a hybrid service recommendation prototype …downloads.hindawi.com/journals/wcmc/2017/6783240.pdfa...

Research ArticleA Hybrid Service Recommendation Prototype Adapted forthe UCWW A Smart-City Orientation

Haiyang Zhang1 Ivan Ganchev12 Nikola S Nikolov13 Zhanlin Ji14 and Maacuteirtiacuten OrsquoDroma1

1Telecommunications Research Centre (TRC) University of Limerick Limerick Ireland2Department of Computer Systems University of Plovdiv ldquoPaisii Hilendarskirdquo Plovdiv Bulgaria3Department of Computer Science and Information Systems University of Limerick Limerick Ireland4North China University of Science and Technology Tangshan China

Correspondence should be addressed to Ivan Ganchev ivanganchevulie

Received 1 April 2017 Revised 11 August 2017 Accepted 20 August 2017 Published 12 October 2017

Academic Editor Damianos Gavalas

Copyright copy 2017 Haiyang Zhang et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

With the development of ubiquitous computing recommendation systems have become essential tools in assisting users indiscovering services they would find interestingThis process is highly dynamic with an increasing number of services distributedover networks bringing the problems of cold start and sparsity for service recommendation to a new level To alleviate theseproblems this paper proposes a hybrid service recommendation prototype utilizing user and item side information which naturallyconstitute a heterogeneous information network (HIN) for use in the emerging ubiquitous consumer wireless world (UCWW)wireless communication environment that offers a consumer-centric and network-independent service operationmodel and allowsthe accomplishment of a broad range of smart-city scenarios aiming at providing consumers with the ldquobestrdquo service instances thatmatch their dynamic contextualized and personalized requirements and expectations A layered architecture for the proposedprototype is described Two recommendation models defined at both global and personalized level are proposed with modellearning based on the Bayesian Personalized Ranking (BPR) A subset of the Yelp dataset is utilized to simulate UCWW dataand evaluate the proposed models Empirical studies show that the proposed recommendation models outperform several widelydeployed recommendation approaches

1 Introduction

With the rapid development of ubiquitous computing peopletoday are able to access any services anytime and anywhereMany studies have been done in exploiting wireless commu-nications models for use in ubiquitous network for exampleNGMN (Next Generation Mobile Network) [1] and MUSE(Mobile Ubiquity Service Environment) [2] Among themthe ubiquitous consumer wireless world (UCWW) [3 4]brings a different approach to the current global wirelessenvironment setting out a generic network-independent andconsumer-centric techno-business model (CBM) foundationfor future wireless communications The primary change theUCWW brings is that the users become consumers insteadof subscribers and thus potentially are able to use the mobileservice of any service provider (SP) via the ldquobestrdquo available

access network of any access network provider (ANP)Figure 1 depicts a high-level view of the UCWW [3]

One of the key UCWW features is related to the provisionof a personalized and customized list of preferred mobileservices to consumers by taking into account their prefer-ences as well as the current network and service context [5]The following are some possible scenarios for utilizing theUCWWwithin the smart-city paradigm [6]

(i) Smart parking service when a consumer in herhiscar enters a universityhospital campus or a similarfacility she will automatically get a recommendationfor the ldquobestrdquo car parking spaces with allocation andreservation options subject to herhis profile prefer-ences and campus parking policies The recommen-dation will come with enhanced functions and infor-mation options if required by the consumer profile

HindawiWireless Communications and Mobile ComputingVolume 2017 Article ID 6783240 11 pageshttpsdoiorg10115520176783240

2 Wireless Communications and Mobile Computing

ANPn

(best for web)

3P-AAA platform$

Internet

VoIPVideo streaming

mLearning

providersproviders

providers

mShopping

SALE $

mGovernment

UCWW

Servicerecommendation

system (SRS) Data management platform (DMP)

ABCampSmobile user(consumer)ANP SDs

ANP1ANP2ANP3ANPn

xSP

SDs

xSP1

xSP2

xSP3

xSP4

m 12

C

ANP3

(best for video streaming)

ANP2

(best for businessincoming callsamp SMSMMS)

ANP1(best for family-related

incoming callsamp international outgoing calls)

Figure 1 The UCWW a high-level view

for example reservation fee payment scheme anddetailed directions to that parking space on a standardnavigator app or other proprietary app Options forprovision of all or part of this service for examplethe key parking space reservation can bemade underother conditions for example as a ldquoyesrdquo response toldquoreserve parking at my work-placerdquo pop-up on amobile device first thing in the morning even beforeleaving from home to go to work

(ii) Personal-health location reminders the goal of thisservice is to present the consumer with up-to-date

notifications about lowest priced consumer-prescribed drugs in drugstorespharmacies withinthe geographic location of the consumerThere wouldbe matching service descriptions (SDs) for apps tocollect and collate the information for example aspart of a cloud-based service recommendation sys-tem from cooperating drugstores In the SD for suchan app alerts or reminders may be set manuallythrough profile policy when the consumer is withineasy reach of a drugstore with the lowest priceddrug There are many consumer-oriented variationsof such a kind of service leading to many ways

Wireless Communications and Mobile Computing 3

personal-health location reminders may work fordifferent people Also this service can potentiallysupport other smart-city healthy living applicationsfor example targeted profile-based real-time alertsabout areas of high and low pollen count pollutionair quality index (AQI) and so on or more specificalerts about consumer moves around the city

In order to support consumer requirements in scenariossuch as those described above recommendation techniquesbecome essential tools assisting consumers in seeking the bestavailable servicesThe services in theUCWWare divided intotwo broad categories access network communication services(ANCSs) and teleservices (TSs) [7] ANCSs are used by theconsumer to find and use the best access network available inthe current location while TSs are more complex containingall non-access-network services from e-learning to onlineInternet shopping email andmultimedia services [4] In thiswork we only focus on TSs recommendation problems Theterms ldquoservicesrdquo and ldquoitemsrdquo are used to refer to TSs andldquousersrdquo is used to refer to consumers in the rest of the paper

In this paper a hybrid recommendation prototype forTSs advertising is proposed working as a platform to assistservice providers to reach their valuable targeted users whileat the same time offering each user a list of ranked serviceinstances they may be interested in To alleviate the cold startand sparsity problems we propose to leverage the rich sideinformation related to users and services constructed asa heterogeneous information network (HIN) to build theproposed recommendation models The proposed modelscan be potentially also utilized in other recommendationsystems The contributions of this paper are summarized asfollows

(i) First we design a layered recommendation frame-work for use in the UCWW consisting of an offlinemodeling part and an online recommendation part

(ii) Second we propose to leverage HIN to model theinformation related to users and services fromwhichrich entity relationships can be generated The richrelationships are combined with implicit user feed-back in a collaborative filtering way to alleviate thecold start and sparsity problems Recommendationmodels are defined at both global and personalizedlevel in this paper and are estimated by the BayesianPersonalized Ranking (BPR) optimization technique[8]

(iii) Third we select a subset of the Yelp dataset toconstruct the HIN which is complementary to theUCWW service recommendation scenario Based onthis dataset extensive experimental investigations areconducted to show the effectiveness of the proposedmodels

The remainder of the paper is organized as followsSection 2 presents some related work in this area Section 3introduces the background and preliminaries for this studySection 4 presents the layered configuration of the rec-ommendation prototype architecture The proposed global

and personalized recommendation models are presented inSection 5 with parameters estimated in Section 6 Section 7presents and analyses the experimental results FinallySection 8 concludes the paper and suggests future researchdirections

2 Related Work

21 Collaborative Filtering with Additional Information Col-laborative filtering (CF) is the most successful and widelyused recommendation approach to build recommendationsystems It focuses on learning user preferences by dis-covering usage patterns from the userndashitem relations [9]CF recommendation algorithms are typically favored overcontent-based filtering (CBF) algorithms due to their overallbetter performance in predicting common behavior patterns[10] In the past few decades huge amount of work was doneon exploiting userndashitem rating matrices to generate recom-mendations [11ndash14]

In recent years there is an increasing trend in exploitingvarious kinds of additional information to solve the cold startand sparsity problems in CF as well as to improve the rec-ommendation quality of CF models With the prevalence ofsocial media social networks have been popular resource toexploit in order to improve recommendation performanceMa et al [15] introduce a novel social recommendationframework fusing the userndashitem matrix with usersrsquo socialtrust networks using probabilistic matrix factorizationGuo et al [16] propose a trust-based matrix factorizationapproach TrustSVD which takes both implicit influence ofratings and trust into consideration in order to improve therecommendation performance and at the same time to reducethe effect of the data sparsity and cold start problems Userand item side information is also a popular informationsource for incorporation into CF models in the form of tags[17 18] user reviews [19 20] and so on

To further improve the recommendation performanceHINs have been used to model information related to usersand items in which entities are of various types and linksrepresent various types of relations [21] Yu et al [22] intro-duce a matrix factorization approach with entity similarityregularization where the similarity is derived from metap-aths in a HIN Luo et al [23] proposed a social collaborativefiltering method HeteCF based on heterogeneous socialnetworks Zheng et al [24] propose a new dual similarityregularization to enforce the constraints on both similar anddissimilar objects based on a HIN Majority of the worksrelated to HINs are based on explicit feedback data fewworks have been done exploiting implicit feedback dataYu et al [25] propose to utilize implicit feedback data todiffuse user preferences along different metapaths in HINsfor recommendation generation However there are somelimitations to this work Firstly the authors learn a low-rank representation for the diffused rating matrix under eachmetapath which makes the computational complexity of themodel training stage relatively high Secondly the authorsmake personalized recommendation based on a group ofusers obtained by clustering However finding a suitablenumber of clusters for a dataset is a challenging problem and


the recommendation performance heavily depends on thequality of the clusters

In this study we propose to use item similarities alongdifferent metapaths in a HIN directly to enrich the item-basedCF Recommendationmodels are defined at both globaland personalized level where different metapath weights arelearned for each user avoiding the use of user clusters

22 Top-N Recommendation with Implicit Feedback Everyrecommendation algorithm relies on the past user feedbackfor example the user profiling in CBF and the user similarityanalysis inCFThe feedback is either explicit (ratings reviewsetc) or implicit (clicks browsing history etc) [26] Althoughit seems more reliable to make recommendations usingthe information explicitly supplied by users themselves theusers are usually reluctant to spend extra time or effort onsupplying such information and sometimes the informationthey provide is inconsistent or incorrect [27] Compared toexplicit feedback implicit feedback can be collected in amuch easier and faster way and at a much larger scale sinceit can be tracked automatically without any user effort Forthis reason there has been an increasing research attentionto the task of making recommendations by utilizing implicitfeedback as opposite to explicit feedback data [28]

Along with recommendation based on implicit feedbackin the last few years great attention was paid to the top-119873recommendation problem Many works have been publishedaddressing both tasks [29 30] While rating predictionattempts to predict unrated values for each user as accurateas possible top-119873 recommendation aims at discovering aranked list of itemswhich are themost interesting for the user

In the UCWW recommendation scenario with con-sumers feedback available the proposed hybrid recommen-dation methods should be able to provide a list of top-119873services for each active consumer

3 Background and Preliminaries

31 Heterogeneous Information Network Most entities in thereal world are interconnected which can be representedwith information networks for example social networks andresearch networksThe entity recommendation problem alsoexists in an information network environment with itemsrecommended by mining different type of relations fromresources that are related to users and items

In real-world recommendation scenarios multiple-typeobjects and multiple-typed links are involved Thus therecommendation problem could be modeled with hetero-geneous information networks (HINs) [21] The followingdefinition of an information network was adopted from [21]

Definition 1 (information network) An information networkis defined as a directed graph 119866 = (119881 119864) with an object typemapping function 120601 119881 rarr 119860 and a linked type mappingfunction 120593 119864 rarr 119877 Each object V isin 119881 belongs to oneparticular object type 120601(V) isin 119860 and each link 119890 isin 119864 belongsto a particular relation 120593(119890) isin 119877 [21]

When the number of object types |119860| is greater than 1or the number of relation types |119877| is greater than 1 the

Group

Consumer interaction Service

Tag

Category

Figure 2 Network schema in UCWW

network is called a heterogeneous information network (HIN)otherwise it is a homogeneous information network

In a HIN an abstract graph is used to represent the entityand relation-type restrictions as per the following definition

Definition 2 (network schema) A network schema 119878119866 of 119866 isthe directed graph defined over the object type 119860 with edgesfrom 119877 denoted as 119878119866 = (119860 119877) [21]

The definition of the network schema sets the rules onwhat types of entities exist and how they are connected inan information network The network schema designed foruse in the UCWW service recommendation is shown inFigure 2 Links between a consumer and a service denotetheir interactions links between a service and a tag or a ser-vice and a category denote the corresponding attributes for aservice and links between a consumer and a group or a con-sumer and another consumer denote their social relation-ships

In a HIN two entity types could be connected viadifferent types of relationships following the network schemathus generating ametapath

Definition 3 (metapath) A metapath 119875 = 11986011198771997888997888rarr 1198602

1198772997888997888rarrsdot sdot sdot119877119897997888rarr 119860 119897+1 is a path defined on a network schema 119878119866 =

(119860 119877) [21]

Each metapath can be considered as a type of a path inan information network representing one relation betweenentity pairs in aHIN An example of service recommendationin the personal-health location reminder scenariomentionedin Section 1 is described in Example 1

Example 1 A drug sale reminder service which advertisesa healthcare product will belong to the ldquopersonal-healthrdquocategory andwill have tags like ldquosalerdquo ldquohealthcarerdquo and so onwhich are supposed to be defined by the service providers Fora consumer 119888 if the recommendation system found that someof this consumerrsquos friends used the same service in the last twoweeks this service will be in the rank list for recommendationto consumer 119888 under a consumer-consumer-servicemetapath

32 Metapath Based Similarity In a HIN rich similari-ties between entities can be generated following different


Offline modeling Online recommendation

Request

Top-N services

Recommendation layer

Global recommendation

Computing layer

Similarity computing

Data layer

HIN construction

SDregistry

Third-party

monitoring

Userbehavior

Global parameters computing

Personalized parameters computing

Personalized recommendation

Figure 3 The UCWW service recommendation architecture

metapaths Different metapaths represent different semanticmeanings for example user-user denotes social relationbetween two users and user-service-usermeans that two usersare similar because they have similar service-usage histo-ries The network mining approaches used in homogeneousinformation networks such as the random walk used inpersonalized PageRank [31] and the pairwise random walkused in the SimRank [32] are not suitable for HINs becausethey are biased to either highly visible or highly concentratedobjects [33] In this study the PathSim approach is utilized toquantitatively measures the same-type objectsrsquo similarity in aHIN along symmetric metapath [33] Given two entities 119909 119910belonging to the same type in a HIN the PathSim is definedas follows [33]

Sim119901 (119909 119910) =2 times 119908 (119909 119910 | 119875)

119908 (119909 119909 | 119875) + 119908 (119910 119910 | 119875) (1)

where 119875 denotes the path type and 119908(119909 119910 | 119875) denotes thenumber of path instances between 119909 and 119910 along metapath119875

4 UCWW Service RecommendationArchitecture

The service recommendation system in the UCWW [3435] works as a platform for connecting service providerswith consumers The service recommendation architectureconsists of three layers (Figure 3) The data layer and thecomputing layer belong to the offlinemodeling part in whichthe similarities between services along different metapathsand their corresponding weights are precomputed In theonline recommendation part the top-119873 services for theactive user are computed at the recommendation layer basedon the results provided by the offline modeling part

41 Data Layer Information related to users and servicesis collected and extracted at this layer to construct a HINwhich works as both a service repository and a knowledgebase Compared to most semantic-based recommendationapproaches utilizing existing knowledge base or ontology[36] recommendation using a HIN as a knowledge base ismore flexible as it is able to define its own rules (networkschema in HIN) for different recommendation requirements

As shown in Figure 3 in the UCWW information aboutconsumers and services is collected from three differentsources

(i) A central registry where service descriptions (SDs)are stored including attributes such as categoryquality of service (QoS) biding price and consumerspackage [37]

(ii) A third-party monitoring platform which providesinformation about the number of clicksrequestsmade by consumers for services

(iii) User interactions with services in the past or socialrelations between users extracted from other socialresources and so on (details about data collection anddata management platform can be found in [34 35])

42 Computing Layer In a HIN items could be similar viadifferent types of relations which represent different reasonsfor similarity Therefore similarity between items in a HINcould be computed from a combination of different relationsrather than only from the rating distributions as in thetraditional item-based CF The main task of this layer is tocompute service similarities along different metapaths in theHIN and learn the weights for each metapath in both globaland personalized recommendation models

43 Recommendation Layer This is the most external user-facing layer presenting system facade to the consumers Allthe queries are performed through this layer When a userhas a request for finding the ldquobestrdquo instance of a particularservice a ranked list (computed according to a certainrecommendation model) is provided as a response back tohimher

5 Semantic Recommendation Model

In the UCWW recommendation scenario the number ofservices and consumers is relatively high which can causeeven more serious cold start and sparsity problems in servicerecommendation In this section we propose to exploit theside information related to services and consumers to allevi-ate this problem The side information is first constructed asa HIN from which rich service similarities under different


semantics are calculated The proposed models incorporatethese similarities into item-based CF to improve the predic-tion accuracy For each user the recommendation systemwillfirst calculate the prediction score for each unrated serviceand then recommend the top-119873 services with the highestscores to that user

51 Global Recommendation Model The item-based CFapproach tries to find similar items to the target item basedon their rating pattern However with an additional datasource related to items and users items could be similarbecause of different reasons based on different features ofitems In the UCWW context within the scope of the HINservices could be similar due to different reasons via differentmetapaths For instance service-consumer-service representsthe relation used in the traditional item-based CF denotingthat two services are similar because they are used by agroup of consumers while service-category-service meansthat two services are similar because they share the samecategory If one can understand the underlying semanticrelations between services and discover services based on richrelations then potentially more accurate recommendationscan be provided to the consumers Based on this observationand the background knowledge presented in Section 3 aglobal recommendation model [38] is proposed which uti-lizes metapaths with the following format servicendashlowastndashservice

Given a metapath 119875 = 11986011198771997888997888rarr 1198602

1198772997888997888rarr sdot sdot sdot119877119897997888rarr 119860 119897+1 with

1198601 = 119904119890119903V119894119888119890 and 119860 119897+1 = 119904119890119903V119894119888119890 the predicted value of aconsumer 119888 for service 119894 could be defined as follows

119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)

where Sim119901(119894 119895) is the PathSim value between service 119894 andservice 119895 along the 119901th metapath 119871 is the number of differentmetapaths considered 120579119901 is the weight of the 119901th metapathamong all 119871 metapaths (since different types of metapathsrepresent different relationship semantics and naturally havedifferent importance in the recommendation model) and 119877+119888denotes the set of services with user interactions in the past

52 Personalized Recommendation Model With the globalrecommendation model proposed in the previous subsec-tion consumers are provided with potentially interesting(for them) services based on both different types of ser-vice relations with rich semantic meanings and service-rating patterns from consumer feedback However in real-world UCWW scenarios consumersrsquo interests in particularfeatures may differ from each other For instance takingthe online shopping case as an example the price of aphoto camera is usually much more important criterionfor buying than its color which could be learned from theglobal recommendation model However it may happen thatone consumer simply wants a camera of a certain colorregardless of the price whichmeans that themetapath whichincludes the corresponding tag (a certain color) should havehigher importance In this case the accuracy of the globalrecommendation model may not be sufficient because it

only considers the overall weights of features without takinginto consideration the consumersrsquo individual preferencesIn order to better capture the consumer preferences andinterests a fine-grained personalized recommendation modelis also elaborated in this work with consideration of everyconsumerrsquos interests It allows a higher degree of personal-ization compared to the global recommendation model Thepersonalized recommendation model applied to consumer 119888and service 119894 is defined as follows

119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)

where 119908119888119901 represents the weight of interest of consumer 119888 inthe pth feature (metapath) and 119908119888 is the vector representingthe consumerrsquos preferences for all features (metapaths)

Compared to the global recommendation model with119871 parameters to learn the personalized recommendationmodel needs to learn |119862| times119871 parameters where 119862 is the set ofcustomers

For both the global and personalized recommendationmodels given a consumer one can calculate the recommen-dation scores for all services by utilizing either (2) or (3) andthen the top-119873 services can be returned to that consumer asthe recommendation result Parameter estimation methodsfor both models are introduced in the next section

6 Recommendation Models Optimization

The objective of the recommendation task is to recommendunrated items with the highest prediction score to each userA large number of previous studies concentrate on predict-ing unrated values for each user as accurately as possibleHowever the ranking over the items is more important [39]Considering a typical UCWW recommendation scenariowith only a binary consumer feedback available a rank-based approach Bayesian Personalized Ranking (BPR) [8]could be utilized to estimate parameters in the proposedrecommendationmodelsThe assumption behind BPR is thatthe user prefers a consumed item to an unconsumed itemaiming to maximize the following posterior probability

119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)

where 119877 is the rating matrix 119901(Θ | 119877) represents the likeli-hood of the desired preference structure for all users accord-ing to 119877 andΘ is the parameter vector of an arbitrary modelThus BPR is based on pairwise comparisons between a smallset of positive items and a very large set of negative items fromthe usersrsquo histories BPR estimates parameters by minimizingthe loss function defined as follows [8]

119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)

where 120590 = 1(1 + 119890minus119909) is the sigmoid function of 119909 119862 is theset of available consumers 119903119888119894 and 119903119888119895 are the predicted scoresof consumer 119888 for items 119894 and j and 119877minus119888 is the set of itemswithout user ratings yet Parameters are estimated by meansof minimization


Input 119877 implicit feedback119866 information network

Output Learned global meta-path weights 120579(1) Initialize 120579(2) Generate triples119863119904 = 119889((119888 119894 119895) | 119894 isin 119877+119888 119895 isin 119877

minus119888 )

(3) while not converged do(4) while 119889 isin 119863119904 do(5) compute 120597119874120597120579 with equation (6)(6) end(7) 120579 larr997888 120579 minus 120572120597119874

120597120579(8) end

Algorithm 1 Global recommendation model learning

61 Global Recommendation Model Learning In the globalrecommendation model the parameter for estimation is120579 = 1205791 120579119871 which represents the global weights of allmetapaths considered

The gradient descent (GD) approach [40] could be usedto estimate this parameterThe gradient with respect to 120579 canbe calculated as follows

120597119874120597120579

= minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)

where 119903119888119894119895 = 119903119888119894 minus 119903119888119895 For each 120579119901 in 120579 = 1205791 120579119871 thegradient of 119903119888119894119895 is

120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)

The process of learning the global recommendation model ispresented in Algorithm 1

62 Personalized Recommendation Model Learning In thepersonalized recommendation model learning process oneneed to learn |119862| times 119871 parameters with a weighted vector ofmetapaths for each consumer Considering the large numberof consumers and services in the UCWW and the corre-sponding huge number of parameters to learn we employ thestochastic gradient descent (SGD) [41] approach to estimatethe parameters for the personalized recommendation model

Similar to (6) for each triple (119888 119894 119895) (119888 119894) ≻ (119888 119895) theupdate step with respect to119908119888119901 is based on BPR and for eachtriple it is computed as follows

120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)

For each 119908119888119901 the gradient for 119903119888119894119895 is estimated as

120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)

The learning algorithm for the personalized recommen-dation model is presented in Algorithm 2

Table 1 Statistics of the dataset used in the experiments

Relations 119886 harr 119887 Numberof 119886

Numberof 119887

Numberof relations

Consumerharr service 2000 5000 8757Consumerharr consumer 2000 2000 2454Consumerharr group 2000 11 9484Serviceharr category 5000 47 49981Serviceharr tag 5000 511 14001

Table 2 Metapaths considered in experiments

Metapath Notationconsumer-(service-consumer-service) Pure item-based CF

consumer-(service-consumer-consumer-service)

Consumer social relationenriched item-based CF

consumer-(service-consumer-group-consumer-service)

Consumer group enricheditem-based CF

consumer-(service-category-service)

CBF with one featurerelated to items consideredconsumer-(service-tag-service)

consumer-(service-tag-service-tag-service)

7 Experiments

71 Experiment Setup In order to simulate a typical UCWWrecommendation scenario we define the network schemafor the proposed recommendation prototype as shown inFigure 2 We select a subset of the Yelp dataset (httpswwwyelpiedataset challenge) which contains user ratingson local business and attributes information related to usersand businesses After preprocessing the new dataset consistsof five matrices representing different relations The detailsof the dataset are shown in Table 1 In this dataset the con-sumer-service matrix contains 2000 consumers with 8757service binary interactions on 5000 services which leads toan extremely sparse matrix with a sparsity of 9991

We randomly take 70 of the consumer-service interac-tion dataset as a training set and use the remaining 30 asa test set Six different types of metapaths were utilized forboth models in the information network in the format ofservicendashlowastndashservice as shown in Table 2 For BPR parameterestimation fifty triples (119888 119894 119895) (119888 119894) ≻ (119888 119895) are randomlygenerated for each consumer in the training set

72 Evaluation Metrics and Comparative Approaches Inthe proposed service recommendation prototype a rankedlist of services with top-119873 recommendation score is pro-vided to the consumers Precision recall and 1198651-Measureare used to measure the prediction quality [42] In theUCWW service recommendation prototype precision indi-cates how many services are actually relevant among allselectedrecommended services whereas recall gives the



Output Learned personalized meta-path weight marix119882(1) Initialize119882(2) Generate triples119863119904 = 119889((119888 119894 119895) | 119894 isin 119877+119888 119895 isin 119877

minus119888 )

(3) while not converged do(4) while 119889 isin 119863119904 do(5) compute 120597119874120597119908119888119901 with equation (8)

119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end

Algorithm 2 Personalized recommendation model learning

number of selectedrecommended services among all rele-vant services

precision

= |recommended services cap used services||recommended services|

recall

= |recommendded services cap used services||used services|

(10)

In the evaluation of the adopted top-119873 recommendationmodel precision is normally inversely proportional to recallWhen119873 increases recall increases as well whereas precisiondecreases Therefore the 1198651-Measure which is the harmonicmean of precision and recall [43] was also used as per thefollowing definition

1198651-Measure = 2 lowast precision lowast recallprecision + recall

(11)

For all the three evaluation metrics a higher scoreindicates better performance of the corresponding approach

To demonstrate the effectiveness of the proposed modelswe evaluated and compared them with the following widelydeployed recommendation approaches

(i) Item-based CF (IB-CF) this is the traditional andwidely used item-based collaborative filtering thatrecommends items based on the itemrsquos 119896-nearestneighbors [11]

(ii) BPR-SVD this method learns the low-rank approxi-mation for the user feedbackmatrix based on the rankof the items with model learning by BPR optimiza-tion technique [8]

We use Hybrid-g and Hybrid-p to denote the pro-posed global and the personalized recommendation modelsrespectively

73 Experimental Results To examine the effectiveness ofthe proposed recommendation models we experimentally

0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p

Figure 4 Precision over different119873 (top-119873) values

computed the top-119873 list containing items with the highesttop-119873 recommendation score for each consumer in thetest set The evaluation and comparison results are shownin Figure 4 (precision) Figure 5 (recall) and Table 3 (1198651-Measure) from which several observations can be drawn

(i) First IB-CF outperforms BPR-SVD for small valuesof 119873 for both precision and recall but BPR-SVDachieves better results when119873 increases (119873 gt 7)

(ii) Second the two proposed recommendation models(Hybrid-g and Hybrid-p) sufficiently outperform theother two methods over a wide range of values of119873

(iii) Third the global model Hybrid-g shows overall bet-ter recommendation accuracy than the personalizedmodel Hybrid-p which may be due to the sparsityof the rating matrix as a relatively small number ofrated items cannot truly reflect the true interests ofconsumers

Similar to the IB-CF the rich similarities generated fromthe HIN in proposed models can be also precomputed andupdated periodically offline as well as the learned weights for


009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p

Figure 5 Recall over different119873 (top-119873) values

Table 3 1198651-Measure for different119873 (top-119873) values

1198651-Measure 11986515 119865110 119865115 119865120IB-CF 000849 0009085 0009833 0010424BPR-SVD 0006279 0011774 0013485 001404Hybrid-g 0017263 0020244 0018482 0017534Hybrid-p 0014582 0017731 0017073 0016001

Table 4 Comparison of computational complexities of comparedrecommendation algorithms

Algorithms Offline OnlineIB-CF 119874(1198982119899) 119874(119898ℎ)BPR-SVD 119874(119905119889119899) 119874(119898119889)Hybrid-g 119874(1198982119899|119871| + 119899119905) 119874(119898ℎ)Hybrid-p 119874((1198982119899 + 119899119905)|119871|) 119874(119898ℎ|119871|)

bothmodels Given 119899 consumers and119898 services for an activeconsumer the upper bound of the computational complexityfor top-119873 recommendation among all algorithms addressedin this paper is shown in Table 4 where ℎ denotes the numberof services the active consumers already used t is the numberof iterations for learning parameters and 119889 is the number oflatent features in the matrix factorization approach and |119871| isthe number ofmetapaths considered in the proposedmodels

As ℎ and 119889 are much smaller than m we can assumethat the proposed global recommendation model has similarcomputational complexity to both the traditional IB-CF andBPR-SVD approaches in the online recommendation stagebut higher computational complexity in the offline modelingstage for achieving better effectiveness The computationalcomplexity of the personalized recommendation model ishigher than the global recommendation model in both theoffline and online stages with a different set of weightsfor each user to learn and combine Between the proposedmodels the global recommendation model provides betterresults than the personalized model and achieves this with

lower computation complexity in both the offline modelingstage and the online recommendation stage

8 Conclusion

Mobile phones are currently the most popular personalcommunication devicesThey have formed a newmedia plat-form for merchants with their anytime-anywhere accessiblefunctionalities However the most important problem formerchants is how to deliver a service to the right mobile userin the right context efficiently and effectively The proposedservice recommendation prototype can potentially providea platform to assist service providers to reach their valuabletargeted consumers

The integration of the proposed service recommendationsystem prototype into the ubiquitous consumer wirelessworld (UCWW) has the potential to create an infrastructurein which consumers will have access to mobile servicesincluding those supporting smart-cities operation with aradically improved contextualization As a consequence thisenvironment is expected to radically empower individualconsumers in their decision making and thus positivelyimpact the society as awhole It will also facilitate and enable adirect relationship between consumers and service providersSuch direct relationship is attractive for the effective develop-ment of smart-city services since it allows for more dynamicadaptability and holds the potential for user-driven serviceevolution Besides benefiting consumers the UCWW opensup the opportunity for stronger competition between serviceproviders therefore creating a more liberal more open andfairer marketplace for existing and new service providersIn such a marketplace service providers can deliver a newlevel of services which are both much more specialized andreaching a much larger number of mobile users

The recommendation prototype proposed in this papercould be potentially employed for discovering the ldquobestrdquoservice instances available for use to a consumer throughthe ldquobestrdquo access network (provider) realizing a consumer-centric always best connected andbest served (ABCampS) expe-rience in UCWW In line with the layered architecture of theservice recommendation prototype two hybrid recommen-dation models which leverage a heterogeneous informationnetwork (HIN) are proposed at a global and personalizedlevel respectively for exploiting sparse implicit data Anempirical study has shown the effectiveness and efficiency ofthe proposed approaches compared to two widely employedapproachesTheproposed recommendationmodels also havethe potential to work under other recommendation scenarioseffectively

However for service recommendation in the UCWWwe only provided the basic recommendation models in thispaper without considering real-time context informationAlso the similarity matrices computed from different meta-paths are still sparse which may cause some inaccuraterating predictions As a future work we intend to conductfurther study on context aware recommendations with a realapplication operating with big dataWe also intend to explorethe study of matrix factorization approach on similaritymatrices derived from different metapaths


Disclosure

This paper is extended from the paper entitled ldquoHybridRecommendation for SparseRatingMatrix AHeterogeneousInformation Network Approachrdquo presented at the IAEAC2017

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This publication has been supported by the Chinese Schol-arship Council (CSC) the Telecommunications ResearchCentre (TRC) University of Limerick Ireland and the NPDof the University of Plovdiv Bulgaria under Grant noΦΠ17-ΦMJ-008

References

[1] North Alliance ldquoNGMN 5G white paperrdquo 2015 httpswwwngmnorg

[2] J Yang Z Ping H Zheng W Xu L Yinong and T XiaoshengldquoTowards mobile ubiquitous service environmentrdquo WirelessPersonal Communications vol 38 no 1 pp 67ndash78 2006

[3] M OrsquoDroma and I Ganchev ldquoToward a ubiquitous consumerwireless worldrdquo IEEE Wireless Communications vol 14 no 1pp 52ndash63 2007

[4] M OrsquoDroma and I Ganchev ldquoThe creation of a ubiquitous con-sumer wireless world through strategic ITU-T standardizationrdquoIEEE Communications Magazine vol 48 no 10 pp 158ndash1652010

[5] I Ganchev M OrsquoDroma N S Nikolov and Z Ji ldquoA ucwwcloud-based system for increased service contextualization infuturewireless networksrdquo inProceedings of the 2nd internationalconference on telecommunications and remote sensing ICTRS2013

[6] H Zhang I Ganchev N S Nikolov and M OrsquoDroma ldquoAservice recommendation model for the Ubiquitous ConsumerWireless Worldrdquo in Proceedings of the 2016 IEEE 8th Interna-tional Conference on Intelligent Systems (IS) pp 290ndash294 SofiaBulgaria September 2016

[7] P Flynn I Ganchev and M OrsquoDroma ldquoWBCs -ADA Vehicleand Infrastructural Support in a UCWWrdquo in Proceedings ofthe 2006 IEEE Tenth International Symposium on ConsumerElectronics pp 1ndash6 June 2006

[8] S Rendle C Freudenthaler Z Gantner and L Schmidt-Thieme ldquoBPR Bayesian personalized ranking from implicitfeedbackrdquo in Proceedings of the 25th conference on uncertaintyin artificial intelligence pp 452ndash461 AUAI Press 2009

[9] Y Shi M Larson and A Hanjalic ldquoCollaborative filteringbeyond the user-item matrix a survey of the state of the artand future challengesrdquo ACM Computing Surveys vol 47 no 1Article ID 2556270 pp 31ndash345 2014

[10] R Ronen N Koenigstein E Ziklik and N Nice ldquoSelect-ing content-based features for collaborative filtering recom-mendersrdquo in Proceedings of the 7th ACM Conference on Recom-mender Systems (RecSys rsquo13) pp 407ndash410 Hong Kong October2013

[11] G Linden B Smith and J York ldquoAmazoncom recommen-dations item-to-item collaborative filteringrdquo IEEE InternetComputing vol 7 no 1 pp 76ndash80 2003

[12] M H Aghdam M Analoui and P Kabiri ldquoCollaborativefiltering using non-negative matrix factorisationrdquo Journal ofInformation Science vol 43 no 4 pp 567ndash579 2017

[13] Y Koren R Bell and C Volinsky ldquoMatrix factorization tech-niques for recommender systemsrdquo Computer vol 42 no 8 pp30ndash37 2009

[14] Y Koren ldquoFactorization meets the neighborhood a multi-faceted collaborative filtering modelrdquo in Proceedings of the14th ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining (KDD rsquo08) pp 426ndash434 New YorkNY USA August 2008

[15] H Ma H Yang and M R Lyu ldquoSorec social recommendationusing probabilistic matrix factorizationrdquo in Proceedings of the17th ACM Conference on Information and Knowledge Manage-ment (CIKM rsquo08) pp 931ndash940NapaValley Calif USAOctober2008

[16] G Guo J Zhang and N Yorke-Smith ldquoA novel recommenda-tion model regularized with user trust and item ratingsrdquo IEEETransactions on Knowledge and Data Engineering vol 28 no 7pp 1607ndash1620 2016

[17] M G Manzato ldquogsvd++ supporting implicit feedback onrecommender systemswithmetadata awarenessrdquo inProceedingsof the 28th Annual ACM Symposium on Applied Computing pp908ndash913 Coimbra Portugal March 2013

[18] I Fernandez-Tobas and I Cantador ldquoExploiting social tagsin matrix factorization models for cross-domain collaborativefilteringrdquo in Proceedings of the International Workshop on NewTrends in Content based Recommender Systems pp 34ndash41 2014

[19] Y Bao H Fang and J Zhang ldquoTopicmf Simultaneouslyexploiting ratings and reviews for recommendationrdquo in Pro-ceedings of the Twenty-Eighth AAAI Conference on ArtificialIntelligence vol 14 pp 2ndash8 2014

[20] K Bauman B Liu and A Tuzhilin ldquoRecommending itemswith conditions enhancing user experiences based on sentimentanalysis of reviewsrdquo in Proceedings of the International Work-shop on New Trends in Content based Recommender Systems pp19ndash22 2016

[21] C Shi Y Li J Zhang Y Sun and P S Yu ldquoA survey of hetero-geneous information network analysisrdquo IEEE Transactions onKnowledge and Data Engineering vol 29 no 1 pp 17ndash37 2017

[22] X Yu X Ren Q Gu Y Sun and J Han ldquoCollaborativefiltering with entity similarity regularization in heterogeneousinformation networksrdquo in Proceedings of the International JointConference on Artificial Intelligence workshop on HeterogeneousInformation Network Analysis vol 27 2013

[23] C Luo W Pang Z Wang and C Lin ldquoHete-CF Social-BasedCollaborative Filtering RecommendationUsingHeterogeneousRelationsrdquo in Proceedings of the 2014 IEEE International Confer-ence on Data Mining (ICDM) pp 917ndash922 Shenzhen ChinaDecember 2014

[24] J Zheng J Liu C Shi F Zhuang J Li and B Wu ldquoRec-ommendation in heterogeneous information network via dualsimilarity regularizationrdquo International Journal of Data Scienceand Analytics vol 3 no 1 pp 35ndash48 2017

[25] X Yu X Ren Y Sun et al ldquoPersonalized entity recommen-dation A heterogeneous information network approachrdquo inProceedings of the 7th ACM international conference on Websearch and data mining pp 283ndash292 New York NY USAFeburary 2014


[26] M R Ghorab D Zhou A OrsquoConnor and V Wade ldquoPersona-lised information retrieval survey and classificationrdquo UserModelling and User-Adapted Interaction vol 23 no 4 pp 381ndash443 2013

[27] A Demiriz ldquoEnhancing product recommender systems onsparse binary datardquoData Mining and Knowledge Discovery vol9 no 2 pp 147ndash170 2004

[28] Y Hu C Volinsky and Y Koren ldquoCollaborative filtering forimplicit feedback datasetsrdquo in Proceedings of the 8th IEEEInternational Conference on Data Mining (ICDM rsquo08) pp 263ndash272 IEEE Pisa Italy December 2008

[29] P Cremonesi Y Koren and R Turrin ldquoPerformance of rec-ommender algorithms on top-N recommendation tasksrdquo inProceedings of the 4th ACM Recommender Systems Conference(RecSys rsquo10) pp 39ndash46 New York NY USA September 2010

[30] V C Ostuni T Di Noia E Di Sciascio and R Mirizzi ldquoTop-n recommendations from implicit feedback leveraging linkedopen datardquo in Proceedings of the 7th ACM Conference onRecommender Systems pp 85ndash92 ACM New York NY USA2013

[31] L Page S Brin R Motwani and T Winograd ldquoPageRank cita-tion ranking bringing order to thewebrdquo Stanford InfoLab 1999

[32] G Jeh and JWidom ldquoSimrank a measure of structural-contextsimilarityrdquo in Proceedings of the 8th ACMSIGKDD internationalconference on Knowledge discovery and data mining pp 538ndash543 Edmonton Alberta Canada July 2002

[33] Y Sun J Han X Yan P S Yu and TWu ldquoPathsim Meta path-based top-k similarity search in heterogeneous informationnetworksrdquo in Proceedings of the VLDB Endowment vol 4 pp992ndash1003 2011

[34] I Ganchev Z Ji and M OrsquoDroma ldquoA distributed cloud-basedservice recommendation systemrdquo in Proceedings of the 2015International Conference on Computing and Network Commu-nications (CoCoNet) pp 212ndash215 Trivandrum December 2015

[35] I Ganchev Z Ji and M OrsquoDroma ldquoA data management plat-form for recommending services to consumers in the UCWWrdquoin Proceedings of the 2016 IEEE International Conference onConsumer Electronics (ICCE) pp 405-406 Las Vegas NVJanuary 2016

[36] S E Middleton D De Roure and N R Shadbolt ldquoOntology-Based Recommender Systemsrdquo inHandbook on Ontologies pp779ndash796 Springer Berlin Heidelberg 2009

[37] Z Ji I Ganchev and M OrsquoDroma ldquoAdvertisement data man-agement and application design in WBCsrdquo Journal of Softwarevol 6 no 6 pp 1001ndash1008 2011

[38] H Zhang I Ganchev N S Nikolov Z Ji and M ODromaldquoHybrid recommendation for sparse rating matrix A hetero-geneous information network approachrdquo in Proceedings of thein 2017 IEEE Advanced Information Technology Electronic andAutomation Control Conference (IAEAC) 2017

[39] J Pessiot T Truong N Usunier M Amini and P GallinarildquoLearning to rank for collaborative filteringrdquo in Proceedingsof the 9th International Conference on Enterprise InformationSystems pp 145ndash151 Citeseer Funchal Madeira Portugal 2007

[40] C Burges T Shaked E Renshaw et al ldquoLearning to rankusing gradient descentrdquo in Proceedings of the 22nd InternationalConference on Machine Learning (ICML rsquo05) pp 89ndash96 ACMNew York NY USA 2005

[41] M Zinkevich M Weimer L Li and A J Smola ldquoParallelizedStochastic Gradient Descentrdquo inAdvances in neural informationprocessing systems pp 2595ndash2603 2010

[42] D M W Powers ldquoEvaluation from precision recall and f-measure to roc informedness markedness correlationrdquo Jour-nal of Machine Learning Technologies vol 2 no 1 pp 37ndash632011

[43] J L Herlocker J A Konstan L G Terveen and J T RiedlldquoEvaluating collaborative filtering recommender systemsrdquoACM Transactions on Information Systems vol 22 no 1 pp 5ndash53 2004

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Active and Passive Electronic Components

Control Scienceand Engineering

Journal of


International Journal of

RotatingMachinery


Hindawi Publishing Corporation httpwwwhindawicom

Journal of

Volume 201

Submit your manuscripts athttpswwwhindawicom

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics


Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation



DistributedSensor Networks



ANPn

(best for web)

3P-AAA platform$

Internet

VoIPVideo streaming

mLearning

providersproviders

providers

mShopping

SALE $

mGovernment

UCWW

Servicerecommendation

system (SRS) Data management platform (DMP)

ABCampSmobile user(consumer)ANP SDs

ANP1ANP2ANP3ANPn

xSP

SDs

xSP1

xSP2

xSP3

xSP4

m 12

C

ANP3

(best for video streaming)

ANP2

(best for businessincoming callsamp SMSMMS)

ANP1(best for family-related

incoming callsamp international outgoing calls)

Figure 1 The UCWW a high-level view

for example reservation fee payment scheme anddetailed directions to that parking space on a standardnavigator app or other proprietary app Options forprovision of all or part of this service for examplethe key parking space reservation can bemade underother conditions for example as a ldquoyesrdquo response toldquoreserve parking at my work-placerdquo pop-up on amobile device first thing in the morning even beforeleaving from home to go to work

(ii) Personal-health location reminders the goal of thisservice is to present the consumer with up-to-date

notifications about lowest priced consumer-prescribed drugs in drugstorespharmacies withinthe geographic location of the consumerThere wouldbe matching service descriptions (SDs) for apps tocollect and collate the information for example aspart of a cloud-based service recommendation sys-tem from cooperating drugstores In the SD for suchan app alerts or reminders may be set manuallythrough profile policy when the consumer is withineasy reach of a drugstore with the lowest priceddrug There are many consumer-oriented variationsof such a kind of service leading to many ways










2 Related Work















Group


Tag

Category









(119860 119877) [21]






Request

Top-N services



Computing layer


Data layer

HIN construction

SDregistry

Third-party

monitoring

Userbehavior






Sim119901 (119909 119910) =2 times 119908 (119909 119910 | 119875)

119908 (119909 119909 | 119875) + 119908 (119910 119910 | 119875) (1)



















119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)




119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)






119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)


119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)





minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation


















2 Related Work















Group


Tag

Category









(119860 119877) [21]






Request

Top-N services



Computing layer


Data layer

HIN construction

SDregistry

Third-party

monitoring

Userbehavior






Sim119901 (119909 119910) =2 times 119908 (119909 119910 | 119875)

119908 (119909 119909 | 119875) + 119908 (119910 119910 | 119875) (1)



















119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)




119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)






119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)


119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)





minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation




















Group


Tag

Category









(119860 119877) [21]






Request

Top-N services



Computing layer


Data layer

HIN construction

SDregistry

Third-party

monitoring

Userbehavior






Sim119901 (119909 119910) =2 times 119908 (119909 119910 | 119875)

119908 (119909 119909 | 119875) + 119908 (119910 119910 | 119875) (1)



















119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)




119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)






119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)


119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)





minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation











Request

Top-N services



Computing layer


Data layer

HIN construction

SDregistry

Third-party

monitoring

Userbehavior






Sim119901 (119909 119910) =2 times 119908 (119909 119910 | 119875)

119908 (119909 119909 | 119875) + 119908 (119910 119910 | 119875) (1)



















119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)




119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)






119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)


119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)





minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation















119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

120579119901Sim119901 (119894 119895) (2)




119903119888119894 = sum119895isin119877+119888

119871

sum119901=1

119908119888119901Sim119901 (119894 119895) (3)






119901 (Θ | 119877) prop 119901 (119877 | Θ) 119901 (Θ) (4)


119874 = minussum119888isin119862

sum119894isin119877+119888 119895isin119877

minus119888

ln120590 (119903119888119894 minus 119903119888119895) + 120582 Θ2 (5)





minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












minus119888 )


120597120579(8) end




120597119874120597120579


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597120579119903119888119894119895 + 120582120579 (6)


120597119903119888119894119895120597120579119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (7)




120597119874120597119908119888119901


sum119894isin119877+119888 119895isin119877

minus119888

119890minus1199031198881198941198951 + 119890minus119903119888119894119895

120597120597119908119888119901

119903119888119894119895 + 120582119908119888119901 (8)


120597119903119888119894119895120597119908119888119901

= sum119896isin119877+119888

(Sim119901 (119894 119896) minus Sim119901 (119895 119896)) (9)




Numberof 119887

Numberof relations











7 Experiments







minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation












minus119888 )


119908119888119901 larr997888 119908119888119901 minus 120572120597119874120597119908119888119901

(6) end(7) end



precision


recall


(10)



(11)







0018

0016

0014

0012

001

0008

0006

0004

0002

Prec

ision

N

6 8 10 12 14 16 18 204Top N

IB-CFBPR-SVD

Hybrid-gHybrid-p








009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










009

008

007

006

005

004

003

002

0

001

Reca

llN

6 8 10 12 14 16 18 2042Top-N

IB-CFBPR-SVD

Hybrid-gHybrid-p









8 Conclusion






Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation










Disclosure




Acknowledgments


References













































RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation




























RoboticsJournal of





Journal of



RotatingMachinery



Journal of

Volume 201


VLSI Design



Shock and Vibration







Journal of



Volume 2014


SensorsJournal of





Propagation









a hybrid service recommendation prototype …downloads.hindawi.com/journals/wcmc/2017/6783240.pdfa...

Documents