Research ArticleLeveraging Image Visual Features in Content-BasedRecommender System

Fuhu Deng Panlong Ren Zhen Qin Gu Huang and Zhiguang Qin

University of Electronic Science and Technology of China Chengdu China

Correspondence should be addressed to Zhen Qin qinzhenuestceducn

Received 26 April 2018 Accepted 17 July 2018 Published 12 August 2018

Academic Editor Jose M Lanza-Gutierrez

Copyright copy 2018 Fuhu Deng et al is 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

Content-based (CB) and collaborative filtering (CF) recommendation algorithms are widely used in modern e-commercerecommender systems (RSs) to improve user experience of personalized services Item content features and user-item rating dataare primarily used to train the recommendationmodel However sparse data would lead such systems unreliable To solve the datasparsity problem we consider that more latent information would be imported to catch usersrsquo potential preferences ereforehybrid features which include all kinds of item features are used to excavate usersrsquo interests In particular we find that the imagevisual features can catch more potential preferences of users In this paper we leverage the combination of user-item rating dataand item hybrid features to propose a novel CB recommendation model which is suitable for rating-based recommenderscenarios e experimental results show that the proposed model has better recommendation performance in sparse datascenarios than conventional approaches Besides training offline and recommendation online make the model has higher ef-ficiency on large datasets

1 Introduction

With the vigorous development of the Internet a largevolume of data is generated everyday People face an arduoustask of finding optimal information which matches theirpreferences from such tremendous amount of informationRSs use data mining and information filtering techniques torecommend items to potential users according to theirpreferences [1] and have been regarded as an important toolto solve the severe problem of information overload [2]Meanwhile business is benefitted with a growth of sales

Generally RS can be categorized as three mainbranchesmdashCB CF and hybrid algorithms [3] CB ap-proaches exploit the content features of items to recommendthe relevant items based on the usersrsquo preferences Mostcontent features are textual features extracted from webpages and product descriptions or tagged by users Unlikestructured data there are too few attributes with well-defined values [4] Besides CB has shortcomings of thenovelty of recommendation and others [5] CF predicts userpreferences based on the known user ratings of items

However when user ratings are scarce it is not sufficient toreflect usersrsquo preferences In such scenario it is unreliable tofind neighbors through user ratings Moreover with theincrease of the interactions between users and systems thegrowth of the rating matrix is extraordinarily fast It needs tocost considerable time to find usersrsquo neighbors on the wholedataset In addition CF suffers from the new user cold-startproblem [6] erefore RSs usually combine CB and CF toovercome each otherrsquos shortcomings [7]

Most RSs typically exploit textual features in order togenerate item recommendation they usually ignore thepositive effects brought by visual features However someresearchers have achieved a huge success in the study ofvisual features in RSs e paper [8] written and researchedby Deldjoo et al uses mise-en-scene visual features based onMPEG-7 and deep learning for movie recommendationeir work in [9 10] discusses video recommendation basedon low-level features and visual features And their work in[11 12] discusses visual features in movie recommendationIn addition Boutemedjet and Ziou [13] propose a graphicalmodel for context-aware visual content recommendation

HindawiScientific ProgrammingVolume 2018 Article ID 5497070 8 pageshttpsdoiorg10115520185497070

Melo et al [14] discuss content-based filtering enhanced byhuman visual attention applied to clothing recommenda-tion Filho et al [15] leverage deep visual features in content-based movie recommender systems eir previous worksinspired us to use image visual features of items to exploreusersrsquo potential preferences for recommendation

When using RS-based websites a majority of peoplechoose items according to itemsrsquo image visual features suchas colors shapes and textures Besides content features andsome other features of items can reflect usersrsquo potentialpreferences to certain extent Due to these reasons in theproposed approach we combine user ratings and hybridfeatures which include all kinds of item features discussedabove In particular by transforming user-item ratings to theratings of hybrid features the usersrsquo potential preferences ofhybrid features are used to find the nearest neighbors In thisway the data sparsity problem and the low efficiencyproblem are solved e primary contributions of this workare listed below

(i) Image visual features are taken into considerationcreatively By transforming user-item ratings to theratings of features we calculate usersrsquo potentialpreferences of hybrid features which include itemcontent features image visual features and so on

(ii) A novel recommendation model based on CB al-gorithms is proposed It is a generic recommen-dation model which is suitable for rating-basedrecommender scenarios

(iii) Plentiful of experiments are performed on the real-world datasets to evaluate the proposed model eresults show that our approach has better recom-mendation performance and higher efficiency

e structure of this paper is organized as followsSection 2 introduces our dataset and the feature descriptionA detailed explanation of the proposed model is given inSection 3 Following that the experimental evaluation andresults are shown in Section 4 Finally we conclude the studyin Section 5

2 Dataset and Feature Description

In this section it introduces the dataset we used and thedescription of hybrid features we used in this work

21 Dataset e dataset used in this paper is based ona public dataset hetrec2011-movielens-2k [16] which is anextension of the original MovieLens 10M dataset [17]published by GroupLens Research Group at the Universityof Minnesota

Table 1 shows some statistics about the hetrec2011-movielens-2k dataset and our extension To summarizethere are 2113 users 10197 movies and 855598 ratingsranging from 05 to 50 in increments of 05 ere is anaverage of 404921 ratings per user and 84637 ratings permoviee density of the dataset is 397ere are a total of13222 unique tags which fall into 47957 tag assignmenttuples of the form (user tag movie) Besides the moviesrsquo

cultural backgrounds are classified into 72 countries and 20genres In the dataset extension it referenced the corre-sponding web pages of movies at the IMDBwebsite [18] and9189 moviesrsquo posters are downloaded from the URLs ofIMDB Some of the URLs have errors in accessing them sothat we could not get every moviesrsquo posters For the fairnessof the experiments we removed the corresponding movieswhich do not contain posters So there are 9189 movies usedin experiments in total More information about the formatand statistics of the data is available at the hetrec2011-movielens-2k website [19]

22 Feature Description In this paper we primarily dividethe hybrid features into three main partsmdasheditorial featuresuser-generated features and image visual features Editorialfeatures are extracted from the textual information of itemsand user-generated features are extracted from the in-teractions between users and systems Specifically in theabove dataset editorial features include all kinds of contentfeatures of items user-generated features include user tagsmainly and image visual features are image features ofmovie posters Next the details about these kinds of featureswill be given

221 Editorial Features In our dataset editorial features aremainly content features of movies specifically moviecontent features consist of movie genres movie actorslanguages and so on As shown in Table 1 movie genres andcountries are chosen as the editorial features of moviesReasonably movie genres have a deep influence on choosingmovies to many people Some people might take a keeninterest in horror movies while others might never watchthem Similarly the countries of movies also have effects onpeoplersquos preferences

222 User-Generated Features In our dataset mainly usertags are the reflection of user interactions e movies thatusers have tagged can reflect that users are attracted by thesemovies to a certain degree Whether the content of tags isgood or bad these tags are something relevant to the moviesespecially the tags that many users have taggederefore bycalculating the term frequency (TF) of each tag we choosetags that are tagged more than 5 times by users to enrich thefeatures of movies In total 1245 tags are chosen as the user-generated features

Table 1 Summary statistics of our datasetNumber of users 2113Number of movies 10197Number of ratings 855598Number of countries 72Number of tags 13222Number of tag assignments 47957Number of movie genres 20Number of movie posters 9189

2 Scientific Programming

223 Image Visual Features When users browse an item onthe Internet in most cases the first thing that catches theirattention is the picture of the item especially for moviessufficient potential information can be revealed from themovie posters For example horror moviesrsquo posters couldlook bleak and cold romantic movies might have warm-toned posters erefore there is a possibility that somepeople may choose movies by their preferences of movieposters

However movie posters are very complex diversifiedand heterogeneous so that it is very difficult to extract sometypical features except for the postersrsquo dominant colorwhich is quite discriminating erefore the dominant hueof each porter is extracted through the RGB color modelMore specifically we calculate the proportion of the differentcolor pixels in the whole image After that as shown inTable 2 the movie posters are classified into 8 categories bythe range of their RGB values and each category can attractthe attention of a specific kind of people

3 Recommendation Model Based onHybrid Features

In conventional CF recommendation algorithms user-itemrating data are usually used to calculate the usersrsquo similaritymatrices and the nearest neighbors However they alwayssuffer from the problems of data sparsity and low efficiencyas described in Section 1 To solve the data sparsity problemwe consider that more latent information would be importedto build usersrsquo similarity matrices According to CB rec-ommendation algorithms instead of building usersrsquo simi-larity matrices by user-item ratings the proposedmodel usesusersrsquo potential interests of hybrid features to calculate thenearest neighbors In particular it transforms user-itemratings into the ratings of hybrid features and then calcu-lates usersrsquo potential interest values of each feature to builda user profile which is the vector representation of userinterests in the feature space spanned In this way userpreferences are reflected from many-sided featuresrsquo ratingdata when user ratings are scarce Moreover when user-itemrating matrix grows extremely large the proposed modelwhich uses hybrid featuresrsquo rating matrix can greatly reducethe time consumption because the number of features ismuch less than items so the hybrid featuresrsquo rating matrix ismuch smaller than the user-item rating matrix

As shown in Figure 1 the detailed workflow of theproposed model is divided into Feature Interest Measureprocess and Recommendation process Firstly our modelinput various kinds of featuresrsquo matrices and user-itemrating matrix And we combine user-item rating matrixwith the input featuresrsquo matrices to build several hybridfeaturesrsquo rating matrices by converting user-item ratings intohybrid featuresrsquo ratings Next we calculate usersrsquo potentialinterest values of each feature to generate a user-featureinterest measure matrix In the above Feature InterestMeasure process all of these calculations and trans-formations are completely offline After that in the onlineRecommendation process the user similarity matrix andneighbor set are trained and generated based on user-feature

interest measure matrix Finally the predicted rating valuesof items are calculated through the known user ratings andneighbor set

31 Feature Interest Measure We have introduced ourdataset in Section 2 the hybrid features include 20 moviegenres 72 movie countries 8 movie poster styles and 1245movie tags Let Gi(i 1 2 20) represent 20 moviegenres Ci(i 1 2 72) represent 72 movie countriesSi(i 1 2 8) represent 8 movie poster styles and Ti(i

1 2 1 245) represent 1245 movie tags e format andexample of the hybrid featuresrsquo matrix is shown in Table 3e value 1 means the movie includes the feature and thevalue 0 means it does not Apparently the ratings of moviescan be converted to the ratings of features e problem isthat there are much more movies than features and theratings of features are probably more than once in mostcases So it needs to use an exact value to reflect usersrsquo realinterest of a feature as much as possible And the example ofuser-feature interest measure matrix is shown in Table 4 Wewere inspired by a calculation method proposed by Wanget al [20] to propose our calculation methods Here wedescribe some related concepts and definitions which areused in our calculations e specific calculation processesare as follows

(i) Feature rating ratio It is defined as the ratio ofuser ursquos total effective rating values for feature k touser ursquos total rating values which is expressed asFRR(u k)

FRR(u k) 1113936iisinIksubIurge σ2rui

1113936iisinIurui

(1)

where Iu is the set of items rated by user u Ik is the set ofitems with feature k rui denotes user ursquos rating for item i andσ denotes the maximum possible rating value in system Inthe model items with the rating larger than σ2 are regardedas the ones that users are interested FRR can reflect howmuch is user u interested in feature k to some extent And thedenominator of FRR can avoid the negative effect brought bydifferent usersrsquo liveness (the amount of ratings)

(ii) Feature rating frequency ratio It is defined as theratio of user ursquos total effective rating times for feature

Table 2 e categories of the movie posters

e range of RGB values Poster styleR isin [0 128) G isin [0 128) B isin [0 128) DarkR isin [0 128) G isin [0 128) B isin [128 256) Deep blueR isin [0 128) G isin [128 256) B isin [0 128) GreenR isin [128 256) G isin [0 128) B isin [0 128) RedR isin [0 128) G isin [128 256) B isin [128 256) Light blueR isin [128 256) G isin [0 128) B isin [128 256) PurpleR isin [128 256) G isin [128 256) B isin [0 128) YellowR isin [128 256) G isin [128 256) B isin [128 256) White

Scientific Programming 3

k to user ursquos total rating times which is expressed asFRFR(u k)

FRFR(u k) 1113936iisinIksubIu

δ rui( 1113857

Tu

11138681113868111386811138681113868111386811138681113868

δ rui( 1113857

1 rui geσ2

0 rui ltσ2

⎧⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎩

(2)

Here user ursquos total rating times |Tu| can remove thenegative effect introduced by the liveness of different usersFRFR also can reflect user ursquos interest for feature k toa certain extent

However FRFR calculates all effective rating times withthe same weight Actually user preferences are also reflectedby the rating values So we modified (2) to reflect userpreferences more reliable as below

(iii) Weighted feature rating frequency ratio It isexpressed as WFRFR(u k)

WFRFR(u k) 1113936iisinIksubIu

ωr lowast δ rui( 1113857

σ2lowast Tu

11138681113868111386811138681113868111386811138681113868

(3)

where

ωr rui minusσ2

+ 1 (4)

where ωr is the weight value of δ(rui) which is calculatedthrough user rating rui In this way WFRFR can overcomethe shortcomings of FRFR

(iv) Feature interest measure It is defined as user ursquosinterest measure value for feature k which isexpressed as FIM(u k)

FIM(u k) 2σ lowast FRR(u k)lowastWFRFR(u k)

FRR(u k) + WFRFR(u k) (5)

where FIM is the harmonic mean value of FRR andWFRFRe maximum possible rating value σ is the normalizationfactor to make the value of FIM ranges from 0 to σ so thiscalculation method is suitable for all kinds of rating-basedRS

32 Recommendation After the generation of user-featureinterest measure matrix offline the proposed model pro-cesses the recommendation online Firstly as shown in (6)by employing Pearson correlation coefficient [21] thesimilarity of users u and v is calculated through their interestvalues of all the features where ruk denotes FIM(u k) whichis the interest value of user u for feature k ru denotes theaverage value of user ursquos interest values and n is the numberof features co-rated by users u and v

sim(u v) 1113936

nk1 ruk minus ru( 1113857 rvk minus rv( 1113857

1113936nk1 ruk minus ru( 1113857

21113936

ni1 rvk minus rv( 1113857

21113969 (6)

When the similarity calculation is completed usersimilarity matrix and neighbor set are generated Based onthe similarity matrix and neighbor set the user-item ratingmatrix is combined to calculate the predicted ratings ofitems [21] 1113954rui denotes user ursquos predicted rating for item ie calculation method of 1113954rui is as follows

Table 4 Feature interest measure matrixUser G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

u1 35 1 25 35 2 45 45 1 2 middot middot middot

u2 45 2 05 15 4 15 25 4 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮un 25 2 15 05 4 35 35 3 2 middot middot middot

Table 3 Input featuresrsquo matrixMovie G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

m1 1 0 1 1 0 0 1 0 0 middot middot middot

m2 0 1 0 0 0 1 0 0 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮mk 1 1 0 0 1 1 0 0 0 middot middot middot

Image visual featuresDynamic features

Static features

Use

r (ra

ting)

Use

r (ra

ting)

Static featuresrsquomatrix

Dynamic featuresrsquomatrix

Image visualfeaturesrsquo matrix

User-itemrating matrix

Input feature matrices

1 Feature interest measure (offline) 2 Recommendation (online)

Item ratingprediction

Use

r (in

tere

st va

lue)

Use

r (ra

ting)

Hybrid features

User-feature interestmeasure matrix

Hybrid featuresrsquo rating matrix

User-itemrating matrix

Similaritymatrix

Neighbor set

Training results

Figure 1 e workflow of the proposed model

4 Scientific Programming

1113954rui ru +1113936

nv1 sim(u v) rvi minus rv( 1113857

1113936nv1|sim(u v)|

(7)

where ru denotes the average rating of user u v is a neighboruser of user u and there are n neighbors in total rvi denotesthe rating of user v for item i

4 Experimental Evaluation

In this section it describes how we evaluate the proposedmodel on the real dataset In order to construct severalsparse data scenarios in the evaluation we randomly selectuser rating data from the full dataset described in Section 2while keeping the rating number proportion of each userunchanged Two methods are used to generate the sparsedatasets one is that the average user rating number changesfrom 10 to 80 in increments of 10 the other is that theaverage user rating number and user number change from25 of full dataset to 100 in increments of 25

41 Experimental Setup e experiments are performed onan operating system of Windows 10 Intel(R) Core(TM) i5-6500 CPU 320GHz and 8GB RAM Primarily user-based K-nearest neighbors (UserKNN) algorithm [22] ischosen as the baseline By importing the proposed model toKNN it is called hybrid feature-based KNN (HFB-KNN)algorithm In addition if the proposed model only uses thecontent features for comparison it is called content feature-based KNN (CFB-KNN) algorithm To evaluate the per-formance of above algorithms to avoid bias we used 10-foldcross validation to avoid any fortunate occurrences and theexperiments are deployed as follows

(1) Selection of user neighbor number

To have a better performance by varying the userneighbor number and setting userrsquos average rating numberas a fixed value (eg 60 ratings) the above algorithms areevaluated to find the optimal user neighbor number

(2) Experiments on sparse datasets

To compare the recommendation performance byvarying the average user rating number from 10 to 80 inincrements of 10 the above algorithms are evaluated on 8sparse datasets

(3) Comparison of recommendation time

To compare the recommendation time by varying theaverage user rating number and user number from 25 of fulldataset to 100 in increments of 25 the above algorithmsare evaluated on 2 groups of different-scale datasets

42 EvaluationMetrics In order to verify the advance of theproposed model mean absolute error (MAE) and rootmean-squared error (RMSE) are adopted as the evaluationmetrics According to Herlocker et al [23] by computing thevalue distinction between predicted values and real valuesMAE and RMSE are usually used to evaluate predictiveaccuracy e smaller values of the MAE and RMSE indicate

a better performance in recommendation because theyamplify the contributions of the absolute errors between thepredictions and the true ratings Assuming ri is the predictedrating 1113954ri is the actual rating n is the total number of ratingsfrom all users MAE and RMSE are defined as follows

MAE 1n

1113944

n

i1ri minus 1113954ri

11138681113868111386811138681113868111386811138681113868

RMSE

1n

1113944

n

i1ri minus 1113954ri( 1113857

2

11139741113972

(8)

43 Results

431 Selection of User Neighbor Number Figure 2 shows theresults of MAE and RMSE with different numbers of userneighbors It can be noticed that the proposed model hasbetter rating prediction accuracy because HFB-KNNrsquos MAEand RMSE values are always smallest in all the scenariosof different user neighbor numbers In addition the numberof neighbor is set as 70 in the following experiments be-cause such configuration leads to better recommendationperformance

432 Experiments on Sparse Datasets Figure 3 shows thecomparison results of rating prediction accuracy on 8sparse datasets Obviously the prediction accuracy of 3algorithms gets better when the average number of userrating changes from 10 to 80 and HFB-KNN outperformsboth UserKNN and CFB-KNN Moreover when the dataare quite sparse (eg the average number of user rating isless than 60) it should be noticed that the performance ofHFB-KNN and CFB-KNN is dramatically better than thatof UserKNN is is because both HFB-KNN and CFB-KNN are based on the proposed model which can enrichthe feature matrix and get more latent information of userpreferences in the sparse data scenarios e reason whyHFB-KNN outperforms CFB-KNN is that CFB-KNN onlyexploits content features however HFB-KNN leverageseditorial features user-generated features and image visualfeatures

433 Comparison of Recommendation Time Figure 4 showsthe comparison of recommendation time on 2 groups ofdifferent-scale datasets Figure 4(a) shows the scenario ofvarying the average user rating number from 14 of all userratings to the full dataset while Figure 4(b) shows thescenario of varying the user number from 14 of all users tothe full dataset We can find that all the three algorithms takemore recommendation time with the increase of datasetscale However HFB-KNN and CFB-KNN take much lesstime than UserKNN in the two scenariosus it verifies thehigher efficiency of the proposed model In particular itshould be noticed that the larger scale the dataset is the moresuperior in time consumption the proposed model is

Scientific Programming 5

068

0675

067

0665

066

0655

065

MA

E

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(a)

089

0885

088

0875

087

0865

086

0855

RMSE

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(b)

Figure 2 e comparison of rating prediction accuracy with different user neighbor numbers

085

08

075

07

065

06

MA

E

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

(a)

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

105

1

095

09

085

08

RMSE

(b)

Figure 3 e comparison of rating prediction accuracy on sparse datasets

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(a)

Figure 4 Continued

6 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

223 Image Visual Features When users browse an item onthe Internet in most cases the first thing that catches theirattention is the picture of the item especially for moviessufficient potential information can be revealed from themovie posters For example horror moviesrsquo posters couldlook bleak and cold romantic movies might have warm-toned posters erefore there is a possibility that somepeople may choose movies by their preferences of movieposters

However movie posters are very complex diversifiedand heterogeneous so that it is very difficult to extract sometypical features except for the postersrsquo dominant colorwhich is quite discriminating erefore the dominant hueof each porter is extracted through the RGB color modelMore specifically we calculate the proportion of the differentcolor pixels in the whole image After that as shown inTable 2 the movie posters are classified into 8 categories bythe range of their RGB values and each category can attractthe attention of a specific kind of people

3 Recommendation Model Based onHybrid Features

In conventional CF recommendation algorithms user-itemrating data are usually used to calculate the usersrsquo similaritymatrices and the nearest neighbors However they alwayssuffer from the problems of data sparsity and low efficiencyas described in Section 1 To solve the data sparsity problemwe consider that more latent information would be importedto build usersrsquo similarity matrices According to CB rec-ommendation algorithms instead of building usersrsquo simi-larity matrices by user-item ratings the proposedmodel usesusersrsquo potential interests of hybrid features to calculate thenearest neighbors In particular it transforms user-itemratings into the ratings of hybrid features and then calcu-lates usersrsquo potential interest values of each feature to builda user profile which is the vector representation of userinterests in the feature space spanned In this way userpreferences are reflected from many-sided featuresrsquo ratingdata when user ratings are scarce Moreover when user-itemrating matrix grows extremely large the proposed modelwhich uses hybrid featuresrsquo rating matrix can greatly reducethe time consumption because the number of features ismuch less than items so the hybrid featuresrsquo rating matrix ismuch smaller than the user-item rating matrix

As shown in Figure 1 the detailed workflow of theproposed model is divided into Feature Interest Measureprocess and Recommendation process Firstly our modelinput various kinds of featuresrsquo matrices and user-itemrating matrix And we combine user-item rating matrixwith the input featuresrsquo matrices to build several hybridfeaturesrsquo rating matrices by converting user-item ratings intohybrid featuresrsquo ratings Next we calculate usersrsquo potentialinterest values of each feature to generate a user-featureinterest measure matrix In the above Feature InterestMeasure process all of these calculations and trans-formations are completely offline After that in the onlineRecommendation process the user similarity matrix andneighbor set are trained and generated based on user-feature

interest measure matrix Finally the predicted rating valuesof items are calculated through the known user ratings andneighbor set

31 Feature Interest Measure We have introduced ourdataset in Section 2 the hybrid features include 20 moviegenres 72 movie countries 8 movie poster styles and 1245movie tags Let Gi(i 1 2 20) represent 20 moviegenres Ci(i 1 2 72) represent 72 movie countriesSi(i 1 2 8) represent 8 movie poster styles and Ti(i

1 2 1 245) represent 1245 movie tags e format andexample of the hybrid featuresrsquo matrix is shown in Table 3e value 1 means the movie includes the feature and thevalue 0 means it does not Apparently the ratings of moviescan be converted to the ratings of features e problem isthat there are much more movies than features and theratings of features are probably more than once in mostcases So it needs to use an exact value to reflect usersrsquo realinterest of a feature as much as possible And the example ofuser-feature interest measure matrix is shown in Table 4 Wewere inspired by a calculation method proposed by Wanget al [20] to propose our calculation methods Here wedescribe some related concepts and definitions which areused in our calculations e specific calculation processesare as follows

(i) Feature rating ratio It is defined as the ratio ofuser ursquos total effective rating values for feature k touser ursquos total rating values which is expressed asFRR(u k)

FRR(u k) 1113936iisinIksubIurge σ2rui

1113936iisinIurui

(1)

where Iu is the set of items rated by user u Ik is the set ofitems with feature k rui denotes user ursquos rating for item i andσ denotes the maximum possible rating value in system Inthe model items with the rating larger than σ2 are regardedas the ones that users are interested FRR can reflect howmuch is user u interested in feature k to some extent And thedenominator of FRR can avoid the negative effect brought bydifferent usersrsquo liveness (the amount of ratings)

(ii) Feature rating frequency ratio It is defined as theratio of user ursquos total effective rating times for feature

Table 2 e categories of the movie posters

e range of RGB values Poster styleR isin [0 128) G isin [0 128) B isin [0 128) DarkR isin [0 128) G isin [0 128) B isin [128 256) Deep blueR isin [0 128) G isin [128 256) B isin [0 128) GreenR isin [128 256) G isin [0 128) B isin [0 128) RedR isin [0 128) G isin [128 256) B isin [128 256) Light blueR isin [128 256) G isin [0 128) B isin [128 256) PurpleR isin [128 256) G isin [128 256) B isin [0 128) YellowR isin [128 256) G isin [128 256) B isin [128 256) White

Scientific Programming 3

k to user ursquos total rating times which is expressed asFRFR(u k)

FRFR(u k) 1113936iisinIksubIu

δ rui( 1113857

Tu

11138681113868111386811138681113868111386811138681113868

δ rui( 1113857

1 rui geσ2

0 rui ltσ2

⎧⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎩

(2)

Here user ursquos total rating times |Tu| can remove thenegative effect introduced by the liveness of different usersFRFR also can reflect user ursquos interest for feature k toa certain extent

However FRFR calculates all effective rating times withthe same weight Actually user preferences are also reflectedby the rating values So we modified (2) to reflect userpreferences more reliable as below

(iii) Weighted feature rating frequency ratio It isexpressed as WFRFR(u k)

WFRFR(u k) 1113936iisinIksubIu

ωr lowast δ rui( 1113857

σ2lowast Tu

11138681113868111386811138681113868111386811138681113868

(3)

where

ωr rui minusσ2

+ 1 (4)

where ωr is the weight value of δ(rui) which is calculatedthrough user rating rui In this way WFRFR can overcomethe shortcomings of FRFR

(iv) Feature interest measure It is defined as user ursquosinterest measure value for feature k which isexpressed as FIM(u k)

FIM(u k) 2σ lowast FRR(u k)lowastWFRFR(u k)

FRR(u k) + WFRFR(u k) (5)

where FIM is the harmonic mean value of FRR andWFRFRe maximum possible rating value σ is the normalizationfactor to make the value of FIM ranges from 0 to σ so thiscalculation method is suitable for all kinds of rating-basedRS

32 Recommendation After the generation of user-featureinterest measure matrix offline the proposed model pro-cesses the recommendation online Firstly as shown in (6)by employing Pearson correlation coefficient [21] thesimilarity of users u and v is calculated through their interestvalues of all the features where ruk denotes FIM(u k) whichis the interest value of user u for feature k ru denotes theaverage value of user ursquos interest values and n is the numberof features co-rated by users u and v

sim(u v) 1113936

nk1 ruk minus ru( 1113857 rvk minus rv( 1113857

1113936nk1 ruk minus ru( 1113857

21113936

ni1 rvk minus rv( 1113857

21113969 (6)

When the similarity calculation is completed usersimilarity matrix and neighbor set are generated Based onthe similarity matrix and neighbor set the user-item ratingmatrix is combined to calculate the predicted ratings ofitems [21] 1113954rui denotes user ursquos predicted rating for item ie calculation method of 1113954rui is as follows

Table 4 Feature interest measure matrixUser G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

u1 35 1 25 35 2 45 45 1 2 middot middot middot

u2 45 2 05 15 4 15 25 4 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮un 25 2 15 05 4 35 35 3 2 middot middot middot

Table 3 Input featuresrsquo matrixMovie G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

m1 1 0 1 1 0 0 1 0 0 middot middot middot

m2 0 1 0 0 0 1 0 0 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮mk 1 1 0 0 1 1 0 0 0 middot middot middot

Image visual featuresDynamic features

Static features

Use

r (ra

ting)

Use

r (ra

ting)

Static featuresrsquomatrix

Dynamic featuresrsquomatrix

Image visualfeaturesrsquo matrix

User-itemrating matrix

Input feature matrices

1 Feature interest measure (offline) 2 Recommendation (online)

Item ratingprediction

Use

r (in

tere

st va

lue)

Use

r (ra

ting)

Hybrid features

User-feature interestmeasure matrix

Hybrid featuresrsquo rating matrix

User-itemrating matrix

Similaritymatrix

Neighbor set

Training results

Figure 1 e workflow of the proposed model

4 Scientific Programming

1113954rui ru +1113936

nv1 sim(u v) rvi minus rv( 1113857

1113936nv1|sim(u v)|

(7)

where ru denotes the average rating of user u v is a neighboruser of user u and there are n neighbors in total rvi denotesthe rating of user v for item i

4 Experimental Evaluation

In this section it describes how we evaluate the proposedmodel on the real dataset In order to construct severalsparse data scenarios in the evaluation we randomly selectuser rating data from the full dataset described in Section 2while keeping the rating number proportion of each userunchanged Two methods are used to generate the sparsedatasets one is that the average user rating number changesfrom 10 to 80 in increments of 10 the other is that theaverage user rating number and user number change from25 of full dataset to 100 in increments of 25

41 Experimental Setup e experiments are performed onan operating system of Windows 10 Intel(R) Core(TM) i5-6500 CPU 320GHz and 8GB RAM Primarily user-based K-nearest neighbors (UserKNN) algorithm [22] ischosen as the baseline By importing the proposed model toKNN it is called hybrid feature-based KNN (HFB-KNN)algorithm In addition if the proposed model only uses thecontent features for comparison it is called content feature-based KNN (CFB-KNN) algorithm To evaluate the per-formance of above algorithms to avoid bias we used 10-foldcross validation to avoid any fortunate occurrences and theexperiments are deployed as follows

(1) Selection of user neighbor number

To have a better performance by varying the userneighbor number and setting userrsquos average rating numberas a fixed value (eg 60 ratings) the above algorithms areevaluated to find the optimal user neighbor number

(2) Experiments on sparse datasets

To compare the recommendation performance byvarying the average user rating number from 10 to 80 inincrements of 10 the above algorithms are evaluated on 8sparse datasets

(3) Comparison of recommendation time

To compare the recommendation time by varying theaverage user rating number and user number from 25 of fulldataset to 100 in increments of 25 the above algorithmsare evaluated on 2 groups of different-scale datasets

42 EvaluationMetrics In order to verify the advance of theproposed model mean absolute error (MAE) and rootmean-squared error (RMSE) are adopted as the evaluationmetrics According to Herlocker et al [23] by computing thevalue distinction between predicted values and real valuesMAE and RMSE are usually used to evaluate predictiveaccuracy e smaller values of the MAE and RMSE indicate

a better performance in recommendation because theyamplify the contributions of the absolute errors between thepredictions and the true ratings Assuming ri is the predictedrating 1113954ri is the actual rating n is the total number of ratingsfrom all users MAE and RMSE are defined as follows

MAE 1n

1113944

n

i1ri minus 1113954ri

11138681113868111386811138681113868111386811138681113868

RMSE

1n

1113944

n

i1ri minus 1113954ri( 1113857

2

11139741113972

(8)

43 Results

431 Selection of User Neighbor Number Figure 2 shows theresults of MAE and RMSE with different numbers of userneighbors It can be noticed that the proposed model hasbetter rating prediction accuracy because HFB-KNNrsquos MAEand RMSE values are always smallest in all the scenariosof different user neighbor numbers In addition the numberof neighbor is set as 70 in the following experiments be-cause such configuration leads to better recommendationperformance

432 Experiments on Sparse Datasets Figure 3 shows thecomparison results of rating prediction accuracy on 8sparse datasets Obviously the prediction accuracy of 3algorithms gets better when the average number of userrating changes from 10 to 80 and HFB-KNN outperformsboth UserKNN and CFB-KNN Moreover when the dataare quite sparse (eg the average number of user rating isless than 60) it should be noticed that the performance ofHFB-KNN and CFB-KNN is dramatically better than thatof UserKNN is is because both HFB-KNN and CFB-KNN are based on the proposed model which can enrichthe feature matrix and get more latent information of userpreferences in the sparse data scenarios e reason whyHFB-KNN outperforms CFB-KNN is that CFB-KNN onlyexploits content features however HFB-KNN leverageseditorial features user-generated features and image visualfeatures

433 Comparison of Recommendation Time Figure 4 showsthe comparison of recommendation time on 2 groups ofdifferent-scale datasets Figure 4(a) shows the scenario ofvarying the average user rating number from 14 of all userratings to the full dataset while Figure 4(b) shows thescenario of varying the user number from 14 of all users tothe full dataset We can find that all the three algorithms takemore recommendation time with the increase of datasetscale However HFB-KNN and CFB-KNN take much lesstime than UserKNN in the two scenariosus it verifies thehigher efficiency of the proposed model In particular itshould be noticed that the larger scale the dataset is the moresuperior in time consumption the proposed model is

Scientific Programming 5

068

0675

067

0665

066

0655

065

MA

E

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(a)

089

0885

088

0875

087

0865

086

0855

RMSE

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(b)

Figure 2 e comparison of rating prediction accuracy with different user neighbor numbers

085

08

075

07

065

06

MA

E

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

(a)

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

105

1

095

09

085

08

RMSE

(b)

Figure 3 e comparison of rating prediction accuracy on sparse datasets

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(a)

Figure 4 Continued

6 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

k to user ursquos total rating times which is expressed asFRFR(u k)

FRFR(u k) 1113936iisinIksubIu

δ rui( 1113857

Tu

11138681113868111386811138681113868111386811138681113868

δ rui( 1113857

1 rui geσ2

0 rui ltσ2

⎧⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎩

(2)

Here user ursquos total rating times |Tu| can remove thenegative effect introduced by the liveness of different usersFRFR also can reflect user ursquos interest for feature k toa certain extent

However FRFR calculates all effective rating times withthe same weight Actually user preferences are also reflectedby the rating values So we modified (2) to reflect userpreferences more reliable as below

(iii) Weighted feature rating frequency ratio It isexpressed as WFRFR(u k)

WFRFR(u k) 1113936iisinIksubIu

ωr lowast δ rui( 1113857

σ2lowast Tu

11138681113868111386811138681113868111386811138681113868

(3)

where

ωr rui minusσ2

+ 1 (4)

where ωr is the weight value of δ(rui) which is calculatedthrough user rating rui In this way WFRFR can overcomethe shortcomings of FRFR

(iv) Feature interest measure It is defined as user ursquosinterest measure value for feature k which isexpressed as FIM(u k)

FIM(u k) 2σ lowast FRR(u k)lowastWFRFR(u k)

FRR(u k) + WFRFR(u k) (5)

where FIM is the harmonic mean value of FRR andWFRFRe maximum possible rating value σ is the normalizationfactor to make the value of FIM ranges from 0 to σ so thiscalculation method is suitable for all kinds of rating-basedRS

32 Recommendation After the generation of user-featureinterest measure matrix offline the proposed model pro-cesses the recommendation online Firstly as shown in (6)by employing Pearson correlation coefficient [21] thesimilarity of users u and v is calculated through their interestvalues of all the features where ruk denotes FIM(u k) whichis the interest value of user u for feature k ru denotes theaverage value of user ursquos interest values and n is the numberof features co-rated by users u and v

sim(u v) 1113936

nk1 ruk minus ru( 1113857 rvk minus rv( 1113857

1113936nk1 ruk minus ru( 1113857

21113936

ni1 rvk minus rv( 1113857

21113969 (6)

When the similarity calculation is completed usersimilarity matrix and neighbor set are generated Based onthe similarity matrix and neighbor set the user-item ratingmatrix is combined to calculate the predicted ratings ofitems [21] 1113954rui denotes user ursquos predicted rating for item ie calculation method of 1113954rui is as follows

Table 4 Feature interest measure matrixUser G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

u1 35 1 25 35 2 45 45 1 2 middot middot middot

u2 45 2 05 15 4 15 25 4 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮un 25 2 15 05 4 35 35 3 2 middot middot middot

Table 3 Input featuresrsquo matrixMovie G1 G2 G20 C1 C2 C72 S1 S2 S8 middot middot middot

m1 1 0 1 1 0 0 1 0 0 middot middot middot

m2 0 1 0 0 0 1 0 0 1 middot middot middot

⋮ ⋮ ⋮ ⋮ ⋮mk 1 1 0 0 1 1 0 0 0 middot middot middot

Image visual featuresDynamic features

Static features

Use

r (ra

ting)

Use

r (ra

ting)

Static featuresrsquomatrix

Dynamic featuresrsquomatrix

Image visualfeaturesrsquo matrix

User-itemrating matrix

Input feature matrices

1 Feature interest measure (offline) 2 Recommendation (online)

Item ratingprediction

Use

r (in

tere

st va

lue)

Use

r (ra

ting)

Hybrid features

User-feature interestmeasure matrix

Hybrid featuresrsquo rating matrix

User-itemrating matrix

Similaritymatrix

Neighbor set

Training results

Figure 1 e workflow of the proposed model

4 Scientific Programming

1113954rui ru +1113936

nv1 sim(u v) rvi minus rv( 1113857

1113936nv1|sim(u v)|

(7)

where ru denotes the average rating of user u v is a neighboruser of user u and there are n neighbors in total rvi denotesthe rating of user v for item i

4 Experimental Evaluation

In this section it describes how we evaluate the proposedmodel on the real dataset In order to construct severalsparse data scenarios in the evaluation we randomly selectuser rating data from the full dataset described in Section 2while keeping the rating number proportion of each userunchanged Two methods are used to generate the sparsedatasets one is that the average user rating number changesfrom 10 to 80 in increments of 10 the other is that theaverage user rating number and user number change from25 of full dataset to 100 in increments of 25

41 Experimental Setup e experiments are performed onan operating system of Windows 10 Intel(R) Core(TM) i5-6500 CPU 320GHz and 8GB RAM Primarily user-based K-nearest neighbors (UserKNN) algorithm [22] ischosen as the baseline By importing the proposed model toKNN it is called hybrid feature-based KNN (HFB-KNN)algorithm In addition if the proposed model only uses thecontent features for comparison it is called content feature-based KNN (CFB-KNN) algorithm To evaluate the per-formance of above algorithms to avoid bias we used 10-foldcross validation to avoid any fortunate occurrences and theexperiments are deployed as follows

(1) Selection of user neighbor number

To have a better performance by varying the userneighbor number and setting userrsquos average rating numberas a fixed value (eg 60 ratings) the above algorithms areevaluated to find the optimal user neighbor number

(2) Experiments on sparse datasets

To compare the recommendation performance byvarying the average user rating number from 10 to 80 inincrements of 10 the above algorithms are evaluated on 8sparse datasets

(3) Comparison of recommendation time

To compare the recommendation time by varying theaverage user rating number and user number from 25 of fulldataset to 100 in increments of 25 the above algorithmsare evaluated on 2 groups of different-scale datasets

42 EvaluationMetrics In order to verify the advance of theproposed model mean absolute error (MAE) and rootmean-squared error (RMSE) are adopted as the evaluationmetrics According to Herlocker et al [23] by computing thevalue distinction between predicted values and real valuesMAE and RMSE are usually used to evaluate predictiveaccuracy e smaller values of the MAE and RMSE indicate

a better performance in recommendation because theyamplify the contributions of the absolute errors between thepredictions and the true ratings Assuming ri is the predictedrating 1113954ri is the actual rating n is the total number of ratingsfrom all users MAE and RMSE are defined as follows

MAE 1n

1113944

n

i1ri minus 1113954ri

11138681113868111386811138681113868111386811138681113868

RMSE

1n

1113944

n

i1ri minus 1113954ri( 1113857

2

11139741113972

(8)

43 Results

431 Selection of User Neighbor Number Figure 2 shows theresults of MAE and RMSE with different numbers of userneighbors It can be noticed that the proposed model hasbetter rating prediction accuracy because HFB-KNNrsquos MAEand RMSE values are always smallest in all the scenariosof different user neighbor numbers In addition the numberof neighbor is set as 70 in the following experiments be-cause such configuration leads to better recommendationperformance

432 Experiments on Sparse Datasets Figure 3 shows thecomparison results of rating prediction accuracy on 8sparse datasets Obviously the prediction accuracy of 3algorithms gets better when the average number of userrating changes from 10 to 80 and HFB-KNN outperformsboth UserKNN and CFB-KNN Moreover when the dataare quite sparse (eg the average number of user rating isless than 60) it should be noticed that the performance ofHFB-KNN and CFB-KNN is dramatically better than thatof UserKNN is is because both HFB-KNN and CFB-KNN are based on the proposed model which can enrichthe feature matrix and get more latent information of userpreferences in the sparse data scenarios e reason whyHFB-KNN outperforms CFB-KNN is that CFB-KNN onlyexploits content features however HFB-KNN leverageseditorial features user-generated features and image visualfeatures

433 Comparison of Recommendation Time Figure 4 showsthe comparison of recommendation time on 2 groups ofdifferent-scale datasets Figure 4(a) shows the scenario ofvarying the average user rating number from 14 of all userratings to the full dataset while Figure 4(b) shows thescenario of varying the user number from 14 of all users tothe full dataset We can find that all the three algorithms takemore recommendation time with the increase of datasetscale However HFB-KNN and CFB-KNN take much lesstime than UserKNN in the two scenariosus it verifies thehigher efficiency of the proposed model In particular itshould be noticed that the larger scale the dataset is the moresuperior in time consumption the proposed model is

Scientific Programming 5

068

0675

067

0665

066

0655

065

MA

E

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(a)

089

0885

088

0875

087

0865

086

0855

RMSE

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(b)

Figure 2 e comparison of rating prediction accuracy with different user neighbor numbers

085

08

075

07

065

06

MA

E

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

(a)

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

105

1

095

09

085

08

RMSE

(b)

Figure 3 e comparison of rating prediction accuracy on sparse datasets

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(a)

Figure 4 Continued

6 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

1113954rui ru +1113936

nv1 sim(u v) rvi minus rv( 1113857

1113936nv1|sim(u v)|

(7)

where ru denotes the average rating of user u v is a neighboruser of user u and there are n neighbors in total rvi denotesthe rating of user v for item i

4 Experimental Evaluation

In this section it describes how we evaluate the proposedmodel on the real dataset In order to construct severalsparse data scenarios in the evaluation we randomly selectuser rating data from the full dataset described in Section 2while keeping the rating number proportion of each userunchanged Two methods are used to generate the sparsedatasets one is that the average user rating number changesfrom 10 to 80 in increments of 10 the other is that theaverage user rating number and user number change from25 of full dataset to 100 in increments of 25

41 Experimental Setup e experiments are performed onan operating system of Windows 10 Intel(R) Core(TM) i5-6500 CPU 320GHz and 8GB RAM Primarily user-based K-nearest neighbors (UserKNN) algorithm [22] ischosen as the baseline By importing the proposed model toKNN it is called hybrid feature-based KNN (HFB-KNN)algorithm In addition if the proposed model only uses thecontent features for comparison it is called content feature-based KNN (CFB-KNN) algorithm To evaluate the per-formance of above algorithms to avoid bias we used 10-foldcross validation to avoid any fortunate occurrences and theexperiments are deployed as follows

(1) Selection of user neighbor number

To have a better performance by varying the userneighbor number and setting userrsquos average rating numberas a fixed value (eg 60 ratings) the above algorithms areevaluated to find the optimal user neighbor number

(2) Experiments on sparse datasets

To compare the recommendation performance byvarying the average user rating number from 10 to 80 inincrements of 10 the above algorithms are evaluated on 8sparse datasets

(3) Comparison of recommendation time

To compare the recommendation time by varying theaverage user rating number and user number from 25 of fulldataset to 100 in increments of 25 the above algorithmsare evaluated on 2 groups of different-scale datasets

42 EvaluationMetrics In order to verify the advance of theproposed model mean absolute error (MAE) and rootmean-squared error (RMSE) are adopted as the evaluationmetrics According to Herlocker et al [23] by computing thevalue distinction between predicted values and real valuesMAE and RMSE are usually used to evaluate predictiveaccuracy e smaller values of the MAE and RMSE indicate

a better performance in recommendation because theyamplify the contributions of the absolute errors between thepredictions and the true ratings Assuming ri is the predictedrating 1113954ri is the actual rating n is the total number of ratingsfrom all users MAE and RMSE are defined as follows

MAE 1n

1113944

n

i1ri minus 1113954ri

11138681113868111386811138681113868111386811138681113868

RMSE

1n

1113944

n

i1ri minus 1113954ri( 1113857

2

11139741113972

(8)

43 Results

431 Selection of User Neighbor Number Figure 2 shows theresults of MAE and RMSE with different numbers of userneighbors It can be noticed that the proposed model hasbetter rating prediction accuracy because HFB-KNNrsquos MAEand RMSE values are always smallest in all the scenariosof different user neighbor numbers In addition the numberof neighbor is set as 70 in the following experiments be-cause such configuration leads to better recommendationperformance

432 Experiments on Sparse Datasets Figure 3 shows thecomparison results of rating prediction accuracy on 8sparse datasets Obviously the prediction accuracy of 3algorithms gets better when the average number of userrating changes from 10 to 80 and HFB-KNN outperformsboth UserKNN and CFB-KNN Moreover when the dataare quite sparse (eg the average number of user rating isless than 60) it should be noticed that the performance ofHFB-KNN and CFB-KNN is dramatically better than thatof UserKNN is is because both HFB-KNN and CFB-KNN are based on the proposed model which can enrichthe feature matrix and get more latent information of userpreferences in the sparse data scenarios e reason whyHFB-KNN outperforms CFB-KNN is that CFB-KNN onlyexploits content features however HFB-KNN leverageseditorial features user-generated features and image visualfeatures

433 Comparison of Recommendation Time Figure 4 showsthe comparison of recommendation time on 2 groups ofdifferent-scale datasets Figure 4(a) shows the scenario ofvarying the average user rating number from 14 of all userratings to the full dataset while Figure 4(b) shows thescenario of varying the user number from 14 of all users tothe full dataset We can find that all the three algorithms takemore recommendation time with the increase of datasetscale However HFB-KNN and CFB-KNN take much lesstime than UserKNN in the two scenariosus it verifies thehigher efficiency of the proposed model In particular itshould be noticed that the larger scale the dataset is the moresuperior in time consumption the proposed model is

Scientific Programming 5

068

0675

067

0665

066

0655

065

MA

E

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(a)

089

0885

088

0875

087

0865

086

0855

RMSE

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(b)

Figure 2 e comparison of rating prediction accuracy with different user neighbor numbers

085

08

075

07

065

06

MA

E

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

(a)

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

105

1

095

09

085

08

RMSE

(b)

Figure 3 e comparison of rating prediction accuracy on sparse datasets

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(a)

Figure 4 Continued

6 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

068

0675

067

0665

066

0655

065

MA

E

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(a)

089

0885

088

0875

087

0865

086

0855

RMSE

10 20 30 40 50 60 70 80 90 100Number of neighbors

UserKNNCFB-KNNHFB-KNN

(b)

Figure 2 e comparison of rating prediction accuracy with different user neighbor numbers

085

08

075

07

065

06

MA

E

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

(a)

10 20 30 40 50 60 70 80Average number of user ratings

UserKNNCFB-KNNHFB-KNN

105

1

095

09

085

08

RMSE

(b)

Figure 3 e comparison of rating prediction accuracy on sparse datasets

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(a)

Figure 4 Continued

6 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

5 Conclusion

In this paper we proposed an approach to calculate usersrsquopotential preferences based on hybrid features In particularby importing editorial features user-generated features andimage visual features of items for consideration trans-forming user-item ratings into hybrid feature ratings andcalculating usersrsquo potential interest values of each featurerecommendation is processed based on the feature interestvalues e experimental results showed that the proposedmethod had better recommendation performance on thesparse datasets and has higher efficiency on the largedatasets In future work we will try to extract more featuresthat can reflect usersrsquo potential preferences especially moreimage visual features based on pixel information rather thandominant hue

Conflicts of Interest

e authors declare that they have no conflicts of interest

Acknowledgments

is work was supported in part by the National NaturalScience Foundation of China (Nos 61672135 and61472064) the National Science Foundation of Chi-namdashGuangdong Joint Foundation (No U1401257) theSichuan Science-Technology Support Plan Program (Nos2018GZ0236 2016JZ0020 and 2017FZ0004) the Funda-mental Research Funds for the Central Universities (Nos2672018ZYGX2018J057 and ZYGX2015KYQD136) andthe Neijiang Science and Technology Incubating Project(No 170676)

References

[1] F Ricci L Rokach B Shapira and P B Kantor Recom-mender Systems Handbook Springer Berlin Germany 2015

[2] H Costa and L Macedo ldquoEmotion-based recommendersystem for overcoming the problem of information overloadrdquoin Proceedings of International Conference on Practical Ap-plications of Agents and Multi-Agent Systems pp 178ndash189Salamanca Spain May 2013

[3] P B orat R M Goudar and S Barve ldquoSurvey on col-laborative filtering content-based filtering and hybrid rec-ommendation systemrdquo International Journal of ComputerApplications vol 110 no 4 pp 31ndash36 2015

[4] P Lops M D Gemmis and G Semeraro Content-basedRecommender Systems State of the Art and Trends SpringerNew York NY USA 2014

[5] M J Pazzani and D Billsus Content-Based RecommendationSystems in 6e Adaptive Web Springer Berlin Germany2007

[6] L H Son ldquoDealing with the new user cold-start problem inrecommender systems A comparative reviewrdquo InformationSystems vol 58 no 5 pp 87ndash104 2016

[7] R Burke ldquoHybrid recommender systems survey and ex-perimentsrdquo User Modeling and User-Adapted Interactionvol 12 no 4 pp 331ndash370 2002

[8] Y Deldjoo M Quadrana M Elahi and P Cremonesi ldquoUsingmise-en-scene visual features based on mpeg-7 and deeplearning for movie recommendationrdquo arXiv preprint arXiv170406109 2018

[9] Y Deldjoo M Elahi M Quadrana and P CremonesildquoToward building a content-based video recommendationsystem based on low-level featuresrdquo in Proceedings of In-ternational Conference on Electronic Commerce and WebTechnology pp 45ndash56 Valencia Spain September 2015

[10] Y Deldjoo M Elahi P Cremonesi F Garzotto P Piazzollaand M Quadrana ldquoContent-based video recommendationsystem based on stylistic visual featuresrdquo Journal on DataSemantics vol 5 no 2 pp 99ndash113 2016

60

50

40

30

20

10

0

Reco

mm

enda

tion

time (

s)

25 50 75 100Proportion of the number of user ratings ()

UserKNNCFB-KNNHFB-KNN

(b)

Figure 4 e comparison of recommendation time (a) Recommendation time with different user rating number (b) Recommendationtime with different user number

Scientific Programming 7

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

[11] Y Deldjoo M Elahi P Cremonesi F B Moghaddam andA L E Caielli ldquoHow to combine visual features with tags toimprove movie recommendation accuracyrdquo in Proceedings ofInternational Conference on Electronic Commerce and WebTechnologies pp 34ndash45 Porto Portugal September 2016

[12] Y Deldjoo M Elahi and P Cremonesi ldquoUsing visual featuresand latent factors for movie recommendationrdquo in Proceedingsof ACM Recommender Systems Conference CEUR-WSBoston MA USA September 2016

[13] S Boutemedjet and D Ziou ldquoA graphical model for context-aware visual content recommendationrdquo IEEE Transactions onMultimedia vol 10 no 1 pp 52ndash62 2008

[14] E V D Melo E A Nogueira and D Guliato ldquoContent-basedfiltering enhanced by human visual attention applied toclothing recommendationrdquo in Proceedings of IEEE In-ternational Conference on TOOLS with Artificial Intelligencepp 644ndash651 San Jose CA USA November 2016

[15] R J R Filho J Wehrmann R C Barros et al ldquoLeveragingdeep visual features for content-based movie recommendersystemsrdquo in Proceedings of International Joint Conference onNeural Networks pp 604ndash611 Anchorage AK USA May2017

[16] I Cantador P Brusilovsky and T Kuflik ldquo2nd workshop oninformation heterogeneity and fusion in recommender sys-tems (HETREC 2011)rdquo in Proceedings of 5th ACM Conferenceon Recommender Systems (RecSys) Chicago IL USA October2011

[17] Grouplens 2018 httpwwwgrouplensorg[18] Imdb 2018 httpwwwimdbcom[19] Hetrec 2011 httpsgrouplensorgdatasetshetrec-2011[20] Q Wang X Yuan and M Sun ldquoCollaborative filtering

recommendation algorithm based on hybrid user modelrdquo inProceedings of International Conference on Fuzzy Systems andKnowledge Discovery vol 4 pp 1985ndash1990 Shandong ChinaAugust 2010

[21] M Elahi F Ricci and N Rubens ldquoA survey of active learningin collaborative filtering recommender systemsrdquo ComputerScience Review vol 20 pp 29ndash50 2016

[22] G Suganeshwari and S S Ibrahim A Survey on CollaborativeFiltering Based Recommendation System Springer In-ternational Publishing Basel Switzerland 2016

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

8 Scientific Programming

Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems

Hindawiwwwhindawicom

Volume 2018

International Journal of

ReconfigurableComputing

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018

Hindawiwwwhindawicom Volume 2018

Civil EngineeringAdvances in

Hindawiwwwhindawicom Volume 2018

Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications

Hindawiwwwhindawicom Volume 2018

Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

International Journal of

Biomedical Imaging

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Engineering Mathematics

International Journal of

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom Volume 2018

Computational Intelligence and Neuroscience

Hindawiwwwhindawicom Volume 2018

Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Hindawiwwwhindawicom Volume 2018

Human-ComputerInteraction

Advances in

Hindawiwwwhindawicom Volume 2018

Scientic Programming

Submit your manuscripts atwwwhindawicom