. Ontology effec-tively deals with the problem of semantic heterogeneity and de-pending upon the preciseness of the specification, the concept ofontology encompasses various data and conceptual models, forexample classification, thesauri, and database schemas [31]. Weemploy ontology to offer a standardized representation of healthinsurance data at different providers’ locations with different for-mats. Besides insurance plans, the user queries indicating the cov-erage preferences and financial aspects are also transformed andrepresented in ontological form.

To query the heterogeneous health insurance plans reposito-ries, we employ the DaaS model [17]. More specifically, the DaaSis as an approach for data integration from different sources. In

102 A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109

Fig. 1. Generic ontology for health insurance plans.

the proposed framework, each of the providers maintains a repos-itory of plans. Based on the user elicited requirements, the SaaSbased system requests the plan data using theDaaS. TheDaaS com-bines the data from multiple providers and offers a standardizedrepresentation to plans data to find the match between the userrequirements and plans. Despite using the third-party cloud infras-tructure, our framework permits the providers to exercise their au-tonomous control over their data because the plans are updated orremoved by the providers themselves. Therefore, apparently no is-sues pertaining to the security and privacy of the providers’ dataarise. Fig. 1 presents a generic ontology for the health insuranceplans over that all the ontologies can bemapped. Due to space lim-itations all the levels of the ontology are not presented in Fig. 1.

To copewith the heterogeneity issues of datasets across varioushealth insurance providers we used XML schemas. Althoughontology and schema refer to different levels of abstractionin representation, when both are applied to online sources ofinformation the relationship becomes obvious [32]. The structureand vocabulary for describing the semantics of information presentin documents is provided by ontology, whereas XML schemas areused to prescribe the structure and contents of the documents [32].The XML documents can be represented in the form of labeledtrees. An XML tree allows a whole document to be representedas a root node. The non-terminal or internal nodes represent theelements whereas the contents are represented at leaf nodes [33].

3. Proposed system architecture for health insurance recom-mendation system

The proposed architecture to manage the massive healthinsurance data across hundreds of providers with thousands ofinsurance plans consists of the following modules: (a) insuranceplans ontology managed and offered by the insurance provider,and delivered as the DaaS, (b) user requirement gathering module,(c) matching module, and (d) ranking module. In the proposedcloud based recommendation framework, each insurance providermaintains its own plan repository autonomously and offers therequired information to the system as the DaaS on demand.The SaaS based implementation permits the user requirementgathering module to elicit the requirements from the users

and transforms the delivered information into ontology thatsubsequently is captured as an XML schema. The XML schemasrepresent the information in hierarchical fashion. Therefore, thecommon representation of the XML documents is in the form oflabeled trees [34]. The user requirements and all the plans residingat the providers’ location are represented as the trees. In thetraditional Document Object Model (DOM) the nodes symbolizethe XML elements, whereas the children represent the attributes[35]. The matching module matches the user requirements treewith multiple plans trees to determine the structural similarities.The structural similarities between the user requirement tree andthe plans tree are computed by comparing the labels or tags whilepreserving the parent–child relationship. The matching moduleonly provides the match details between the user requirementstree and the plan trees. Therefore, to make the recommendationprocess more effective, the MAUT based approach is used to rankthe plans according to the criteria laid down by the users. TheMAUT is an important phenomenon used in decision theory basedon Multi-criteria Decision Making [36]. In the MAUT, the decisionsare made in such a way that the utility function based on theattributes or criteria is maximized [37]. The utility of each of thealternatives can be calculated by the decision makers through amulti-attribute utility function and the function with the highestutility value is selected [38]. In the proposed work, the MAUTuses nine attributes to help users evaluate the recommended plansbased on their ranking scores. Fig. 2 presents the architecture ofthe proposed system. The notations used throughout the text arepresented in Table 1.

3.1. The matching module

The matching module matches the user requirements withmultiple plans to determine the similarities. In the proposedframework, both the user requirements and the insurance plansare represented in the form of trees. To describe the problem oftree matching in our scenario, we first present some preliminaryconcepts related to the rooted labeled trees.

A labeled tree is defined as a tree T = (N, E, r, ∂), where N =

{n1, n2, . . . , nK } is a finite set of nodes of the tree and E =

{e1, e2, . . . , ek} is a set of edges between the nodes of the labeled

A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109 103

Table 1Notations used and their meanings.

Notation Meaning Notation Meaning Notation Meaning

R Requirements tree e Essential match γ Sum of matching and non-matching nodesP Plans tree d Desirable match ∪RC Union of R and CRe Essential requirements o Optional match δsi Requirements similarityRd Desirable requirements ℵe Essential non-match Pn Providers NetworkRo Optional requirement ℵd Desirable non-match ρ Actual value requested by the userPr Requested premium ℵo Optional non-match µ−e Weight of the missing attributeDr Requested deductibles Wi Weight of ith attribute e Essential matchCPr Requested copay W norm

i Normalized weight d Desirable matchCIr Requested coinsurance δri Requirements satisfiability measure o Optional matchMBr Maximum benefit µ Desired attribute value requested by the user ρ Actual value requested by the user

Matching nodes ℵ Non-matching nodes ∂ Labeling function

Fig. 2. Cloud based health insurance recommendation system architecture.

tree. The root node is represented as r and ∂ represents a labelingfunction to map each node to a set of labels {L = l1, l2, . . . , lK }.

In the text to follow, R and P be the trees representing the userrequirements and insurance plan, respectively. Moreover, eachsingle plan in P is represented as Pk. The tree matching problemis to find an exact mapping while preserving the ancestry. For anexact matching, if the label of node in R matches the label of nodein P at the corresponding level only then the descendants of thenode in R will be matched to descendants of node in P [33].

In our approach, while eliciting the insurance requirements,the users also indicate three types of coverage requirementsnamely: (a) Essential Requirements, (b) Desirable Requirements,and (c) Optional Requirements. The set R = {Re, Rd, Ro} is a setwhere the essential requirements are represented by Re, desirablerequirements are represented byRd, and the optional requirementsare represented by Ro. For each Ri ∈ R, different weight is assignedto observe the effect of a match or non-match on the overallsimilarity value. The set C = {Pr ,Dr , CPr , CIr ,MBr} represents thecustomer requirements in terms of cost. The Pr ,Dr , CPr , CIr , Pnrepresent the amount in terms of premium, deductibles, co-pay,and co-insurance, respectively, whereas MBr is the maximumbenefit that a user expects from a plan. The variable Pn representsthe providers’ network that users may opt as their healthcareproviders. Providers’ network is an important qualitymeasure thatbecomes more critical in presence of multiple plans with similarfeatures. The algorithm to match the user requirements tree withthe plan tree is presented below.

The user requirement tree R and the plan tree P are provided asinput to the algorithm in line 1. Lines 2 and 3 initialize the variablesused to calculate the total number of matching and non-matching

nodes, respectively. From lines 4–10, the algorithm matches thelabel of the node in R with the node at the same level in Pk whilepreserving the ancestry. If a match is found is incremented atline 9 and the procedureMatchTree() is recursively called at line 14to find the matches between the sub-trees of R and P. If the labelsof R and P at subsequent levels do not match, it means that theirsub-trees are not matched and the total number of non-matchingchildren in the tree R is calculated at line 16.

We explain thematching processwith the help of an illustrativeexample. Fig. 3 represents the requirement tree (�) on left side andthe plan tree (�k) on the right side. For the sake of simplicity, allthe nodes belonging to both the trees are not presented in Fig. 3.The tree matching algorithm is applied recursively to performmatching of the corresponding nodes by comparing labels of thetrees � and �k. If the nodes in both the trees have the samelabel, then the sub-trees of � and �k are compared. A match isconsidered, if any of the child nodes of a matched parent matcheswith the requirement tree label at the same level. If the labelsof the roots of two sub-trees do not match, the algorithm doesnot compare the subsequent levels of the mismatching nodes. Forexample, the node c with the label ‘‘outpatient coverage’’ in R doesnot match with the corresponding level in Pk that has ‘‘InpatientCoverage’’ and ‘‘Minor event coverage’’ at the same level under thesame parent. Therefore, the subsequent levels of node c will not becompared. The matching algorithm requires the nodes to be at thesame level and should be decedents of the parents with the samelabels in bothR and Pk. However, in both trees the two nodes beingmatched should not be necessarily in the same order. As can be

104 A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109

Fig. 3. An illustrative example for tree matching.

observed in Fig. 3 that node j at level 4 inR has label ‘‘Herpes Zoster ’’whereas in Pk the corresponding node m has label ‘‘Hepatitis C’’.However, node l in Pk has label ‘‘Herpes Zoster ’’ under the sameparent (‘‘Routine physical exam’’) as in the requirement tree. Thematching algorithm exhaustively compares the label of a node inthe� to all the nodes at the same level in Pk under the same parentand finds a match for the label ‘‘Herpes Zoster ’’. The ‘‘✓’’ and ‘‘×’’symbols in Fig. 3 represent the matching and non-matching node.The similarity between the two trees is calculated as below:

δsi =Mi

Ti, (1)

where,

Mi = e + d + o, (2)

Ni = ℵe + ℵd + ℵo, (3)Ti = Mi + Ni. (4)

In Eq. (1), Mi and Ti represent the number of matching require-ments, and the total requested requirements in the user query. Theδsi is the requirement satisfiability measure that can have maxi-mum value of 1. The measure represents the percentage of the re-quirements that are met by a plan. If all the requirements statedby the users are met, the measure would have themaximum valueof 1.

3.2. Plan ranking using the MAUT

The matching module only calculates the similarities betweenthe user requirements and the stored plans. However, consideringthe diverse user requirements, in terms of cost and coverage, thereis a need to provide users a ranked list of plans to make effectivechoice of insurance plan. The framework allows the users to specifythe relative importance or priorities of various decision attributes.Ranking in the proposed scenario is imperative because it helpsusers to evaluate several plans by altering the relative importanceof attributes to find the best suitable plan.

We utilize the MAUT approach for ranking the plans that aresimilar to the customer coverage requirements as well as the

cost requirements. The MAUT involves the customers in decisionmaking. While stating the coverage needs, the users also indicatethe relative importance of ranking criteria or attributes from boththe sets C and S as well as for Pn. The purpose of using the relativeimportance is to determine that exactly what attributes should begiven higher weights during the ranking process. The higher therelative importance of the particular criteria, the higher weight itis assigned as compared to the others. Consequently, the rankingdecisions are biased towards the criteria with higher priorities.

3.2.1. Attribute weight assignmentA key task in ranking the plans to select the best alternatives

using the MAUT approach is weight assignment to attributes. Theproposed approach is user centered that allows the users to specifythe relative importance of decision attributes in relation to otherattributes. The attributes that are given higher relative importanceby the users while specifying requirements are assigned higherweights during the ranking process. We employ two methods forweight assignment, namely: (a) Rank Order Centroid (ROC) and(b) ratio method. Both of the weight assignment methods aredescribed below.

3.2.1.1. Weight assignment using the Rank Order Centroid (ROC)method. To rank the identified health insurance plans we usedthe Rank Order Centroid (ROC) method [39] to assign weights tothe users’ specified criteria or attributes. The ROC method assignsweights to a number of attributes that are ranked according torelative importance [40]. Combining sets R and C with Pn we geta complete set of requirements called ∪RC as below:

∪RC = {Re, P r , Rd, Ro, Dr , CPr , CIr ,MBr , Pn} .

The order of elements in the set ∪RC = {Re, P r , Rd, Ro, Dr ,CPr , CIr ,MBr , Pn} indicates the relative importance of require-ment to the user. For instance, Re has highest relative importanceas compared to the remaining eight attributes in the set ∪RC . Sim-ilarly, the attribute Pr is the attribute with the second highestpriority. The plan ranking decisions largely depend upon the im-portance of the attributes to the consumers or users. The proposed

A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109 105

Fig. 4. User requirement specification interface.

approach allows the users to test the ranking alternatives by vary-ing the relative importance of different attributes (see Fig. 4, whereuser can change the order of the criteria elements). The weights ofthe attributes are calculated using the following equation for theROC:

Wi =

1K

kn=i

1n, (5)

where k is the number of attributes and Wi represents the weightof the ith attribute to be ranked using the ROC.

3.2.1.2. Weight assignment using the ratio method. The ratiomethod proposed by Edwards [41] is another method to assignweights to the attributes for ranking decisions. Like the ROCmethod, the decision attributes are ranked in the order of relativeimportance. The weights are assigned as multiples of 10 andthe attributes with the lowest importance is assigned weight 10.Typically, the weights to the attributes are assigned at a jump of10. However, assigning weights more than the prescribed jumpis usually based on the subjective judgments that sometimes maylead to higher normalizedweights. The normalizedweights of eachof the attributes are calculated as follows:

W normi =

wik

j=1wj

, (6)

where k is the number of attributes being used in the decision.The weight assignment procedure is elaborated with the examplebelow. The following raw weights are assigned to each of theelements of set ∪RC = {Pr , Re, Rd,Dr , Pn, CPr , CIr ,MBr , Ro, Pn} =

{90, 80, 70, 60, 50, 40, 30, 20, 10}. The normalized weight for theattribute Pr using Eq. (6) is calculated as below.

Pr = 90/(90+ 80+ 70+ 60+ 50+ 40+ 30+ 20+ 10) = 0.2.The weights for the other attributes are calculated similarly.

The final ranking of a particular plan is computed by using theattribute function R as below:

R =

(Wi × δri), (7)

where Wi represents the weights of the attributes calculatedthrough either the ROC or ratio method and δri is a measure usedto determine the satisfiability of particular set of requirements. Theranking score of a plan is calculated by multiplying the weights ofeach element of∪RC to the satisfiability value of each attribute. Themeasure δri is calculated as:

δri =µ

ρ, (8)

where µ and ρ are the desired values requested by the user andthe actual value of a particular attribute present in the plan, re-spectively. For example, if the user requests a plan with monthlypremium of $150 whereas the actual premium of the plan beingoffered by the insurance provider is 175, then the value of the sat-isfiability measure will be 0.86. If δri = 1, then the particular cri-teria has the highest satisfiability and is consequently assigned thehighestweight. If δri > 1, themaximumvalue of δri is still regardedequal to 1. Our approach also permits users to reduce the elementsof ∪RC for making ranking decisions. Moreover, in set R, if any ofthe elements is not indicated in the user query, the weights areadjusted by distributing the weights among the remaining two el-ements. The weights are adjusted using the adjustment factor as:

α =

Kn=i

WRi + µ−e × Ri

, (9)

whereµ−e is the weight of themissing element of R andWRi refersto the same weights that we used in Eq. (7). However, if two ele-ments of set R are missing then all the entire weight is assigned tothe remaining one element. Adjusting weights is beneficial in thesense that the ranking decisions are not substantially biased to-wards one or two elements of R. Therefore, substantially fair rank-ing outcomes are achieved.

106 A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109

Table 2Importance of attributes in the test runs.

Attribute importance T1 T2 T3 T4 T5 T6 T7

1 ER ER PR PR PR PR ER2 DR PR ER ER ER ER PR3 OR DR DR DD DD CP CP4 PR OR CP CP CP DD DD5 DD DD DD DR CI CI MB6 CP CP OR MB MB PN PN7 CI CI PN OR OR OR DR8 MB PN CI PN DR DR OR9 PN MB MB CI PN MB CI

Essential Requirements: ER, Desirable Requirements: DR, Optional Requirements:OR, Premium: PR, Co-pay: CP, Co-insurance: CI,Max. Benefit:M, ProvidersNetwork:PN.

3.3. Prototype implementation

We implemented a prototype system that provides users an in-terface to the cloud environment. A requirement engine is usedto help users specify their coverage needs and cost expectations,and the prioritized criteria for decision making. The framework isimplemented as SaaS using modular service oriented architecture.In the SaaS architecture, the software is hosted as a service thatis provided to customers via the aforementioned interface acrossthe Internet [42]. The SaaS can considerably reduce the customersIT cost and meets the flexible business requirement, especially forbusiness management services. One common feature of the SaaSbusiness services is that the customers’ business data are storedand processed at the service provider side [43]. The SaaSmodel re-lieves the users or organizations using cloud services of the tasksof installation and maintaining the software. Instead the users paythe cloud service providers for the services. In the proposed frame-work, the users access the cloud services through a Web interfacemodule. The interface module collects the requirements informa-tion from the users. The collected information is directed to thecloud based framework. Subsequently, the user requirements in-formation is transformed into XML based ontology for compari-son with the insurance plans. All the insurance providers maintainthe cloud based ontology repository. The plans from the respectiveontology repositories are extracted based on the user requirementsusing the DaaS. On receiving the plan ontologies, the user require-ments are matched with the plan ontologies to determine thesimilarity. However, the similarity matching is not the true char-acterization of the effectiveness of a plan to the users becausematching does not take into account the cost criteria. Therefore, the

ranking module ranks the matched plans according to the crite-ria specified by the user. The experiments were conducted on ourlocally administered Ubuntu cloud computing setup running on96 core Supermicro SuperServer SYS-7047GR-TRF systems. Fig. 4shows the screenshot of the user requirement capturing module.

4. Results and discussion

To test the validity of the system, we used real data of healthinsurance plans that were shortlisted as qualified health plans un-der the insurance marketplace released by the health department[25]. The data comprises of more than 78,000 different individualand family health insurance plans and over 45,000 dental plans.However, the data was not properly organized and therefore wasnot directly usable. Consequently, we used our system to createhealth insurance plans by using the aforesaid data. Therefore, wecurrently chose a subset of plans to test the system. The informa-tion depicted in the planswas transformedmanually by keying thedata to our system. All the generated insurance plans were storedas XML schemas. Around one hundred plans were created to testthe system performance. A user study was conducted to test theeffectiveness of recommendations provided by the system. Duringthe user study the userswere guided about the procedures of inter-acting with the interface. The users were also asked to conduct thetest runs by changing the importance level of desired attributes aslisted in Table 2. The underlying reason behind providing the userswith the flexibility to test the ranking results with different prior-itized criteria was to observe the variations in the ranked results.Tomake the ranking processmore explicit different prioritieswereassigned to the attributes during different tests for the selection ofhealth insurance plans. The weights were assigned using the ROCand ratio methods. Both of the methods were tested on the sameset of requirements to determine the effects of weight changes onthe overall decision quality and plan recommendation. The column‘‘Attributes Importance’’ in Table 1 depicts the relative importanceof the attributes during the seven test runs, namely, T1–T7. Theuser attributes are abbreviated as below in Table 2.

The weight assignment by the ROCmethod and the normalizedweight assignment by the ratio method are presented in Table 3.

The ranking scores obtained for different plans using the ROCand ratio methods for weight assignment are presented in Table 4and Table 5, respectively. As can be observed from Tables 4 and5, that altering the priority and relative importance of attributesresulted in different ranking score for the same plan in differenttests. For example, in Table 2 during the test T1 the attribute

Table 3Weight assignment using the ROC and the ratio method.

Weight assignment method Attributes1 2 3 4 5 6 7 8 9

ROC 0.31 0.20 0.14 0.12 0.097 0.078 0.063 0.048 0.036Ratio 0.2 0.18 0.16 0.13 0.11 0.089 0.067 0.044 0.022

Table 4Plan ranking using the ROC.

TestNo.

Plan name

AK Aetna Classic 5000PD

Be ConnectedBronze

BlueDirect 704000

Humana National Preferred Bronze 4850/6350

Premera Preferred Plus Bronze HSA 5250

T1 0.75 0.65 0.68 0.63 0.71T2 0.74 0.69 0.72 0.67 0.75T3 0.75 0.74 0.79 0.71 0.81T4 0.79 0.80 0.83 0.77 0.85T5 0.78 0.81 0.87 0.80 0.89T6 0.72 0.75 0.83 0.73 0.85T7 0.78 0.74 0.79 0.70 0.82

A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109 107

Table 5Plan ranking using the ratio method.

TestNo.

Plan name

AK Aetna Classic 5000PD

BeConnectedBronze

BlueDirect 704000

Humana National Preferred Bronze 4850/6350

Premera Preferred Plus Bronze HSA 5250

T1 0.71 0.65 0.70 0.64 0.72T2 0.70 0.66 0.71 0.65 0.73T3 0.72 0.67 0.74 0.65 0.75T4 0.77 0.74 0.78 0.71 0.80T5 0.76 0.76 0.83 0.67 0.79T6 0.74 0.78 0.77 0.79 0.76T7 0.74 0.72 0.77 0.68 0.78

Fig. 5. Plan ranking using the ROC method for weight assignment.

Fig. 6. Plan ranking using the ratio method for weight assignment.

‘‘ER’’ was assigned the highest importance while the ‘‘DR’’ wasat the second highest importance level. Therefore, they wererespectively assigned the highest and second highest weights bythe weight assignment methods and consequently, the plan AKAetna Classic 5000 (AKC5) PD had the highest rank value. In testT2, the importance level was altered and the attribute ‘‘PR’’ wasassigned the second highest importance and the attribute ‘‘DR’’was at importance level 3 and the plan Premera Preferred PlusBronze HAS 5250 turned out with the highest ranking score. Withthe ratio method, in test T1 with the same user requirements theplans Premera Preferred Plus Bronze HAS 5250 and AK Aetna Classic5000 PD had ranking scores of 0.72 and 0.71 respectively. However,later from test T3 to test T7, changing the relative importance ofdecision attributes resulted in more significant differences among

the ranking scores of different plans. Figs. 5 and 6 present theranking scores for five plans during the seven conducted tests.Another important observation is pertaining to the performanceof the two weight assignment methods with each other. As can beobserved from Tables 4 and 5 that the ranking score achieved usingthe ROCmethodwas slightly higher as compared to those obtainedusing the ratio method. The reason is that the weight assignmentin the ROC method is dependent on the number of attributes orcriteria for making a decision. Since, there are nine attributes;therefore, the weights of the attributes with high importance aremuch dispersed, while the attributes with the lowest importanceare assigned very small weights. Alternatively, the normalizedweights for the ratiomethodwere obtained bymanually specifyingthe initial weights for all of the attributes. Before normalizing, the

108 A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109

rawweights assigned to attributes with the highest and the lowestimportance were 90 and 10, respectively. However, increasing thehighest raw weight value may result in an increased normalizedvalue. The reason is that the weight assignments in ratio method isbased on the strong splitting bias that eventually results in higherranking score of the alternatives. Consequently, the higher rawweights in ratio method could result in higher ranking score forthe plans while with the lower rawweights the differences amongthe ranking score are very slight. Therefore, presumably we canclaim that the ranking results obtained through the ROC weightassignment method were more balanced as compared to the ratiomethod.

5. Related work

5.1. Ontology and XML representation

Over the past few years, various approaches have been pro-posed for deploying the electronic health data in the cloud plat-form due to the ever increasing volumes of the health data, suchas patient electronic medical records, lab reports, and insuranceclaims. Moreover, to efficiently process and integrate geographi-cally dispersed health data, there are also several proposals. Wepresent an ontology based approach for a standardized represen-tation of the health plans across multiple healthcare providers.Ontology based approaches in distributed environment have beenused in various proposals. An ontology based approach to dealwiththe emergency management that unifies the datasets distributedacross various locations is presented in [28]. The approach is ca-pable of mapping the XML schemas to ontology. An XML retrievalapproach using tree matching is presented in [44]. There are vari-ous tree matching algorithms, for example the exact matching andapproximate matching algorithm to determine the structural sim-ilarity among the XML documents. The exact matching algorithmsused in Refs. [45,46] employ sequential tree matching approachesthat first apply query decomposition process and the query twigis transformed into paths from root to leaf. Also there are vari-eties of approaches that have been used for approximate XML treematching. However, contrary to exact tree matching approachesthese approaches are designed to rank and select elements withrespect to their probability of matching the queries. In Ref. [44], anapproach that uses edge relaxation for indexing XML documentsis presented and weighs the parent–child relationships accordingto a maximal score of 1. We use the exact tree matching algo-rithm to determine the number of matching and non-matching re-quirements to calculate the structural similarity among the trees.Moreover, the user requirements are categorized as ‘‘Essential’’,‘‘Desirable’’, and ‘‘Optional’’. The ‘‘Essential’’ requirements are as-signed higher weights whereas the ‘‘Optional’’ requirements areassigned the lowest weight in the interval [0, 1]. The weights ofthe ‘‘Desirable’’ requirements are in between the ‘‘Essential’’ and‘‘Optional’’ requirements.

5.2. Multi-attribute utility theory

Apart from tree matching aspect, another important dimensionof our proposed work is decision support while ranking the healthinsurance plans. The ranking is performed based on the weightsof the attributes. The MAUT is an important analytical tool fordecision analysis that captures the decision makers’ preferencesto make decisions based on multiple independent objectives [18].The decisionmakers’MAUT functions aremodeled using the utilityelicitation methods. The MAU function can be determined byemploying holistic or decomposed approaches [47]. The holisticapproaches, such as multiple regression analysis and artificialneural networks require a decision maker to evaluate all the

alternatives. On the other hand, the decomposed approaches,such as Simple Multi-Attribute Rating Technique (SMART) [48]and Analytic Hierarchy Process (AHP) require the decision makerto compare the relative importance of various attributes. Huang[18] used the SMART to rank user preferences in terms of theirimportance. The approach uses the ROC to assign weights tothe attributes. Our approach for eliciting the weights of variousattributes uses the ROC and the ratio method. Moreover, there arealso several AHPbasedproposals for recommendation anddecisionmaking based on multiple attributes, such as [49–52]. However,the SMART exhibits better performance as compared to the AHPwhen the decisions to be made are complex enough. In addition,the AHP method compares every two alternatives based on eachsingle attribute that makes it less suitable when there are largenumbers of alternatives.

6. Conclusions and future work

In this paper, we presented a cloud based recommendationsystem for health insurance plans based on the user specifiedcriteria and priorities. Testing the framework at a limited leveldepicts that the proposed framework is highly effective inoffering customized recommendations about health insuranceplans. Particularly, the flexibility to test the insurance plans byaltering the priorities of different attributes is certainly a beneficialfeature that allows comparison among various plans based onmultiple criteria. It is also expected that in near future, the researchon health insurance recommendation systems will also increasein context of the PPACA when more users will start accessing theinsurance marketplace. Therefore, the need for development oftechniques andmethods for big data in the healthcare domain willsignificantly increase.

We intend to extend the current work by deploying machinelearning techniques to predict the best suited health insuranceproducts to new users by considering the existing users’ charac-teristics. An important capability of machine learning techniquesis that they improve with time and learn from both the success-ful and unsuccessful recommendations that eventually results inhigher recommendation accuracy. We also aim to standardize thehealth insurance plan representation by proposing a generic on-tology to capture and represent the entire information particularto the health insurance terms. The standardization will help theresearchers to use the information in more effective way.

References

[1] K. Miller, Big data analytics in biomedical research, Biomed. Comput.Rev. (2014) http://biomedicalcomputationreview.org/content/big-data-analyticsbiomedicalresearch, 2013 (accessed 8.01.13).

[2] J. Bughin, M. Chui, J. Manyika, Clouds, big data, and smart assets: Ten tech-enabled business trends to watch, McKinsey Q. 56 (1) (2010) 75–86.

[3] H. Chen, R.H.L. Chiang, V.C. Storey, Business intelligence and analytics: frombig data to big impact, MIS Q. 36 (4) (2012) 1165–1188.

[4] M. Cottle, W. Hoover, S. Kanwal, M. Kohn, T. Strome, N.W. Trelster,Transforming healthcare through big data, http://assets.fiercemarkets.com/public/newsletter/fiercehealthit/iht2bigdata.pdf, 2014 (accessed on 15.01.14).

[5] D. Slamanig, C. Stingl, Privacy aspects of e-Health, in: 3rd IEEE internationalConference on Availability, Reliability and Security, (ARES’08), March 2008,pp. 1226–1233.

[6] R. Zhang, L. Liu, Security models and requirements for healthcare applicationclouds, 3rd IEEE International Conference onCloudComputing,Miami, FL, USA,July 2010, pp.268–275.

[7] L. Wang, D. Chen, Y. Hu, Y. Ma, J. Wang, Towards enabling cyber infrastructureas a service in clouds, Comput. Electr. Eng. 39 (1) (2013) 3–14.

[8] A. Abbas, S.U. Khan, A review on the state-of-the-art privacy preservingapproaches in e-health clouds, IEEE J. Biomed. Health Inform. 18 (4) (2014)1431–1441.

[9] Insurance Marketplace, http://www.hhs.gov/healthcare/insurance/index.html.

[10] S. Haeder, D.L. Weimer, You can’t make me do it: state implementation ofinsurance exchanges under the affordable care act, Publ. Adm. Rev. (2013)S34–S47.

A. Abbas et al. / Future Generation Computer Systems 43–44 (2015) 99–109 109

[11] L. Wang, G.V. Laszewski, M. Kunze, J. Tao, J. Dayal, Provide virtual distributedenvironments for grid computing on demand, Adv. Eng. Softw. 41 (2) (2010)213–219.

[12] L. Wang, G.V. Laszewski, D. Chen, J. Tao, M. Kunze, Provide virtual machineinformation for grid computing, IEEE Trans. Syst. Man Cybern. A 40 (6) (2010)1362–1374.

[13] L. Wang, M. Kunze, J. Tao, G.V. Laszewski, Towards building a cloud forscientific applications, Adv. Eng. Softw. 42 (9) (2011) 714–722.

[14] L. Wang, J. Tao, R. Ranjan, H. Marten, A. Streit, J. Chen, D. Chen, G-Hadoop:MapReduce across distributed data centers for data-intensive computing,Future Gener. Comput. Syst. 29 (3) (2013) 739–750.

[15] L. Wang, G.V. Laszewski, A. Younge, X. He, M. Kunze, J. Tao, C. Fu, Cloudcomputing: a perspective study, New Gener. Comput. 28 (2) (2010) 137–146.

[16] A.N. Khan, M.L.M. Kiah, S.A. Madani, M. Ali, Enhanced dynamic credentialgeneration scheme for protection of user identity inmobile–cloud computing,J. Supercomput. 66 (3) (2013) 1687–1706.

[17] R. Mokadem, F. Morvan, C.G. Guegan, D. Benslimane, DSD: A DaaS servicediscovery method in P2P environments, in: New Trends in Databases andInformation Systems, Springer International Publishing, 2014, pp. 129–137.

[18] S.-L. Huang, Designing utility-based recommender systems for e-commerce:Evaluation of preference-elicitation methods, Electr. Commer. Res. Appl. 10(4) (2011) 398–407.

[19] R.K. Sarin, Multi-attribute utility theory, in: Encyclopedia of OperationsResearch and Management Science, 2013, pp. 1004–1006.

[20] TechAmerica Foundation Big Data Commission, http://www.techamericafoundation.org/bigdata, 2014 (accessed 10.01.14).

[21] P.B. Jensen, L.J. Jensen, S. Brunak, Mining electronic health records: towardsbetter research applications and clinical care, Nature Rev. Genet. 13 (6) (2012)395–405.

[22] A. McAfee, E. Brynjolfsson, Big data: the management revolution, Harv. Bus.Rev. 90 (10) (2012) 60–66.

[23] S.P. Ahuja, S. Mani, J. Zambrano, A survey of the state of cloud computing inhealthcare, Netw. Commun. Technol. 1 (2) (2012) 12–19.

[24] QHP landscape individual market, https://data.healthcare.gov/dataset/QHP-Landscape-Individual-Market-Medical/b8in-sz6k, 2013(accessed on 20.12.13).

[25] Dental plan information for individuals and families, https://www.healthcare.gov/dental-plan-information/, 2014 (accessed on 11.01.14).

[26] A. Michael, A. Fox, R. Griffith, A.D. Joseph, R. Katz, A. Konwinski, G. Lee, A viewof cloud computing, Commun. ACM 53 (4) (2010) 50–58.

[27] I. Foster, Y. Zhao, I. Raicu, S. Lu, Cloud computing and grid computing 360-degree compared, in: IEEE Grid Computing EnvironmentsWorkshop (GCE’08),2008. pp. 1-10.

[28] J. Li, Q. Li, C. Liu, S.U. Khan, N. Ghani, Community-based collaborativeinformation system for emergencymanagement, Comput. Oper. Res. 42 (2012)116–124.

[29] N. Shadbolt, W. Hall, T.B. Lee, The semantic web revisited, IEEE Intell. Syst. 21(3) (2006) 96–101.

[30] T.R. Gruber, Toward principles for the design of ontologies used for knowledgesharing, Int. J. Hum.-Comput. Stud. 43 (5) (1995) 907–928.

[31] P. Shvaiko, J. Euzenat, Ontology matching: state of the art and futurechallenges, IEEE Trans. Knowl. Data Eng. (2012) 158–176.

[32] M. Klein, D. Fensel, F.V. Harmelen, I. Horrocks, The relation between ontologiesand XML schemas, Electron. Trans. Artif. Intell. (2001) 1–14.

[33] M.A. Tahraoui, K.P. Sauvagnat, C. Laitang, M. Boughanem, H. Kheddouci, L.Ning, A survey on tree matching and XML retrieval, Comput. Sci. Rev. (2013)1–23.

[34] J. Tekli, R. Chbeir, A novel XML document structure comparison frameworkbased-on sub-tree commonalities and label semantics,Web Semant.: Sci. Serv.Agents World Wide Web 11 (2012) 14–40.

[35] Document Object Model, http://www.w3.org/DOM, 2014 (accessed on16.01.14).

[36] S.D. Pohekar, M. Ramachandran, Application of multi-criteria decisionmakingto sustainable energy planning—a review, Renew. Sustain. Energy Rev. 8 (4)(2004) 365–381.

[37] D. Claudio, G.O. Kremer, W.B. Llerena, A. Freivalds, A dynamic multi-attributeutility theory based decision support system for patient prioritization in theemergency department, IIE Trans. Healthc. Syst. Eng. (2014) 1–15.

[38] Y.-S. Huang, W.-C. Chang, W.-H. Li, Z.-L. Lin, Aggregation of utility-basedindividual preferences for group decision-making, European J. Oper. Res. 229(2) (2013) 462–469.

[39] T. Solymosi, J. Dombi, A method for determining the weights of criteria: thecentralized weights, European J. Oper. Res. 26 (1) (1986) 35–41.

[40] B.S. Ahn, K.S. Park, Comparing methods for multi-attribute decision makingwith ordinal weights, Comput. Oper. Res. 35 (5) (2008) 1660–1670.

[41] W. Edwards, How to use multi-attribute utility measurement for socialdecision making, IEEE Trans. Syst. Man Cybern. 7 (5) (1977) 326–340.

[42] J.Y. Lee, J.W. Lee, S.D. Kim, A quality model for evaluating software-as-a-service in cloud computing, in: 7th ACIS International Conference on SoftwareEngineering Research, Management and Applications, 2009, pp. 261-266.

[43] W.-T. Tsai, X. Sun, J. Balasooriya, Service-oriented cloud computing architec-ture, in: 2010 Seventh International Conference on Information Technology:New Generations (ITNG), pp. 684-689.

[44] M.B. Aouicha, M. Tmar, M. Boughanem, M. Abid, XML information retrievalbased on tree matching, in: IEEE International Conference on Engineering ofComputer Based Systems, ECBS, Belfast, Ireland, 2008, pp. 499–500.

[45] P. Zezula, F. Mandreoli, R. Martoglia, Tree signatures and unordered XMLpattern matching, in: 30th Conference on Current Trends in Theory andPractice of Computer Science, Merin, Czech Republic, 2004, pp. 122–139.

[46] P. Zezula, G. Amato, F. Debole, F. Rabitti, Tree signatures for XML querying andnavigation, in: Database and XML Technologies, 2003, pp. 149–163.

[47] J.-C. Pomerol, S.B. Romero, Multicriterion Decision in Management: Principlesand Practice, Kluwer Academic Publishers, Boston, 2000.

[48] W. Edwards, H.F. Barron, SMARTS and SMARTER: improved simple methodsfor multi-attribute utility measurement, Org. Behav. Hum. Decis. Process. 60(3) (1994) 306–325.

[49] C. Schmitt, D. Dengler, M. Bauer, The MAUT machine: an adaptive recom-mender system, in: Proceedings of the ABIS Workshop, Hannover, Germany,2002.

[50] M.F. Frimpon, A multi-criteria decision analytic model to determine the bestcandidate for executive leadership, J. Polit. Law 6 (1) (2013) 111–127.

[51] D.-R. Liu, Y.-Y. Shih, Integrating AHP and data mining for product recommen-dation based on customer lifetime value, Inf. Manage. 42 (3) (2005) 387–400.

[52] Z. Hua, B. Gong, X. Xu, A DS–AHP approach for multi-attribute decisionmaking problemwith incomplete information, Expert Syst. Appl. 34 (3) (2008)2221–2227.

Assad Abbas completed Master of Science in Informaticsfrom University of Skovde, Sweden in 2010. He is a Ph.D.candidate at the Department of Electrical and ComputerEngineering, North Dakota State University, USA. Heis affiliated with COMSATS Institute of InformationTechnology, Pakistan since 2004. His research interests aremainly but not limited to Cloud Computing, InformationSystems, Information Retrieval, Knowledge Engineering,and Knowledge Representation.

Kashif Bilal received Ph.D. in Electrical and ComputerEngineering from North Dakota State University, Fargo,ND, USA in May 2014. He is affiliated with COMSATSInstitute of Information Technology, Abbottabad, Pakistansince 2004. His research domain encompasses topicsrelated to data center networks, distributed computing,wireless networks, and expert systems.

Limin Zhang is an associate professor of ManagementInformation Systems in the Accounting, Finance, andInformation Systems Department at North Dakota StateUniversity. She receives her Ph.D. in Management Infor-mation Systems from the University of Arizona. Her re-search interests include context-basedWeb search, virtualteam collaboration, and using technology to support com-petitive intelligence and decision-making. Her researchwork has appeared in a number of journals and variousconference proceedings. Her teaching interests includedatabase design, Web system development, and informa-

tion technology management.

Samee U. Khan received a BS degree in 1999 fromGhulamIshaq Khan Institute of Engineering Sciences and Technol-ogy, Topi, Pakistan, and a Ph.D. in 2007 from the Universityof Texas, Arlington, TX, USA. Currently, he is Associate Pro-fessor of Electrical and Computer Engineering at the NorthDakota State University, Fargo, ND, USA. Prof. Khan’s re-search interests include optimization, robustness, and se-curity of: cloud, grid, cluster and big data computing, socialnetworks, wired and wireless networks, power systems,smart grids, and optical networks. His work has appearedin over 250 publications. He is a Fellow of the Institution

of Engineering and Technology (IET, formerly IEE), and a Fellow of the British Com-puter Society (BCS).