research article research on e-commerce platform-based...

Research ArticleResearch on E-Commerce Platform-Based PersonalizedRecommendation Algorithm

Zhijun Zhang Gongwen Xu and Pengfei Zhang

School of Computer Science and Technology Shandong Jianzhu University Jinan Shandong 250101 China

Correspondence should be addressed to Zhijun Zhang zzjsdcn163com

Received 22 February 2016 Accepted 26 June 2016

Academic Editor Francesco Carlo Morabito

Copyright copy 2016 Zhijun Zhang et al This 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

Aiming at data sparsity and timeliness in traditional E-commerce collaborative filtering recommendation algorithms whenconstructing user-item rating matrix this paper utilizes the feature that commodities in E-commerce system belong to differentlevels to fill in nonrated items by calculating RFIRF of the commodityrsquos corresponding level In the recommendation predictionstage considering timeliness of the recommendation system time weighted based recommendation prediction formula is adoptedto design a personalized recommendation model by integrating level filling method and rating time The experimental resultson real dataset verify the feasibility and validity of the algorithm and it owns higher predicting accuracy compared with presentrecommendation algorithms

1 Introduction

With the rapid development of the Internet and continuousexpansion of E-commerce scale commodity number andvariety increase quickly Merchants provide numerous com-modities through shopping websites and customers usuallytake a large amount of time to find their commoditiesBrowsing lots of irrelevant information and products willmake consumers run off due to the information overload Inthe E-commerce age users need an electronic shopping assis-tant which can recommend possible interesting or satisfyingcommodities according to interests and hobbies of usersTo solve all these problems a personalized recommendationsystem emerges [1]

Personalized recommendation recommends informationand commodities to users according to interests and purchas-ing behaviors of users Personalized recommendation systemis an advanced business intelligence platform established onthe basis of massive dataset mining and it aims at help-ing E-commerce websites provide completely personalizeddecision-making support and information service for cus-tomer purchase E-commerce platform-based personalizedrecommendation technology has been widely mentioned inacademia and industry The recommendation factors are

usually based on website best seller commodities user citypast purchase behaviors and purchase history to predict thepossible purchase behaviors of users

Traditional collaborative filtering (CF) algorithms haveproblems of data sparsity and cold start With the rapiddevelopment of the network technology personalized rec-ommendation in E-commerce environment faces new chal-lenges faster timeliness higher accuracy and stronger userpersonalization Its major feature is considering the influ-ences of real-time situation On the basis of traditionalcollaborative filtering algorithms three innovation pointsare added a more proper filling data method for nonratedcommodity adding time and giving highweight on data closeto evaluation time and lowweight on data far from evaluationtime and exploring the influences of the number of near-est neighbors on recommendation accuracy and obtainingoptimal nearest-neighbor set Through the abovementionedchanges the prediction accuracy of the algorithm can beimproved and the needs of usersrsquo personalized services canbe satisfied

The remainder of this paper is organized as followsIn Section 2 we provide an overview of related work athome and abroad Section 3 introduces the key technologyof E-commerce recommendation system Section 4 provides

Hindawi Publishing CorporationApplied Computational Intelligence and So ComputingVolume 2016 Article ID 5160460 7 pageshttpdxdoiorg10115520165160460

2 Applied Computational Intelligence and Soft Computing

Table 1 The typical recommendation systems

Field Personalized recommendation systemsElectronic mall Amazoncom eBay AlibabaMovie MovieLens Netflixcom MoviefindercomMessage PHOAKS GroupLens p-TangoWeb page Siteseer QuIC R2P METIOREWMusic MusicYahoocom Ringo CoCoA

the experiment dataset and evaluation metrics which intro-duces the experimental scheme experiment results and itsanalysis followed by the conclusion and future work inSection 5

2 Related Work

With the continuous improvements of E-commerce plat-form E-commerce personalized recommendation systemhas gradually formed into a perfect system Academia andE-commerce enterprises have paid more and more attentionto the recommendation system At present many large-scalewebsites at home and abroad have provided recommendationfunction for users and many prototypes of personalizedrecommendation systems have emerged and obtained goodapplication effects A lot of reprehensive recommendationsystems are shown as in Table 1

Utilizing various social relations in social networkingservices for recommendation studies has achieved greatprogress and becomes the hotspot field of personalizedrecommendation studies Bonhard and Sasse studied theinfluences of social background on recommendation resultsand results proved that when users purchase commoditiesthey tend to accept the recommendation of acquaintance [2]Sinha and Swearingen carried out experiments by the aidof multiple online recommendation systems and the experi-mental results indicated that when online systems and friendsboth provided recommendations users tended to select thelatter [3] Caverlee et al [4] constructed the trust frameworkby online social networking services which adopted trustand feedback information to generate recommendation resultlists with higher recommendation precision Adomaviciusand Tuzhilin [5] proposed multidimensional space recom-mendation algorithm and pointed out that it was necessaryto add recommendation feature dimensions according tospecific conditions Nguyen et al proposed a nonlinearprobability algorithm GPFM for context recommendationmodel In the recommendation process this algorithm usedGaussian process which can not only display feedback infor-mation but also use implicit feedback information Gradientdescent method is used for optimization which improvesthe model expansibility [6] Paper [7] compares four mainrecommendation technologies and introduces the reviewE-commerce research hot topic in the field of personal-ized recommendation The study [8] proposes personalizedproduct recommendations based on preference similarityrecommendation trust and social relations

Aiming at personalized problems of E-commerce manydomestic scholars carried out thorough studies Huang and

Benyoucef made a review on relevant literatures of E-commerce personalized recommendation illustrating theconcept of social commerce discussing the relevant designcharacteristics of social commerce and E-commerce andputting forward a new model and a set of principles to guidethe design of social commerce [9] Li et al proposed anE-commerce personalized recommendation algorithm thatintegrated commodity similarity recommendation trust andsocial relations [8] Experiment results indicate that socialrelations in social networking can be used to improve therecommendation algorithm accuracy Zhang and Liu raiseda personalized recommendation algorithm integrating trustrelationship and time series [10] In another paper [11] asocial networking recommendation algorithm integrating allkinds of context information was proposed On the basisof usersrsquo geographical location and time information thisalgorithm deeply explores the social relations of potentialusers and helps users to seek other users with similar pref-erence Then corresponding recommendations are made bycombining social relations of mobile users which effectivelysolves the recommendation accuracy In literature [12] theauthor comprehensively considered the influences of userpreference geographical convenience and friends and putforward a group purchase discount coupon recommendationsystem to promote the commodity with sensitive locations

It is not difficult to find out through deep analysis ofthe abovementioned algorithms that existing personalizedrecommendation algorithms still have many deficienciespoor expansibility of preference models inability to adaptto dynamic change of datasets resulting in lack of timeinformation that can be used and inability to solve cold startproblems verywell Aiming at the abovementioned problemsbased on a comprehensive consideration of factors such astimeliness of the recommendation system time weightedbased recommendation prediction formula is adopted anddifferent weights are given to rating data according to ratingtime so as to improve the recommendation quality of E-commerce recommendation system

3 Key Technology of Recommendation System

In order to better solve the problems of data sparsity andrating time factor this paper adopts level filling methodto predict the nonrated items and finally combines timeweights in the recommendation prediction stage to improvethe recommendation accuracy of the algorithm

31 Hierarchical Filling Method Traditional collaborativefiltering algorithm CF sets the nonrated items as the averageor a fixed value for example 3 (rating between 1 and 5)shown as in Table 2 Three is set for it is the middle ratingof 1ndash5 It does not consider the user preference and it purelysets the median as the prediction rating Different users givedifferent item rating The advantage of this method is simplebut it cannot solve the problems of traditional collaborativefiltering methods in sparse user matrixes

To reduce the sparsity of the rating matrix this paperadopts level fillingmethod to construct the ratingmatrix For

Applied Computational Intelligence and Soft Computing 3

Table 2 User-item rating matrix

Useritem Item 1 Item 2 Item 3 Item 4User 1 1 3 4 3User 2 3 2 3 3User 3 3 5 2 3

One

Two 1 Two n

Three 1 Three p Three 1 Three q

Item 1 Item r

middot middot middot

middot middot middotmiddot middot middot


Figure 1 Commodity hierarchy in E-commerce

E-commerce websites each commodity owns its categorywhich has a parent category Namely commodities in E-commerce own the concept of level and different commodi-ties own different hierarchies shown as in Figure 1 Thelevel of one commodity is considered in the constructionof rating matrix For commodities at different levels thispaper supposes its subordinate commodities to fill differentprediction rating through calculation This paper combinestraditional classification methods with user rating data andthrough calculation a preliminary rating is made on thenonrated data of recommendation usersThismethod is usedto construct a new user-item rating matrix

For rated data ratings are extracted to the belongingcategory In the construction of ratingmatrix by collaborativefiltering technology for one category its Rated Frequency(RF) is calculated and the calculation method is shown asfollows

RF = the number of rated itemsthe total number to be rated items

(1)

Item Rated Frequency (IRF) represents the weight ofrated items and the calculation method is shown as follows

IRF = log(SUMallrate

SUMdefaultrate) (2)

where SUMallrate represents the total amount of rated data andautomatically filled prediction rating SUMdefaultrate representsthe total number of automatically filled prediction ratings

This paper proposes a user-item rating matrix construc-tion algorithm which automatically fills ratings of nonrateddata in RF lowast IRF greater than the threshold The design flowof hierarchical filling method is shown as follows

Input initial user-item rating matrix 119875Output user rating matrix 119876 after the predictionrating is filled

Step 1 Calculate RF lowast IRF of each category in matrix 119875

Step 2 Fill in the average item rating of RFlowast IRF greater thanthreshold of the rating matrix

Step 3 For new items in ratingmatrix 3 is automatically filledand finally constructs the user-item rating matrix 119876

Through calculating RFlowastIRF of specific category of itemsthis paper fills scores of the top-119873 category items of RFlowast IRFinstead of simply filling dataset 0 which can reduce the datasparsity Finally the algorithm automatically fills new itemswith 3 aiming at solving cold start of new items

32 Improvement of Recommendation Timeliness CF algo-rithm does not take the influences of time on rating datainto consideration and it treats item ratings of different usersvisited at differentmoments equally Interests and preferencesof different users dynamically change with time so thetime when different users have interests in the same itemdiffers However if the rating is the same they are likely tobe regarded as similar neighbor users further influencingthe recommendation quality This paper introduces timefunction shown as follows

119891 (119905119886119894) = 119890

minus119905119886119894

(3)

where 119905119886119894

represents the time that user a has interests initem 119894 Time function119891(119905) is a monotone decreasing functionand decreases with the increase of time 119905 and time weightmaintains (0 1) Namely the newer the data is the greater theweight is and the time function is In this way the influenceof time on recommendation quality is solved

33 Time Weighted Based Prediction Function CF algorithmpredicts the rating of item 119894 by current user 119906 according to therating similarity of users or items shown as follows

119877119906119894= 119877119906+

sum

119873

119886=1(119877119886119894minus 119877119886) sdot sim (119906 119886)

sum

119873

119886=1|sim (119906 119886)|

(4)

where sim(119906 119886) represents the rating similarity between users119906 and 119886 and (5) calculates the similarity between users 119906 and119886 by Pearsonrsquos correlation coefficient

sim (119906 119886)

=

sum119894isin119868119906119886

(119877119906119894minus 119877119906) (119877119886119894minus 119877119886)

radicsum119894isin119868119906119886

(119877119906119894minus 119877119906)

2

radicsum119894isin119868119906119886

(119877119886119894minus 119877119886)

2

(5)

where 119868119906119886

represents the item set rated by users 119906 and 119886 119877119906119894

and 119877119886119894 respectively represent the score of item 119894 by users 119906

and 119886 and 119877119906and 119877

119886 respectively represent the mean score

of all items by users 119906 and 119886Time function 119891(119905) is added in (4) and the weighted

prediction rating of item 119894 by target user 119906 is improved

119875119906119894= 119877119906+

sum

119899

119886=1(119877119886119894minus 119877119886) sdot sim (119906 119886) sdot 119891 (119905

119886119894)

sum

119899

119886=1|sim (119906 119886)| sdot 119891 (119905

119886119894)

(6)


Client

Registration login

Shopping DabaBase

Sales data

Product segmentation

Clientrsquos file

Rating matrixpreprocessing

Hierarchical filling reduce sparse

Produce thenearest neighbors

Recommender engine

Client DB Trade DB

Sales record

Marketing management

Sales

Potential customer

Purchasing commodity

Data preprocessing

Figure 2 NewRec recommendation model

Here 119891(119905119886119894) is shown as (3) and each rating item owns

only oneweight Latest ratings are givenwith greatweight andpast ratings are given with small weight which helps predictmore accurately

34 New Recommendation Model (NewRec) Aiming at datasparsity and timeliness in traditional collaborative filteringrecommendation algorithms this paper integrates hierar-chical filling method and time on the basis of CF andputs forward a newpersonalized recommendation algorithmNewRec NewRec recommendation model is shown as inFigure 2 The whole recommendation model is divided intothree main modules data preprocessing module sparsityreduction module and nearest-neighbor recommendationmodule

Data preprocessing module input user informationincluding user purchase records user rating on commoditiesand user duration time on websites This useful informationis converted into acceptable data format of the recommenda-tion method forming user-item rating matrix

In sparsity reduction module for all the items in user-item ratingmatrix RFIRFof the commodityrsquos correspondinglevel is calculated and filled in the specific value of ratingmatrix which solves the problem of data sparsity

In nearest-neighbor recommendation module consider-ing timeliness of the recommendation system time weightedbased recommendation prediction formula is adopted tocalculate the prediction ratings of the target items rank themand select top-119873 items as recommendation set

4 Experimental Analysis

41 Dataset The dataset in this paper is from httpsmoviel-ensorg which is collected by GroupLens research groupin University of Minnesota This dataset realizes sites ofuser personalized recommendation by collaborative filteringtechnology The system adopts the user ratings ranging from1 to 5 The higher the rating is the more interested the usersare This dataset contains the ratings of 1682 movies by 943users According to the latest statistics there are over 70000users and 6600 rated movies in the database of MovieLenssite At present datasets in MovieLens site are abundantclear real and accurate so they have been widely used in thesimulation test of the personalized recommendation systemand authoritative test data sources in this field Taking this asthe simulation dataset this paper designs a reasonable andfeasible evaluation standard and carries out a comparativeanalysis on the recommendation quality of the improvedalgorithm The experimental results prove the validity andrationality of the improved algorithm

42 Experimental Scheme For collaborative filtering rec-ommendation algorithm its actual effects in E-commercepersonalized recommendation system are mainly influencedby two factors data sparsity and the number of the nearestneighbors Thus this experiment designs the following twoschemes

CF algorithm time-based function recommendation(TimeRec for short) hierarchical filling (HF for short) andNewRec in this paper under different degrees of data sparsityare compared Different degrees of data sparsity can trulysimulate the working condition of E-commerce recommen-dation system and verify the changes of recommendationeffects under different conditions of effective information

Under different numbers of nearest neighbors recom-mendation performances of CF HF TimeRec and NewRecare compared This process can verify the changes of rec-ommendation effects of each recommendation algorithmunder different numbers of nearest neighbors and helpeach recommendation algorithm select optimal number ofnearest neighbors for convenience of operation in futureexperiments

This section designs 5 experiments to verify the superior-ity of the algorithm in this paper

(1) The influences of different degrees of sparsity on rec-ommendation quality in the experiment this paperselected three degrees of data sparsity for comparison

(2) MAE comparison between hierarchical fillingmethod and traditional collaborative filtering CF

(3) The influences of time on recommendation accuracy(4) The influences of numbers of nearest neighbors on

recommendation algorithms the influences of differ-ent scales of nearest-neighbor sets on recommenda-tion quality are observed

(5) The recommendation qualities with the same num-ber of neighbors the recommendation qualities ofdifferent algorithms are compared


092081074

075

08

085

09

MA

E

20 30 40 50 6010NeighbourNum

Figure 3 The impact of data sparsity on recommendation algo-rithm

43 Baseline To test the performance ofNewRec recommen-dation model and time function-based improved algorithmTimeRec this paper will verify the validity of the model byexperiment Traditional collaborative filtering recommenda-tion algorithm CF [13] is taken as baseline CF algorithmutilizes the similarities between items to recommend similarcommodities for target users The similarities between usersor items can be calculated by (5)

44 Metrics To compare the algorithm performance thispaper adopts MAE and RMSE to evaluate the recommen-dation performance of the recommendation algorithm Thedefinition of MAE is shown as follows

MAE =sum119894119895

10038161003816100381610038161003816

119877119894119895minus119877119894119895

10038161003816100381610038161003816

119873

(7)

where119877119894119895represents the actual rating of commodity 119895 by user

119894 and 119877119894119895represents the prediction rating of commodity 119895 by

user 119894119873 represents the number of all prediction ratings Thedefinition of RMSE is shown as follows

RMSE = radicsum119894119895(119877119894119895minus119877119894119895)

2

119873

(8)

45 Experiment Results451 Influences of Data Sparsity on Recommendation Algo-rithm Data sparsity refers to the ratio of nonrated itemsto the elements in the whole rating matrix To verify theinfluences of data sparsity on recommendation accuracythis paper fills the prediction ratings in original user-itemrating matrix for recommendation calculation Datasets withsparsity of 092 081 and 074 are selected and CF algorithmwas used for verificationThe experimental results are shownas in Figure 3

It can be seen from Figure 3 that the recommendationquality does not increase with the decrease of sparsity In thisexperiment when the sparsity is 081 the recommendationquality is the highest In the following experiment datasetswith sparsity of 081 are taken for experiment

452 Analysis of Hierarchical Filling (HF) Method in Recom-mendation Accuracy To verify the influences of data sparsityon recommendation accuracy MAE is calculated before and

CFHF


075

08

085

09

095

MA

E

Figure 4 Analysis of HF in recommendation accuracy

CFTimeRec


075

08

085

09

MA

E

Figure 5 Influences of time on recommendation accuracy

after hierarchical filling (HF) method through experiment Itcan be seen from Figure 4 that with the increasing numberof focused users among neighbor users MAE of HF methodand MAE of CF method both decrease and MAE of HF issmaller than that of CF algorithm under the same number ofneighbors Thus HF method is better than CF algorithm

453 Influences of Time on Recommendation Accuracy Toguarantee the recommendation accuracy influences of timeon prediction rating stage shall be considered and each rateditem owns only one weight Latest ratings are endowed withgreater weight and past ratings are endowed with smallerweight which helps better forecast To verify the influencesof time on recommendation accuracy this section comparesMAE between CF algorithm and TimeRec algorithm

It can be seen from Figure 5 that MAE of the improvedTimeRec algorithm with time function is lower than thatof CF without time function Through comparison it isproved that time does have influences on recommendationprediction and the use of time function improves the recom-mendation quality of the recommendation system

454 Influences of the Number of User Neighbors on Recom-mendation Accuracy It is easy to calculate the nearest neigh-bor of each user by calculating the similarity between usersTo verify the influences of the number of user neighbors onrecommendation accuracy this section makes comparisonthrough experiment and the number of nearest neighborsincreased from 10 to 60 with interval of 10The experimentalresults are shown in Figure 6

It can be seen from Figure 6 that with the increasingnumber of nearest neighbors MAE of four algorithms alltend to decrease firstly and increase then However MAE


CFHF

TimeRecNewRec


075

08

085

09

MA

E

Figure 6 Influences of the number of user neighbors on accuracy

CFHF

TimeRecNewRec


07

08

09

1

RMSE

Figure 7 Comparison among different algorithms on accuracy

of the improved algorithm NewRec is lower than those ofthe other three algorithms which indicates that the NewReccan provide better recommendation quality than the otherthree From the further analysis it can be seen that fourrecommendation algorithms own lowest MAE when thenumber of nearest neighbors is 40 Namely when the numberof nearest neighbors is 40 four recommendation algorithmsall can achieve good recommendation quality

455 Comparison among Different Recommendation Algo-rithms on Recommendation Accuracy To verify the recom-mendation accuracy of NewRec algorithm proposed in thispaper this section calculates RMSE of algorithms throughexperiments and the experimental results are shown inFigure 7

It can be seen from Figure 7 that compared withtraditional collaborative filtering recommendation algorithmCF level filling-based improved algorithm HF and timefunction-based improved algorithm TimeRec the improvedalgorithm NewRec owns the highest recommendation accu-racy

To sum up the abovementioned experimental resultsthe following conclusion can be drawn Compared with theother three algorithms the recommendation quality of theimproved algorithm NewRec is significantly improved afterhierarchical filling and time function are added

This paper utilized the features that commodities in E-commerce system belong to different levels to fill in specificscore in rating matrix by calculating RFIRF of the commod-ityrsquos corresponding level which solves problems of data spar-sity and cold start to certain extent In the recommendation

prediction stage in consideration of timeliness of the rec-ommendation system timeweighted based recommendationprediction formula is adopted and different weights are givento rating data according to rating time so as to improve therecommendation quality of E-commerce recommendationsystem The experiment results in real dataset indicate thatthe algorithm in this paper is better than the traditionalcollaborative filtering recommendation algorithm in runningefficiency and recommendation accuracy

5 Conclusions

Collaborative filtering is a common recommendation tech-nology of E-commerce personalized recommendation sys-tem However it also owns many problems For data sparsityin user-item rating matrix and timeliness of user evalua-tion this paper proposes an improved collaborative filteringrecommendation algorithm NewRec and verifies the feasi-bility of NewRec algorithm through experiment simulationproving that it can improve the recommendation quality ofE-commerce recommendation system At present there arestill many problems and shortcomings in the studies of E-commerce personalized recommendation For user person-alized recommendation the improved collaborative filteringalgorithm in this paper fails to consider the influences ofcontext and user interaction behaviors which need furtherthorough studies in the future

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thispaper is supported by the Science andTechnologyDevel-opment Planning of Shandong Province (2014GGX1010112015GGX101018) A Project of Shandong Province HigherEducational Science and Technology Program (J12LN31J13LN11 and J14LN14) and Jinan Higher Educational Inno-vation Plan (201401214 201303001)

References

[1] Z Zhang and H Liu ldquoApplication and research of improvedprobability matrix factorization techniques in collaborativefilteringrdquo International Journal of Control amp Automation vol 7no 8 pp 79ndash92 2014

[2] P Bonhard and M A Sasse ldquolsquoKnowing me knowing yoursquomdashusing profiles and social networking to improve recommendersystemsrdquo BT Technology Journal vol 24 no 3 pp 84ndash98 2006

[3] R Sinha and K Swearingen ldquoComparing recommendationsmade by online systems and friendsrdquo inProceedings of theDelos-NSFWorkshop on Personalization and Rocommonder Systems inDigital Libraries 2001

[4] J Caverlee L Liu and S Webb ldquoThe SocialTrust frameworkfor trusted social information management architecture andalgorithmsrdquo Information Sciences vol 180 no 1 pp 95ndash1122010

[5] G Adomavicius and A Tuzhilin ldquoMultidimensional recom-mender systems a data warehousing approachrdquo in Electronic


Commerce Second International Workshop WELCOM 2001Heidelberg Germany November 16-17 2001 Proceedings vol2232 ofLectureNotes inComputer Science pp 180ndash192 SpringerBerlin Germany 2001

[6] T V Nguyen A Karatzoglou and L Baltrunas ldquoGaussian pro-cess factorization machines for context-aware recommenda-tionsrdquo in Proceedings of the 37th International ACM SIGIR Con-ference on Research and Development in Information Retrieval(SIGIR rsquo14) pp 63ndash72 ACM Gold Coast Australia July 2014

[7] Y Zhang ldquoE-commerce personalized recommendationrdquoAdvanced Materials Research vol 989ndash994 pp 4996ndash49992014

[8] Y-M Li C-T Wu and C-Y Lai ldquoA social recommendermechanism for e-commerce combining similarity trust andrelationshiprdquo Decision Support Systems vol 55 no 3 pp 740ndash752 2013

[9] Z Huang and M Benyoucef ldquoFrom e-commerce to socialcommerce a close look at design featuresrdquo Electronic CommerceResearch and Applications vol 12 no 4 pp 246ndash259 2013

[10] Z Zhang and H Liu ldquoSocial recommendation model com-bining trust propagation and sequential behaviorsrdquo AppliedIntelligence vol 43 no 3 pp 695ndash706 2015

[11] Z-J Zhang and H Liu ldquoResearch on context-awareness mobileSNS recommendation algorithmrdquo Pattern Recognition and Arti-ficial Intelligence vol 28 no 5 pp 404ndash410 2015

[12] Y-M Li C-L Chou and L-F Lin ldquoA social recommendermechanism for location-based group commercerdquo InformationSciences vol 274 pp 125ndash142 2014

[13] J Wang A P De Vries and M J T Reinders ldquoUnifyinguser-based and item-based collaborative filtering approaches bysimilarity fusionrdquo in Proceedings of the 29th Annual Interna-tional ACM SIGIR Conference on Research and Development inInformation Retrieval pp 501ndash508 ACM August 2006

Submit your manuscripts athttpwwwhindawicom

Computer Games Technology

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Distributed Sensor Networks


Advances in

FuzzySystems

Hindawi Publishing Corporationhttpwwwhindawicom

Volume 2014


ReconfigurableComputing

Hindawi Publishing Corporation httpwwwhindawicom Volume 2014


Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

Artificial Intelligence

HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014

Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi Publishing Corporation

httpwwwhindawicom Volume 2014

Advances in

Multimedia


Biomedical Imaging


ArtificialNeural Systems

Advances in


RoboticsJournal of



Computational Intelligence and Neuroscience

Industrial EngineeringJournal of


Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Human-ComputerInteraction

Advances in

Computer EngineeringAdvances in



Table 1 The typical recommendation systems

Field Personalized recommendation systemsElectronic mall Amazoncom eBay AlibabaMovie MovieLens Netflixcom MoviefindercomMessage PHOAKS GroupLens p-TangoWeb page Siteseer QuIC R2P METIOREWMusic MusicYahoocom Ringo CoCoA

the experiment dataset and evaluation metrics which intro-duces the experimental scheme experiment results and itsanalysis followed by the conclusion and future work inSection 5

2 Related Work

With the continuous improvements of E-commerce plat-form E-commerce personalized recommendation systemhas gradually formed into a perfect system Academia andE-commerce enterprises have paid more and more attentionto the recommendation system At present many large-scalewebsites at home and abroad have provided recommendationfunction for users and many prototypes of personalizedrecommendation systems have emerged and obtained goodapplication effects A lot of reprehensive recommendationsystems are shown as in Table 1

Utilizing various social relations in social networkingservices for recommendation studies has achieved greatprogress and becomes the hotspot field of personalizedrecommendation studies Bonhard and Sasse studied theinfluences of social background on recommendation resultsand results proved that when users purchase commoditiesthey tend to accept the recommendation of acquaintance [2]Sinha and Swearingen carried out experiments by the aidof multiple online recommendation systems and the experi-mental results indicated that when online systems and friendsboth provided recommendations users tended to select thelatter [3] Caverlee et al [4] constructed the trust frameworkby online social networking services which adopted trustand feedback information to generate recommendation resultlists with higher recommendation precision Adomaviciusand Tuzhilin [5] proposed multidimensional space recom-mendation algorithm and pointed out that it was necessaryto add recommendation feature dimensions according tospecific conditions Nguyen et al proposed a nonlinearprobability algorithm GPFM for context recommendationmodel In the recommendation process this algorithm usedGaussian process which can not only display feedback infor-mation but also use implicit feedback information Gradientdescent method is used for optimization which improvesthe model expansibility [6] Paper [7] compares four mainrecommendation technologies and introduces the reviewE-commerce research hot topic in the field of personal-ized recommendation The study [8] proposes personalizedproduct recommendations based on preference similarityrecommendation trust and social relations

Aiming at personalized problems of E-commerce manydomestic scholars carried out thorough studies Huang and

Benyoucef made a review on relevant literatures of E-commerce personalized recommendation illustrating theconcept of social commerce discussing the relevant designcharacteristics of social commerce and E-commerce andputting forward a new model and a set of principles to guidethe design of social commerce [9] Li et al proposed anE-commerce personalized recommendation algorithm thatintegrated commodity similarity recommendation trust andsocial relations [8] Experiment results indicate that socialrelations in social networking can be used to improve therecommendation algorithm accuracy Zhang and Liu raiseda personalized recommendation algorithm integrating trustrelationship and time series [10] In another paper [11] asocial networking recommendation algorithm integrating allkinds of context information was proposed On the basisof usersrsquo geographical location and time information thisalgorithm deeply explores the social relations of potentialusers and helps users to seek other users with similar pref-erence Then corresponding recommendations are made bycombining social relations of mobile users which effectivelysolves the recommendation accuracy In literature [12] theauthor comprehensively considered the influences of userpreference geographical convenience and friends and putforward a group purchase discount coupon recommendationsystem to promote the commodity with sensitive locations

It is not difficult to find out through deep analysis ofthe abovementioned algorithms that existing personalizedrecommendation algorithms still have many deficienciespoor expansibility of preference models inability to adaptto dynamic change of datasets resulting in lack of timeinformation that can be used and inability to solve cold startproblems verywell Aiming at the abovementioned problemsbased on a comprehensive consideration of factors such astimeliness of the recommendation system time weightedbased recommendation prediction formula is adopted anddifferent weights are given to rating data according to ratingtime so as to improve the recommendation quality of E-commerce recommendation system

3 Key Technology of Recommendation System

In order to better solve the problems of data sparsity andrating time factor this paper adopts level filling methodto predict the nonrated items and finally combines timeweights in the recommendation prediction stage to improvethe recommendation accuracy of the algorithm

31 Hierarchical Filling Method Traditional collaborativefiltering algorithm CF sets the nonrated items as the averageor a fixed value for example 3 (rating between 1 and 5)shown as in Table 2 Three is set for it is the middle ratingof 1ndash5 It does not consider the user preference and it purelysets the median as the prediction rating Different users givedifferent item rating The advantage of this method is simplebut it cannot solve the problems of traditional collaborativefiltering methods in sparse user matrixes

To reduce the sparsity of the rating matrix this paperadopts level fillingmethod to construct the ratingmatrix For




One

Two 1 Two n


Item 1 Item r








(1)



SUMdefaultrate) (2)









119891 (119905119886119894) = 119890

minus119905119886119894

(3)

where 119905119886119894



119877119906119894= 119877119906+

sum

119873

119886=1(119877119886119894minus 119877119886) sdot sim (119906 119886)

sum

119873

119886=1|sim (119906 119886)|

(4)


sim (119906 119886)

=

sum119894isin119868119906119886

(119877119906119894minus 119877119906) (119877119886119894minus 119877119886)

radicsum119894isin119868119906119886

(119877119906119894minus 119877119906)

2

radicsum119894isin119868119906119886

(119877119886119894minus 119877119886)

2

(5)

where 119868119906119886



and 119886 and 119877119906and 119877




119875119906119894= 119877119906+

sum

119899


119886119894)

sum

119899

119886=1|sim (119906 119886)| sdot 119891 (119905

119886119894)

(6)


Client

Registration login

Shopping DabaBase

Sales data


Clientrsquos file




Recommender engine

Client DB Trade DB

Sales record


Sales

Potential customer


Data preprocessing




















092081074

075

08

085

09

MA

E





MAE =sum119894119895

10038161003816100381610038161003816

119877119894119895minus119877119894119895

10038161003816100381610038161003816

119873

(7)





2

119873

(8)




CFHF


075

08

085

09

095

MA

E


CFTimeRec


075

08

085

09

MA

E








CFHF

TimeRecNewRec


075

08

085

09

MA

E


CFHF

TimeRecNewRec


07

08

09

1

RMSE








5 Conclusions


Competing Interests


Acknowledgments


References























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in






One

Two 1 Two n


Item 1 Item r








(1)



SUMdefaultrate) (2)









119891 (119905119886119894) = 119890

minus119905119886119894

(3)

where 119905119886119894



119877119906119894= 119877119906+

sum

119873

119886=1(119877119886119894minus 119877119886) sdot sim (119906 119886)

sum

119873

119886=1|sim (119906 119886)|

(4)


sim (119906 119886)

=

sum119894isin119868119906119886

(119877119906119894minus 119877119906) (119877119886119894minus 119877119886)

radicsum119894isin119868119906119886

(119877119906119894minus 119877119906)

2

radicsum119894isin119868119906119886

(119877119886119894minus 119877119886)

2

(5)

where 119868119906119886



and 119886 and 119877119906and 119877




119875119906119894= 119877119906+

sum

119899


119886119894)

sum

119899

119886=1|sim (119906 119886)| sdot 119891 (119905

119886119894)

(6)


Client

Registration login

Shopping DabaBase

Sales data


Clientrsquos file




Recommender engine

Client DB Trade DB

Sales record


Sales

Potential customer


Data preprocessing




















092081074

075

08

085

09

MA

E





MAE =sum119894119895

10038161003816100381610038161003816

119877119894119895minus119877119894119895

10038161003816100381610038161003816

119873

(7)





2

119873

(8)




CFHF


075

08

085

09

095

MA

E


CFTimeRec


075

08

085

09

MA

E








CFHF

TimeRecNewRec


075

08

085

09

MA

E


CFHF

TimeRecNewRec


07

08

09

1

RMSE








5 Conclusions


Competing Interests


Acknowledgments


References























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




Client

Registration login

Shopping DabaBase

Sales data


Clientrsquos file




Recommender engine

Client DB Trade DB

Sales record


Sales

Potential customer


Data preprocessing




















092081074

075

08

085

09

MA

E





MAE =sum119894119895

10038161003816100381610038161003816

119877119894119895minus119877119894119895

10038161003816100381610038161003816

119873

(7)





2

119873

(8)




CFHF


075

08

085

09

095

MA

E


CFTimeRec


075

08

085

09

MA

E








CFHF

TimeRecNewRec


075

08

085

09

MA

E


CFHF

TimeRecNewRec


07

08

09

1

RMSE








5 Conclusions


Competing Interests


Acknowledgments


References























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




092081074

075

08

085

09

MA

E





MAE =sum119894119895

10038161003816100381610038161003816

119877119894119895minus119877119894119895

10038161003816100381610038161003816

119873

(7)





2

119873

(8)




CFHF


075

08

085

09

095

MA

E


CFTimeRec


075

08

085

09

MA

E








CFHF

TimeRecNewRec


075

08

085

09

MA

E


CFHF

TimeRecNewRec


07

08

09

1

RMSE








5 Conclusions


Competing Interests


Acknowledgments


References























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




CFHF

TimeRecNewRec


075

08

085

09

MA

E


CFHF

TimeRecNewRec


07

08

09

1

RMSE








5 Conclusions


Competing Interests


Acknowledgments


References























Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in




















Advances in

FuzzySystems


Volume 2014












Journal of

Journal of





Advances in

Multimedia


Biomedical Imaging



Advances in


RoboticsJournal of










Advances in



research article research on e-commerce platform-based...

Documents