deep cnn-assisted personalized recommendation over big...

Research ArticleDeep CNN-Assisted Personalized Recommendation overBig Data for Mobile Wireless Networks

Yu Zheng 1 Xiaolong Xu 1 and Lianyong Qi 2

1School of Computer amp Software Jiangsu Engineering Center of Network MonitoringNanjing University of Information Science amp Technology Nanjing 210044 Jiangsu China2School of Information Science and Engineering Qufu Normal University Qufu 273165 Shandong China

Correspondence should be addressed to Lianyong Qi lianyongqigmailcom

Received 30 October 2018 Accepted 20 February 2019 Published 24 April 2019

Guest Editor Huaming Wu

Copyright copy 2019 Yu Zheng et alThis is an open access article distributed under the Creative CommonsAttribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

At present to improve the accuracy and performance for personalized recommendation in mobile wireless networks deeplearning has been widely concerned and employed with social and mobile trajectory big data However it is still challenging toimplement increasingly complex personalized recommendation applications over big data In view of this challenge a hybridrecommendation framework ie deepCNN-assisted personalized recommendation namedDCAPR is proposed formobile usersTechnically DCAPR integrates multisource heterogeneous data through convolutional neural network as well as inputs variousfeatures including image features text semantic features and mobile social user trajectories to construct a deep prediction modelSpecifically we acquire the location information and moving trajectory sequence in the mobile wireless network first Then thesimilarity of users is calculated according to the sequence of moving trajectories to pick the neighboring users Furthermorewe recommend the potential visiting locations for mobile users through the deep learning CNN network with the social andmobile trajectory big data Finally a real-word large-scale dataset collected from Gowalla is leveraged to verify the accuracy andeffectiveness of our proposed DCAPR model

1 Introduction

At present mobile wireless networks are moving towardsthe interconnection of all things intelligent interconnectionand social production organizations to accelerate customiza-tion decentralization and service transformation directionGlobal mobile wireless network users reach 3 billion 70million and smartphone penetration rate reaches 56 Theglobal population of 72 billion shows that the global mobilewireless network market as a whole has a demographic divi-dend and is considerable [1]The popularity ofmobile deviceswill take up a large number of usersrsquo time time fragmentswhich have been verified in China Europe the United Statesand other developed countries Whether on the subway oron the bus or even in the bathroom mobile devices playthe role of information acquisition tools always accompaniedby users Therefore mobile devices have become a majorgateway to recommending information on mobile wirelessnetworks

With the explosive growth of mobile traffic data andunprecedented demand for computing power usersrsquo behaviorin the mobile network is no longer limited to accessinginformation but more interacting with other users on thesocial network Mobile social networking as an open publicinformation exchange and business service platform hasquickly entered peoplersquos daily work and life [2] Famoussocial networks include Facebook LinkedIn Twitter Sinamicroblog Renren network Tencent QQ and WeChatIn social networking sites users are no longer individualindividuals but have intricate relationships withmany peopleon the network The most important resource in a socialnetwork is the relationship data between the user and the userUsing technologies such as GPS positioning the geographiclocation andmovement trajectory of themobile network usercan be obtained very accurately Due to the rapid develop-ment of GPS technology in the global positioning systemit is convenient to obtain the current location informationof the user The trajectory formed during the movement

HindawiWireless Communications and Mobile ComputingVolume 2019 Article ID 6082047 16 pageshttpsdoiorg10115520196082047

2 Wireless Communications and Mobile Computing

of people during activities can also be saved by collectingGPS data In a mobile social network the user is not justan individual and the behavior of the user in the socialnetwork is affected by these user relationships By mining theoriginal GPS data to find out the information between usersusers can not only find users who are similar to their ownactivities to establish user social networks but also predictother usersrsquo destinations through similar users thus givingusers some activity recommendations such as restauranttourist destination and gym recommendation Thereforethe research and application of the recommendation systemin such mobile social networking sites should consider theinteraction of user social relationships

The purpose of recommender system is to help con-sumers focus on products they care about and avoid overselection According to a large number of experimentaldata it is found that most people in the daily work anddecision-making always rely on other peoplersquos suggestionsRecommender systems are particularly important for thosewho lack sufficient personal experience and ability Whenthey cannot find the most needed information from a largeamount of information data personalized recommendationsystem will help them filter information According to theuserrsquos personal preferences and requirements different usersor user groups receive different recommendationsThereforepersonalization is a basic strategy to promote user experience

With the deepening of the era of big data the applicationof deep learning in recommendation systems has been paidmore and more attention by academics and industry In amobile wireless network the relationship between users isdifferent The users who establish the relationship may berelatives classmates colleagues friends in the real world orvirtual friends in the network such as members of social net-works with common interests This information constitutesa huge dataset So far the combination of deep learning andsocial network-based recommendation systems has triggereda series of research results and the recommendation oflocation-based social network sequence modeling based ondeep learning is in the ascendant Deep learning has shownoutstanding performance in many research fields such ascomputer vision natural language processing and so onwhich has aroused great interest At present how to securelyimplement multisource data efficient recommendation rec-ommendation service based on big data environment has alsoattracted the attention of many scholars [3ndash8] Obviously thefield of deep learning in recommendation system is booming

In this paper we deeply studied the personalized recom-mendation of social network based on the trajectory data ofmobile wireless network users This paper proposes a newlocation-based social network recommendation frameworkwhich combines the userrsquos location trajectory sequence user-shared images and text information together to recommendthe location of mobile social network users more accuratelyThe main contributions of this paper are as follows

(i) Firstly the location information of mobile networkusers in different time periods is analyzed then theuser trajectory in a certain time period is constructedthen the space-time similarity calculation method of

the location trajectory of the mobile network user isselected and finally the location-based neighbor ispicked out for the user

(ii) Study the problem of how to label user according tothe extract features of pictures and text We use thedual CNN network to extract the characteristics andsemantics of the images posted by the users to judgethe userrsquos interest points so as to find the neighborsin the mobile social network

(iii) Conduct experiments on a dataset collected fromGowalla to demonstrate the effectiveness of the pro-posed framework

The rest of this paper is organized as follows Section 2describes the progress of the deep learning algorithm and therecommended algorithm Section 3 specifies the preparationsthat the algorithm model needs to perform Section 4 detailsthe three main components of the proposed model InSection 5 we conducted extensive experiments and casestudies Finally in Section 6 we summarize and look forwardto the related work

2 Related Work

Recently a lot of research has been done in the field of deeplearning recommendation Recommender system estimatesuser preferences for projects and recommends items thatusers may like [9ndash11] Recommendation models are usuallyclassified into three types content based collaborative fil-tering and hybrid recommendation system [12] Contentbased recommendation is mainly based on the comparisonof project and user assistance information Collaborativefiltering provides recommendations by learning from userproject history interactions whether explicit (for examplethe userrsquos previous rating) or implicit feedback (for examplebrowsing history) Various auxiliary information (such astext image and video) can be considered A hybrid modelis a recommendation system that integrates two or morerecommendation strategies [12]

Zheng et al [13] based on deep collaborative neuralnetworks use comment information to jointly learn projectattributes and user behavior The model uses the shared layerto couple the project characteristics with user behavior Themodel was compared with five lines based on matrix decom-position probability matrix decomposition LDA coopera-tive topic regression hidden factor and cooperative in-depthlearning using three real-world datasets yelp review Amazonreview and Beer reviewThemodel outperforms all baselineson all benchmark datasets [14]

In terms of collaborative filtering based e-commercerecommendation system Li et al [15] first proposed theframework of combining in-depth learning features with CFmodels (such as matrix decomposition) in 2015 Kriegeskorteet al [16] developed a probabilistic rating autoencoder forunsupervised feature learning The autoencoder generatesuser profiles based on user item rating data effectivelyenhancing collaborative filtering methods Hidasi first usedRNN to recommend data based on short sessions rather

Wireless Communications and Mobile Computing 3

Figure 1 User check-in distribution map from the Gowalla dataset

than long historical data in 2016 [17] And in 2018 Hidasiet al used item features such as images and text to furtherenhance RNN based session recommendation [18] Jannachet al showed that the combination of RNN and KNNcan effectively improve the recommendation accuracy of e-commerce applications [19] Chatzis et al [20] used Bayesianstatistical variational reasoning model to improve recurrentneural network model based on session prediction Boginaet al [21] proposed a RNN model Merge dwell time (thetime users check for specific items) to improve the accuracyof session-based recommendation in e-commerce datasetsYoochoose Ebesu et al [22] showed that to solve the coldstart problem of cooperative recommendation system aneural network semantic personalized sorting method basedon deep neural network and pairwise learning is proposed

Based on hybrid recommendation system Kim et al[23] studied a model based on convolutional neural net-work which combines the metadata information of usersor projects to achieve the purpose of improving the matrixdecomposition method Wu et al [24] proposed a denoisingcollaborative filtering method based on automatic encoderThismodel serves as a general framework for all collaborativefiltering methods but with more flexible adjustments Themodel performs better on the MovieLens Yelp and Netflixdatasets than the baseline such as ItemPop ItemCF MatrixDecomposition BPR and FISM

Since convolution neural network has powerful functionsin image text audio video and other types of multisourcefeature representation learning most CNNS-based recom-mendation models use CNNs for feature extraction In [25]Wang et al studied the problem of using visual content toenhance POI recommendation In particular [26] proposeda new framework Visual Content Enhanced POI Recom-mendation (VPOI) which combines the visual content ofPOI recommendation and validates the effectiveness of the

proposed framework with real-world datasets In [27] Chu etal used pretrained deep network VGG-f from MatConvNettoolbox to extract CNN features and used support vectormachine (SVM) to classify images into four categories foodbeverage indoor and outdoor Different types of images mayvary in restaurant recommendation By combining contentbased approach and collaborative filtering method a hybridrestaurant recommendation system is constructed

3 Preliminaries

Nowadays many people use mobile social networks to postpraise share comment browse news and organize offlineactivities through social networks so that people with thesame hobbies can gather together If the userrsquos preferencesare learned from these behaviors and the user is accuratelyportrayed then personalized content recommendation canbe made according to personal preferences habits andother information For example if we open the news classapp because there is personalized content everyone seesthat the news home page is different In this chapter weanalyze the composition of recommendation data from threeperspectives Firstly usersrsquo potential hobby information isobtained by extracting the pictures of users in social net-works Secondly we judge the places that users often visitaccording to the usersrsquo moving track Finally we determinethe usersrsquo interest points by the posted picture and forwardinginformation in social networks

31 Trajectory Marking Scheme Figure 1 depicts the distri-bution of users around the world with data sources comingfrom the Gowalla datasetThis dataset contains 196591 nodes950327 edges and 6442890 check-ins

We classify social network data into three categoriesone-way social network data two-way social network data


3340

47

3022

55

159

41

176

0

20

40

60

80

100

120

140

160

180

200

467910

1214151620

(a) Attention degree

-7399542582

-1223861504-1201283097

-1059385872

-9874755001

minus130minus125minus120minus115minus110minus105minus100

minus95minus90minus85minus80minus75minus70minus65minus60

29 31 33 35 37 41 43

162015

(b) Motion trail

Figure 2 Social network attention We selected the 10 users as the representatives from the Gowalla edges dataset The ordinate representsthe ID number of the person concerned and the abscissa coordinates the ID number of fans As can be seen from the figure the user whoseID number is 20 received more than 176 peoplersquos attention the highest degree of attention and the user whose number is 10 only received theattention of 22 people Figure 2(b) shows the trajectory of the 3 users (their user Id are 15 16 and 20 respectively) The abscissa representslatitude and the ordinate represents longitude

and community-based social network data In social networkdata the amount of attention and the amount of fans of eachuser can be regarded as a complex directed graph Each noderepresents a user and the total number of users that eachuser pays attention to is recorded as the output degree ofthe node and the total number of fans is recorded as theinput degree of the node The social impact of users can be

judged according to the userrsquos output and the degree of inputUserrsquos degree reflects the social impact of users the greaterthe degree the greater the impact Userrsquos degree indicates thenumber of usersrsquo fans As you can see from Figure 2 the real-world phenomenon is that the most influential users in socialnetworks are always in the minority while the majority ofusers who pay attention to many people are in the minority


and the vast majority of users only pay attention to a fewpeople

It is easy to see in Figure 2(a) that users 16 and 20 havemore than 150 fans But their trajectory in Figure 2(b) showsthat there is no intersection in the place they are goingThis means that in social networks even if many people areconcerned about the same kind of people it does not meanthat there must be a common interest between these peopleTo mine the POI between them some information must beadded such as the userrsquos age education gender nationalityetc Experience shows that users from the same region tendto have the same tastes people with the same educationalexperience tend to focus on the same hot newsTherefore theuser data we set is as follows UserID Age Sex Native placeand Educational background

Place marking is an important condition for our DCAPRmodel We use a potential factor to represent the locationeffect at a given time and then learn from the potential factormodel The site marking scheme determines how to allocatepotential factors to specific locations

To capture site features on different time scales we repre-sent a site with a five-tuple representation and then aggregatetheir contributions Based on the empirical data analysis weconsider the characteristics of three site scales time longi-tude and latitudeThey are described by three different latentvectorsTherefore place Li is marked by five tuples (m119908 loi119897119892119894 and lID) which satisfies m (1 12) 119908 (1 7)119897119900119894 isin minus180 +180) and 119897119892119894 isin minus90 +90) and lIDis the place label In addition L1 h8timesW L2 h16timesW andL3 h24timesS are defined to represent the corresponding sitepotential factor matrix L1 h8timesW represents the trajectory ofuser activity within 8 hours of the working day L2 h16timesWrepresents the trajectory of user activity beyond 8 hours andL3 h24timesS represents the trajectory of user activity during theSunday and SaturdayW is the dimension of potential vectorrepresenting the working days in a week

After defining the location information of users we useCosine clustering algorithm to cluster the location informa-tionmatrix in order to obtain the friends with the same inter-est points in the community In this way runningCosine clus-tering algorithmcan get the user group and each user belongsto only one group In fact users in the same group generallyhave the same preferences and then they can recommend theinformation based on the past information of the users in thegroup Then we can recommend information to users moreaccurately according to the information of these friends

The Cosine clustering algorithm uses distance as similar-ity index to find 119870 classes in a given dataset and the centerof each class is obtained according to all the values in theclass Each class is described by clustering center For a givendataset119883 containing N d-dimensional data points and a class119870 to be partitioned the Euclidean distance is chosen as thesimilarity index The clustering objective is to minimize thesum of squares of all kinds of clustering as shown in formula(1)

119869 = 119896sum119896=1

119899sum119894=1

1003817100381710038171003817119909119894 minus 11990611989610038171003817100381710038172

(1)

In the past mobile trajectory model data sparsity is abig problem From Figure 2(b) we can see 6 users movingtrajectories within one day Observations show that each useris basically only active in a fixed number of places and someusers have repetitive movement paths indicating that theirbehavior is similar between 119871 119894 and 119871119895 in different places(L denotes location i and 119895 denote the number of differentplaces) However it is also easy to see that user with no 9 isbasically fixed in two places of activity not intersected withothers similarity is zero In addition we find that there areother changes User preferences vary with climate and mood

Check-in variations at different spatial scales can describeuser preferences from different perspectives (1) Users canlog on to their home system to communicate or shop withfriends or they can log on to APP in the office during the dayto communicate with colleagues or they can log in at nightwhen they have a good time at the bar (2)Users can visitmoreplaces in his her home or office on weekdays At weekendsheshe can checkmore information in some shopping centersor resorts (3) Users may have different habits in differentseasons For example he or she would ski in the cold northduring the hot summer or visit the south coast in the hotsummerTherefore it is impossible to capture all user featuresthat need to be represented in different scales by modelingonly the heterogeneity on a single scale

32 Comments Scheme Traditional machine learning meth-ods mainly use the n-gram concept in natural languageprocessing to extract text features and use TFIDF to adjust theweight of n-gram features and then input the extracted textfeatures into the classifier such as Logistic regression SVMfor training However the above feature extraction methodshave the problems of sparse data and dimension explosionwhich is disastrous for the classifier and makes the trainingmodel generalization ability limited Therefore it is oftennecessary to take some strategies to reduce dimension suchas stop word filtering low-frequency n-gram filtering LDAetc

WeuseCNN to classify sentences in our recommendationalgorithmA sentence ismade up ofmanywords If a sentencehas 119899 words and the ith word is 119908119894 and the word 119908119894 isexpressed as a vector of d-dimension after embedding thenthe matrix of a sentence 1199081n is n times d can be formalized asfollows

1198821119899 = 1199081 oplus 1199082 oplus sdot sdot sdot oplus 119908119899 (2)

A word window containing m words is represented as119882119894119894+119898minus1 and a convolution kernel is a matrix of sizem times d Afeature119891119894 can be extracted by extracting a word window froman activation function as follows

119891119894 = 119865 (119872 sdot 119882119894119894+119898minus1 + 119887) (3)

where 119887 is the corresponding intercept and 119865 is Sigmoidactivation function A convolution kernel matrix is used toscan the whole sentence from the beginning of the clauseto the end of the clause to extract the features of each wordwindow and a feature vector can be obtained which is


represented as follows (where the default is not to paddingthe sentence)

119883 = 1199091 1199092 119909119899minus119889+1 (4)

If there are119898 filters a vector of length119898 can be obtainedby a layer of convolution and a layer of pooling

119911 = [1198621 1198622 119862119898] (5)

where 119862119894 isin R it is the result of Max pooling afterextracting a feature map from a filter Next we carry out Maxpooling for feature map extracted from a convolution kernelFinally the vector 119883 is input to the full link layer to get thefinal feature extraction vector y

119910 = 119882 sdot 119911 + 119887 (6)

33 Image Feature Extract In social networks especiallyTwitter QQ WeChat and other online social apps usersoften share some pictures in the circle of friends Some ofthese pictures were taken by the users themselves and somewere taken by other users Some of these shared pictures havetext descriptions and some have no Regardless of wherethese images come from they represent the userrsquos interestpreferences at that moment If we can accurately analyzeand capture these points of interest from these images wecan provide relevant recommendation to users in a timelymanner

The Alexnet network structure model proposed by Alexin 2012 triggered a boom in neural network applications andwon the championship of the 2012 Image Recognition Com-petition making CNN the core algorithm model in imageclassification [28ndash30] So here we use the CNN network toextract the semantic features of the image

For CNN networks for processing user-image informa-tion the input data of Layer 1 is represented by R G andB of the original image For convolution operations thesize of convolution kernel is as follows 11lowast11lowast3 5lowast5lowast963lowast3lowast256 3lowast3lowast384 For example on the first layer if theoriginal image size is 227lowast227 then the image is convolutedby the convolution kernel of 11lowast11lowast3 Each convolution of theoriginal image generates a new pixel The convolution kernelmoves along the x-axis and y-axis directions of the originalimageThemoving step is 4 pixelsTherefore the convolutionkernel generates (227-11) 4 + 1 = 55 pixels (227 pixels minus11 exactly 54 pixels plus 11 subtracted to generate one pixel)and 55 lowast 55 pixels of rows and columns form the pixel layerafter convolution of the original image

As ReLU deep convolution network is much faster thanTanh and sigmoid based network training we have chosenthe ReLU function in our proposed model These pixel layersare processed by pool operation (pool operation) The scaleof pool operation is 3lowast3 and the step size of pool operationis 2 Then the image after pooling is normalized and thenormalized operation scale is 5lowast5 The Dropout operation ismore effective in preventing overfitting of neural networksRegular methods are used to prevent overfitting of modelsas generally as linear models while Dropout is implemented

in neural networks by modifying the structure of the neuralnetwork itself For a certain layer of neurons some neuronsare randomly deleted by the defined probability while keep-ing the individuals of the input layer and the output layerneurons unchanged and then the parameters are updatedaccording to the learning method of the neural network Inthe next iteration some neurons are rerandomly deleted untilthe end of the training The fully connected layer is actuallya convolution operation in which the convolution kernel sizeis the feature size of the upper layer output The result of theconvolution is a node which corresponds to a point of thefully connected layer The convolution takes local featuresand the full join is to reassemble the previous local featuresinto a complete graph through the weight matrix

4 Deep CNN-AssistedPersonalized Recommendation

41 DCAPR Framework In this paper we propose a noveldeep CNN-assisted personalized recommendation DCAPRAs shown in Figure 3 DCAPR consists of three layers ofprogressively progressive recommendation layers a roughrecommendation layer an enhanced recommendation layerand an accurate recommendation layer

The first layer is a rough recommendation layer Bycomparing the user trajectory sequence of the mobile socialnetwork the similarity of the userrsquos moving trajectorysequence is compared and several candidate buddy usersare picked out But among these candidate users theremay be ldquofake-friendsrdquo that is although the two people havesimilar movement trajectories the points of interest arecompletely different and cannot be regarded as true friendsFor example user A and user B have the same trajectorywithin a certain period of time and are all active in a certainmall However User A is concerned with clothing whileUser B is concerned with the e-sports game upstairs in theclothing store Therefore DCAPR built a second layer ofrecommendation framework to improve this problem

The second layer is the enhancement layer Based on thecandidate friends selected in the previous layer the CNNconvolutional neural network is used to extract features ofvarious image content uploaded by the candidate users onthemobile social platform According to the visual content ofthe image the interest association between the users can befurther explored so that the candidate friends can be refinedand filtered

The third layer is the accurate recommendation layerFor the text the deep learning CNN classification methodis combined with the context to extract and retrieve thesemantic content of the text and the vocabulary definedas illegal is deleted or the illegal vocabulary is occupied bythe recommendation Based on the previous two layers thesemantic comparison of the posts posted by the user is carriedout to construct a deep hierarchical prediction model formore accurate recommendation

Themodel integrates the location information of the userin the real world the pictures shared by the user in the socialnetwork and the text information published or forwarded by


Rough Layer

Extract allusers delete

uselessinformation

form theuser group U

Building atrajectorysequence

for mobilesocial

networkusers

Location Crossing Mode

Common Location Frequency Statistical Mode

User Moving Trajectory Matching Mode

Pick out kneighbor

users withthe sameinterestpoints

Enhanced Layer

Constructm datasets of

picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork

Pick out pneighbors who have

similar points of interest

Accurate Layer

CNN-rand

F1-Multiplefilter features

map

Max Pooling

Fullconnected layer

Preference Prediction

Figure 3 The framework of DCAPR model The framework consists of three layers a rough recommendation layer an enhancedrecommendation layer and an accurate recommendation layer

the user on a platformTherefore in the same space the useris recommended for images news and places by calculatingthe similarity among the semantic features of the charactersthe semantic features of the images and the auxiliary locationinformation

42 Rough Recommendation Layer In order to recommenda location point that may be of interest to a mobile socialnetwork user first of all look for his neighbors in the mobilesocial network Since his neighbors and the user may havesimilar points of interest we can recommend the place wherethe friend has been to the user and vice versa In this layer wetemporarily do not consider the context of the userrsquos locationsequence and only calculate and analyze the userrsquos behaviorcharacteristics from the perspective of time and space so as toroughly filter out several friends of the mobile social networkusers to prepare for future recommendation informationSince mobile social network users have different check-intimes and ways for location points we divide the roughrecommendation layer into two modes frequency positionpoint mode and trajectory sequence matching mode

421 Frequency Position PointMode Thedegree of interest ofthe user at the location point is determined based on the userrsquosfrequency of check-in at a certain pointWe first calculate thefrequency of each userrsquos access to a certain location compareit with the preset frequency threshold and then select theusers who visit the location with a frequency greater thana fixed threshold to form a user neighbor group Since thenature of each userrsquos work may be different the working timemay be different and the labor intensity may be different

such statistics may cause large errors For example user Aand user B frequently go to a famous gym but user A is acourier he is a customer who delivers courier items to thegym and user B is a member of the gym he is going toexercise every time Therefore it is easy to generate misjudgewhether two users are neighboring users only by the numberof occurrences at a certain place In order to avoid this defectwe have improved the statistical method by using the userrsquoscheck-in frequency ratio instead of the check-in frequencyThat is we count the ratio of the number of times each userhas a checkpoint li (1leilen) to the total number of check-insof the user in a fixed time range (for example 1 week) and thespecific calculation is as shown as formulas (7) and (8)

119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)

where 119899 represents the total number of location pointsand 119877119894119895 indicates the check-in frequency ratio of user 119894 atthe location point 119895 And 119901119894119895 is the percentage of user 119894 whochecked in at location j 119877119894 is the average percentage of eachuser who checked in at all locations

According to common sense of life we know that thegreater the proportion indicates that the user is more inter-ested in the location According to the probability of sign-inat each location point we can list each locationrsquos interest pointtable for each user in order of high to lowproportion and then


0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05

Figure 4 The check-in ratios for five users the probability of each userrsquos check-in at this location and the standard deviation of the numberof check-ins

calculate the variance according to the location interest pointtable to calculate the similarity between users

Table 1 lists the frequency of check-in frequencies for fiverandomly selected users at specific locations

Table 1 lists the check-in frequency ratios of five usersrandomly selected in the Gowalla dataset at the location ofthe tag 420315 As can be seen from Table 1 in terms of thenumber of times the user numbered UserId 66 has checkedin 47 times at the place and the number of check-ins is greaterthan the remaining four users However it is obviously wrongto judge that the user is very interested in the location point420315 because the user has a sign-in ratio of 171 at thatlocationThe other user UserId 7 has only 21 check-ins at thislocation This number of check-ins is the least compared tothe number of other four users However hisher check-inratio at position 420315 is 28 which clearly indicates thathisher interest in the location is very strong

Figure 4 shows the check-in ratios for five users theprobability of each userrsquos check-in at this location and thestandard deviation of the number of check-ins The bluecolor in the figure indicates the sign-in ratio of each userat the location point 420315 red indicates the proportion ofeach userrsquos ratio of the check-in at this location comparedto the total check-in ratio of the five users green indicatesthe calculated standard deviation The closer the standarddeviation to the sign-in ratio the more intense his or herinterest in the location

422 Trajectory Sequence Matching Mode According to thesequence of moving trajectories we can analyze from twodimensions in space and time and by comparing the motiontrajectories of the users we can find the nearest neighborssimilar to the trajectory sequence of the user And thenthe location contained in the nearest neighborrsquos trajectory

sequence is recommended to users who are similar to theirtrajectory but have not been to the location For mobilenetwork user location recommendation we divide it intothree steps The first step is the preprocessing stage Weobtain the movement trajectory and movement time intervalof each user by preprocessing the dataset thus forming theuserrsquosmovement trajectory sequence as shown as Figure 5 InStep 2 we regard the sequence of moving tracks as a stringeach character representing a place and setting a thresholdWhen comparing the motion trajectories between two usersonce there is a common substring whose length exceedsthe threshold in their trajectory it is considered that thetwo users find each other as the nearest neighbor If thecommon substringrsquos length is less than the threshold step3 is performed that is the similarity is simply consideredspatially We first count the number of times each user hasbeen to each location and then use the Cosine method tocalculate the similarity between users

Cosine Clustering for User Location How to accurately extractthe personalized information demand preference model ofmobile users with location changes according to the changerule of usersrsquo personalized demand for information changeswith location changes will become the key of location-basedmobile communication network information recommenda-tion service In the proposed model we learn the userrsquos per-sonalized demand for information according to the cyclicalchanges of the userrsquos position with time and extract the userrsquospersonalized information demand preference model Theuserrsquos geographical location is constantly changing within acertain period of time (one day oneweek or onemonth) andthe information services required in different geographicallocations are also different However within a plurality oftime periods (a few days) there is a certain regularity in thechange of the geographical location of the mobile user


Table 1 Check-in status of 5 users of mobile social network at location point 420315

UserId LId Check-in frequency Total Check-in Frequency Ratio Probability0 420315 28 224 0125 01171165357 420315 21 75 028 026234103831 420315 43 100 043 040288087952 420315 21 342 0061404 00575313966 420315 47 275 0170909 0160130159

Table 2 The locations and times of three users visited in the past week

UserID LocationsNatatorium Gym Hospital Bank Museum Restaurant Starbucks Library Bowling alley

UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1

In location-based social networks all POIs have loca-tion attributes and user behavior has temporal and spatialsequential patterns At present the social network can obtainthe userrsquos trajectory through technical means such as check-in and GPS in the social network According to the crossinformation of the userrsquos trajectory and combined with therating of the location the preference of the user can be foundHowever the recommendation system based on location-based social network should not only focus on the userrsquos owntrajectory sequence but also focus on the social relationshipbetween users so as to select the top k sites to recommend tousers through the ratings of other users with high similarityFor instance as shown in Figure 6 according to the userrsquostrajectory the user UA has visited Natatorium Gym Hos-pital Bank Museum etc in the past week Also in the pastweek user UB has visited Natatorium Restaurant HospitalMuseum Starbucks and Library respectively Another userUC went to Bowling alley Restaurant Museum Library andStarbucks

Table 2 shows the places where the three users in Figure 6have been visited and the number of times each place hasbeen visited From Table 2 we can see the social relationshipand similarity between UA UB and UC Therefore we canrecommend to users UA UB and UC the sites that they maybe interested in according to the similarity

We divide each time period into 119873 segments based onthe number of user activitiesThen the sequence of change ofthe geographical location of the mobile user in a time periodis 119897119894 i=12 N and in all119872 time periods the sequence ofposition change sequence of each mobile user is

119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)

The location-based mobile user preference model is atwo-tuple 119880119896=(119906119894 119871119895) where 119880119896 represents the kth user ina mobile social network And the two-tuple 119880119896=(119906119894 119871119895)represents the ith user at a certain location 119871119895 Suppose thereare two mobile social network users A and BThe applicationcharacteristics of all network service items in the locations119871119886 and 119871119887 are 119880119886=(119906119886 119871119886) and 119880119887=(119906119887 119871119887) respectively119906119886 and 119906119887 which are all network service multidimensional

feature vectors used by the two mobile social network usersat locations 119871119909 and 119871119910 are normalized such that they have thesame length The location-based user preference similaritycan be defined as follows

119904119894119898 (119880119886 119880119887) = 1119890119889119894119904(119871119886 119871119887)times sum119899119894=1 (119906119886119894 times 119906119887119894)radicsum119899119894=1 (119906119886119894)2 times radicsum119899119894=1 (119906119887119894)2

(10)

Obviously on the one hand when two mobile users arein the same position the distance between them is 0 dis(119871119886119871119887)=0 at this time 119890119889119894119904(119871119886 119871119887) = 1 For any two differentlocations of mobile users due to dis(119871119886 119871119887)gt0 then 0lt119890119889119894119904(119871119886 119871119887) lt1 If and only if a=b sim(119906119886 119906119887)=l Thereforefor any two mobile users the similarity 119904119894119898(119906119894 119906119895) isin [0 1]According to Table 2 we can calculate the similarity between119880119860 119880119861 and 119880119862 the result is shown in Table 3

On the basis of the similarity calculation results in Table 3we can judge the userrsquos preference from the trajectory of theplace where the user has been and calculate the similaritybetween the trajectory of the user and the user As can be seenfrom Table 3 the similarity between User B and User C issignificantly higher than that between user A and user C andbetween user A and user B In this way we can recommendthe places where User B has been to User C according to theinterests of User C

43 Enhanced Recommendation Layer CNN network forimage processing adopts seven-layer structure andCNNnet-work for text processing adopts three-layer frame structureFirstly we rescale images to 227lowast227 And thenwe use 8-layerVGGNet to extract an image feature map

As shown as Figure 7 semantic information is extractedfrom pictures which are posted by different users and theuser is tagged with various categories For example from thepicture that user 1 has posted we can deduce that the usermay not only like to travel but also may be a photographyenthusiast Therefore the user 1 can be affixed with a travel-loving label or a photographerrsquos label similarly the user 3


Table 3 Cosine formula is used to calculate the similarity of 3 users

UserID UserIDUA UB UC

UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3

Figure 5 An example about three usersrsquo trajectory

Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C

Figure 6 Three users UA UB and UC outdoor trajectory of the past week


Label 1 TourismLabel 2 Photography

Label 1 YogaLabel 2 TourismLabel 3 Photography


CNN

CNN

CNN

User 1

User 2

User 3

Figure 7 An example of the extracted semantic information from posted pictures by different users and label these users

is the same The user 2 can not only be tagged with traveland photography but also can derive the userrsquos preferredsport according to the content in the figure If the motiontag continues to be subdivided information about the userrsquospreference to practice yoga can be obtained Therefore if theuser has just arrived in the city there is no local trajectorygenerated that is when the recommendation based on thelocation information is a cold start we can recommend thelocation that the user may be interested in according to thepicture that the user has posted

44 CNN Network for Comments The third layer of ourmodel is the extraction of text features from comments orforwarded articles from users in social networking forumsThe text extraction method refers to the extraction of textfeatures using the CNN convolution network First theoriginal text is preprocessed including word segmentationdeactivation etc and then the preprocessed text is vector-ized using the skip-gram model in word2vec Finally eachsentence is transformed into a matrix form Next the featureextraction and classification of the comment statements canbe performed using the CNN network This process is verysimilar to the image feature extraction using CNN Whenconvolving the text matrix the text matrix is convolved usingfilters of different lengths The width of the filter is equal tothe length of the longest word vector in the sentence andthen the vector extracted by each filter is operated using Maxpooling Finally each filter corresponds to a number and theresults of these filters are spliced together to obtain a vectorcharacterizing the sentence

5 Experiments

51 Dataset and Experimental Settings Using technologiessuch as user check-in information and GPS positioning the

Table 4 Statistics of dataset We separated images from geographicinformation from 196591 users

Nodes 196591Edges 950327Nodes in largest WCC 196591 (1000)Edges in largest WCC 950327 (1000)Nodes in largest SCC 196591 (1000)Edges in largest SCC 950327 (1000)Average clustering coefficient 02367Number of triangles 2273138Fraction of closed triangles 0007952Diameter (longest shortest path) 1490-percentile effective diameter 57Check-ins 6442890

geographic location and movement trajectory of the mobilenetwork user can be obtained very accurately

We consider using a publicly available Gowalla datasetfor our proposed model Gowalla dataset is a location-basedsocial networking website where users share their locationsby checking-inThe friendship network is undirected andwascollected using their public API and consists of 196591 nodesand 950327 edges We have collected a total of 6442890check-ins over the period of Feb 2009-Oct 2010

Table 4 presents the statistics of the datasetrsquos detail Thedataset provides information such as user identificationage sex occupation time location image comments etcFollowing [31] we removed all users who have less than 10check-ins and locations which have fewer than 15 check-ins Finally the collection constructed contained 837352


Table 5 Comparison of the evaluation results of four models on test sets

Method Precision Recall F1PACE 00976 00913 0094344944VPOI 01283 01208 0124437094SEER 01462 01483 0147242513DCAPR 01528 01567 0154725428

Table 6 Results of recommendation index in the case of recommendation number increase and recommendation dimension fixed

Method Number of Recommended Precision Recall F1PACE 10 00924 00815 00866PACE 20 00976 00913 00943VPOI 10 01064 01059 01061VPOI 20 01283 01208 01244SEER 10 01305 01297 01301SEER 20 01362 01383 01372DCAPR 10 01398 01387 01392DCAPR 20 01528 01567 01547

subtrajectories with corresponding locations comments andimages Table 3 presents the statistics of the datasetrsquos detail

52 Baselines For comparison with the proposed model weconsider the following baselines

(i) Preference and Context Embedding (PACE) Reference[31] pointed out the current POI recommendationmethods are designed for specific data and problemsand a general semisupervised learning model is pro-posedThat is the preference and context embeddingmodel can utilize the information of neighboringusers and locations to alleviate the data sparse prob-lem of the recommendation system

(ii) Visual Content Enhanced POI Recommendation(VPOI) Reference [25] proposed a POI recommen-dation model with visual content enhancement basedon CNN and probability matrix factorization Theauthor studied how to incorporate image contentinformation to improve the POI recommendationVPOI uses CNN to extract features from imagecontent and constructs a probabilistic thememodel through user-image relationship POI-imagerelationship and user-POI relationship Finally theimage feature extraction and probability topic modelare integrated into one unified The optimizationfunction is built in the framework and the NegativeSampling method is used to optimize the parameters

(iii) Sequential Embedding Rank (SEER) Reference [32]made a point of interest recommendation based onthe userrsquos interest preferences and mobile modeSpecifically SEER model uses distributed representa-tion technology to learn the embedded representationof the user and then embed the user as a constraintinto the paired sorting model to capture the sequencepattern of the userrsquos behavior At the same time it alsoincorporates time and space information

53 Experimental Results and Analysis The proposedmethod is evaluated based on Precision Recall andAccuracy using a real-world dataset We adopt the evaluationindex in information retrieval to evaluate our method andcontrast model method Specifically we used Precisionand Recall two values to evaluate the two formulas Thedefinitions are as follows

Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)

1198651 = 2 lowast Pr119890119888119894119904119894119900119899 lowast Re119888119886119897119897Pr119890119888119894119904119894119900119899 + Re119888119886119897119897 (13)

where 119871visited represents the set of locations containedin the Gowalla dataset and 119871NR represents the set of placeswith the recommended number of M The final values forPrecision and Recall are averaged over the dataset for allusers The related experimental results are shown in Table 5

Figure 8 shows the Precision Recall and F1-Score ofdifferent models From Table 5 and Figure 8 we can seethat our model DCAPR is significantly better than theother three benchmark comparison algorithms because weincorporate multisource heterogeneous information suchas images text geographic location information etc Theintegration of multisource heterogeneous information helpsto more accurately characterize the userrsquos access behaviorwhich in turn enables more accurate modeling

In Table 6 when the dimensions remain the same andwhen the number of recommendations increases from 10 to20 the results of eachmodel on the corresponding evaluationindicators (Precision and Recall) are also improved This isdefined by the calculation formulas of Precision and RecallWhen more places are recommended to the user it is easierto hit the already visited records of the user in the test datasetthus causing the value to be large


Table 7 Results of recommendation index in the case of dimension increase and recommendation dimension number fixed

Method Dimension Precision Recall F1PACE 100 00924 00815 00866PACE 500 00965 00902 00932VPOI 100 01064 01059 01061VPOI 500 01279 01264 01271SEER 100 01305 01297 01301SEER 500 01358 01376 01367DCAPR 100 01398 01387 01392DCAPR 500 01525 01563 01544

0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)

Figure 8 Precision Recall and F1-Score with different number of recommendations


Table 7 shows that when the number of recommendationsis consistent and the dimension is increased from 100 to 500the values of the respective models on the correspondingevaluation indicators are correspondingly increased This isbecause more dimensions can describe the hidden featurevalues more carefully which will make the model effectincreaseHowever as can be seen fromTable 7 the increase inthe dimension does notmake themodel continue to improvebecause the oversized dimension leads to overfitting

6 Conclusion and Future Work

The development of intelligent mobile devices has driven therapid development of mobile social networks Deep learning-driven algorithms and models can promote wireless networkanalysis and resource management and help to cope withthe growth of communication and computing in emergingmobile applications In this paper by means of in-depthlearning the user behavior sequence pattern is integrated intothe recommendation system which is helpful to discover thedependencies between user behaviors and improve the qual-ity of recommendation It is for this purpose we presented anovel social network recommendation algorithm frameworkbased on mobile wireless network Finally a comprehensiveexperiment of the DCAPR method is carried out using theuser dataset from Gowalla The results show that the baselineimprovement is more significant when the userrsquos behaviorsequence is fused with the userrsquos posted images text and soon through DCAPR framework

Now the recommendation systembased on deep learningfaces two main problems one is how to better combinemultisource data for recommendation the other is how toanalyze the intermediate process and the final result froma mathematical perspective The deep learning-based rec-ommendation system usually uses the end-to-end model topredict the userrsquos preference for the project by using the mul-tisource heterogeneous data as input The recommendationsystem involves many auxiliary data comments tags userportrait information user socialization and recommendedsituation information (time location) It can be seen thatthe current recommendation system needs many modelingfactors In the future if the multiobjective optimization [33ndash37] and multisource heterogeneous data can be combinedto dynamically evolve user preferences and project featuresthe performance of the recommendation system can beimproved For the second question we are inspired by theresearch of Sun et al [38ndash48] and we may be able to find outthe answer we want

At present learning algorithms in mobile wireless sys-tems are immature and inefficient More endeavors areneeded to bridge the gap between deep learning and wirelesscommunications and mobile computing research Specifi-cally for mobile wireless network recommendation systemthe application of in-depth learning in location-based socialnetwork recommendation systemmainly focuses on sequen-tial pattern modeling How to integrate a large number ofimplicit and explicit heterogeneous spatiotemporal data ofmobile wireless network users through in-depth learning

so as to build a unified recommendation framework is thefuture direction of development

Data Availability

The data used to support the findings of this study areincluded within the article

Conflicts of Interest

The authors declare that they have no conflicts of interest

Acknowledgments

The work was funded by the National Natural ScienceFoundation of China (Grants nos 61702277 and 61872219)

References

[1] X Zhang and Q Zhu ldquoHierarchical caching for statistical QoSguaranteed multimedia transmissions over 5G edge computingmobilewireless networksrdquo IEEEWireless CommunicationsMag-azine vol 25 no 3 pp 12ndash20 2018

[2] Z Sheng C Mahapatra V C M Leung M Chen and P KSahu ldquoEnergy efficient cooperative computing in mobile wire-less sensor networksrdquo IEEE Transactions on Cloud Computingvol 6 no 1 pp 114ndash126 2018

[3] L Qi R Wang S Li et al ldquoTime-aware distributed service rec-ommendationwith privacy-preservationrdquo Information Sciencesvol 480 pp 354ndash364 2018

[4] Y Xu L QiW Dou and J Yu ldquoPrivacy-preserving and scalableservice recommendation based on simhash in a distributedcloud environmentrdquo Complexity vol 2017 Article ID 34378549 pages 2017

[5] X Xu Q Liu Y Luo et al ldquoA computation offloading methodover big data for IoT-enabled cloud-edge computingrdquo FutureGeneration Computer Systems vol 95 pp 522ndash533 2019

[6] W Gong L Qi and Y Xu ldquoPrivacy-aware multidimensionalmobile service quality prediction and recommendation indistributed fog environmentrdquo Wireless Communications andMobile Computing vol 2018 Article ID 3075849 8 pages 2018

[7] X Xu S Fu L Qi et al ldquoAn IoT-Oriented data placementmethod with privacy preservation in cloud environmentrdquoJournal of Network and Computer Applications vol 124 pp 148ndash157 2018

[8] S Zhang L Yao and A Sun ldquoDeep learning based rec-ommender system A survey and new perspectivesrdquo ACMComputing Surveys vol 1 no 1 pp 1ndash35 2018

[9] M Gruteser and D Grunwald ldquoAnonymous usage of location-based services through spatial and temporal cloakingrdquo in Pro-ceedings of the 1st International Conference on Mobile SystemsApplications and Services MobiSys 2003 pp 31ndash42 May 2003

[10] G Adomavicius and A Tuzhilin ldquoToward the next generationof recommender systems a survey of the state-of-the-art andpossible extensionsrdquo IEEE Transactions on Knowledge and DataEngineering vol 17 no 6 pp 734ndash749 2005

[11] F Ricci L Rokach and B Shapira ldquoRecommender systemsintroduction and challengesrdquo in Recommender Systems Hand-book pp 1ndash34 Springer US 2015


[12] R Burke ldquoHybrid recommender systems survey and experi-mentsrdquoUser Modeling and User-Adapted Interaction vol 12 no4 pp 331ndash370 2002

[13] Y Zheng L Zhang ZMa X Xie andW-YMa ldquoRecommend-ing friends and locations based on individual location historyrdquoACM Transactions on the Web (TWEB) vol 5 no 1 article no5 2011

[14] C Chen P Zhao L Li J Zhou X Li and M Qiu ldquoLocallyconnected deep learning framework for industrial-scale rec-ommender systemsrdquo in Proceedings of the 26th InternationalConference on World Wide Web Companion InternationalWorld Wide Web Conferences Steering Committee pp 769-770Perth Australia 2017

[15] S Li J Kawale and Y Fu ldquoDeep collaborative filtering viamarginalized denoising auto-encoderrdquo in Proceedings of the24th ACM International Conference on Information and Knowl-edge Management CIKM 2015 pp 811ndash820 ACM AustraliaOctober 2015

[16] N Kriegeskorte ldquoDeep neural networks a new framework formodeling biological vision and brain information processingrdquoAnnual Review of Vision Science vol 1 no 1 pp 417ndash446 2015

[17] B Hidasi M Quadrana A Karatzoglou and D Tikk ldquoParallelrecurrent neural network architectures for feature-rich session-based recommendationsrdquo in Proceedings of the 10th ACMConference on Recommender Systems RecSys 2016 ACM pp241ndash248 USA September 2016

[18] B Hidasi and A Karatzoglou ldquoRecurrent neural networks withtop-k gains for session-based recommendationsrdquo inProceedingsof the 27th ACM International Conference on Information andKnowledge Management ACM pp 843ndash852 Torino ItalyOctober 2018

[19] D Jannach L Lerche F Gedikli and G Bonnin ldquoWhatrecommenders recommendan analysis of accuracy popularityand sales diversity effectsrdquo in Proceedings of the InternationalConference on User Modeling Adaptation and Personalizationpp 25ndash37 Springer Berlin Heidelberg 2013

[20] S P Chatzis P Christodoulou and A S Andreou ldquoRecurrentlatent variable networks for session-based recommendationrdquoin Proceedings of the 2nd Workshop on Deep Learning forRecommender Systems (DLRS 2017) ACM pp 38ndash45 ComoItaly August 2017

[21] V Bogina and T Kuflik ldquoIncorporating dwell time in session-based recommendations with recurrent neural networksrdquo inProceedings of the 1st Workshop on Temporal Reasoning inRecommender Systems in CEUR Workshop pp 57ndash59 ComoItaly August 2017

[22] T Ebesu and Y Fang ldquoNeural semantic personalized rankingfor item cold-start recommendationrdquo Information RetrievalJournal vol 20 no 2 pp 109ndash131 2017

[23] Y Kim ldquoConvolutional neural networks for sentence classifica-tionrdquo httpsarxivorgabs14085882 2014

[24] Y Wu C DuBois A X Zheng and M Ester ldquoCollaborativedenoising auto-encoders for top-N recommender systemsrdquo inProceedings of the 9th ACM International Conference on WebSearch and Data Mining WSDM 2016 pp 153ndash162 ACM SanFrancisco USA February 2016

[25] S Wang Y Wang J Tang K Shu S Ranganath and HLiu ldquoWhat your images reveal exploiting visual contents forpoint-of-interest recommendationrdquo in Proceedings of the 26thInternationalWorldWideWebConferenceWWW2017 pp 391ndash400 Australia April 2017

[26] J Davidson B Liebald J Liu P Nandy and T Van Vleet ldquoTheYouTube video recommendation systemrdquo in Proceedings of the4th ACM Recommender Systems Conference (RecSys rsquo10) pp293ndash296 Barcelona Spain September 2010

[27] W-T Chu and Y-L Tsai ldquoA hybrid recommendation systemconsidering visual information for predicting favorite restau-rantsrdquoWorld Wide Web vol 20 no 6 pp 1313ndash1331 2017

[28] Y Zheng J Zhu W Fang and L Chi ldquoDeep learning hashfor wireless multimedia image content securityrdquo Security andCommunication Networks vol 2018 Article ID 8172725 13pages 2018

[29] Z Ligang and C Song ldquoFast near-duplicate image detection inriemannianspace by a novel hashing schemerdquoCMCComputersMaterials amp Continua vol 56 no 3 pp 529ndash539 2018

[30] W Fang Z Feihong S Victor and D Yewen ldquoA methodfor improving CNN-based image recognition using DCGANrdquoCMC Computers Materials amp Continua vol 57 no 1 pp 167ndash178 2018

[31] C Yang L Bai C Zhang Q Yuan and J Han ldquoBridgingcollaborative filtering and semi-supervised learning a neuralapproach for POI recommendationrdquo in Proceedings of the23rd ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining KDD 2017 ACM pp 1245ndash1254Halifax Canada August 2017

[32] S Zhao T Zhao I King and M R Lyu ldquoGeo-teaser geo-temporal sequential embedding rank for point-of- interestrecommendationrdquo in Proceedings of the 26th International Con-ference on World Wide Web Companion International WorldWide Web Conferences Steering Committee pp 153ndash162 PerthAustralia April 2017

[33] Y Yuan and W Banzhaf ldquoARJA automated repair of javaprograms via multi-objective genetic programmingrdquo IEEETransactions on Software Engineering 2018

[34] Y Yuan Y-S Ong A Gupta and H Xu ldquoObjective reductionin many-objective optimization evolutionary multiobjectiveapproaches and comprehensive analysisrdquo IEEE Transactions onEvolutionary Computation vol 22 no 2 pp 189ndash210 2018

[35] Y Yuan and H Xu ldquoMultiobjective flexible job shop schedulingusing memetic algorithmsrdquo IEEE Transactions on AutomationScience and Engineering vol 12 no 1 pp 336ndash353 2015

[36] Y Yuan H Xu BWang B Zhang and X Yao ldquoBalancing con-vergence and diversity in decomposition-based many-objectiveoptimizersrdquo IEEE Transactions on Evolutionary Computationvol 20 no 2 pp 180ndash198 2016

[37] Y Yuan H Xu B Wang and X Yao ldquoA new dominancerelation-based evolutionary algorithm for many-objective opti-mizationrdquo IEEE Transactions on Evolutionary Computation vol20 no 1 pp 16ndash37 2016

[38] W W Sun ldquoStabilization analysis of time-delay Hamiltoniansystems in the presence of saturationrdquoAppliedMathematics andComputation vol 217 no 23 pp 9625ndash9634 2011

[39] M Han X Hou L Sheng and C Wang ldquoTheory of rotatedequations and applications to a populationmodelrdquoDiscrete andContinuousDynamical Systems - Series A vol 38 no 4 pp 2171ndash2185 2018

[40] W Sun and L Peng ldquoObserver-based robust adaptive controlfor uncertain stochastic Hamiltonian systems with state andinput delaysrdquoNonlinear Analysis Modelling and Control vol 19no 4 pp 626ndash645 2014


[41] F Li and G Du ldquoGeneral energy decay for a degenerateviscoelastic Petrovsky-type plate equation with boundary feed-backrdquo Journal of Applied Analysis and Computation vol 8 no1 pp 390ndash401 2018

[42] J Liu and A Qian ldquoGround state solution for a Schrodinger-Poisson equation with critical growthrdquoNonlinear Analysis RealWorld Applications vol 40 pp 428ndash443 2018

[43] J Jiang L Liu and Y Wu ldquoPositive solutions to nonlinearfractional differential equations involving Stieltjes integralsconditionsrdquo Journal of Nonlinear Sciences and ApplicationsJNSA vol 10 no 10 pp 5360ndash5372 2017

[44] H Liu and H Gao ldquoGlobal well-posedness and long timedecay of the 3D Boussinesq equationsrdquo Journal of DifferentialEquations vol 263 no 12 pp 8649ndash8665 2017

[45] K M Zhang ldquoOn a sign-changing solution for some fractionaldifferential equationsrdquo Boundary Value Problems vol 2017 no59 8 pages 2017

[46] Y Guo ldquoGlobally robust stability analysis for stochastic cohen-grossberg neural networks with impulse and time-varyingdelaysrdquoUkrainianMathematical Journal vol 69 no 8 pp 1220ndash1233 2017

[47] H Tian andM Han ldquoBifurcation of periodic orbits by perturb-ing high-dimensional piecewise smooth integrable systemsrdquoJournal of Differential Equations vol 263 no 11 pp 7448ndash74742017

[48] Y A Amer A M S Mahdy and E S M Youssef ldquoSolv-ing fractional integro-differential equations by using sumudutransform method and hermite spectral collocation methodrdquoComputers Materials and Continua vol 54 no 2 pp 161ndash1802018

International Journal of

AerospaceEngineeringHindawiwwwhindawicom Volume 2018

RoboticsJournal of

Hindawiwwwhindawicom Volume 2018


Active and Passive Electronic Components

VLSI Design



Shock and Vibration


Civil EngineeringAdvances in

Acoustics and VibrationAdvances in



Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwwwhindawicom

Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of



Journal ofEngineeringVolume 2018

SensorsJournal of



RotatingMachinery


Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018


Chemical EngineeringInternational Journal of Antennas and

Propagation




Navigation and Observation


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia

Submit your manuscripts atwwwhindawicom


of people during activities can also be saved by collectingGPS data In a mobile social network the user is not justan individual and the behavior of the user in the socialnetwork is affected by these user relationships By mining theoriginal GPS data to find out the information between usersusers can not only find users who are similar to their ownactivities to establish user social networks but also predictother usersrsquo destinations through similar users thus givingusers some activity recommendations such as restauranttourist destination and gym recommendation Thereforethe research and application of the recommendation systemin such mobile social networking sites should consider theinteraction of user social relationships

The purpose of recommender system is to help con-sumers focus on products they care about and avoid overselection According to a large number of experimentaldata it is found that most people in the daily work anddecision-making always rely on other peoplersquos suggestionsRecommender systems are particularly important for thosewho lack sufficient personal experience and ability Whenthey cannot find the most needed information from a largeamount of information data personalized recommendationsystem will help them filter information According to theuserrsquos personal preferences and requirements different usersor user groups receive different recommendationsThereforepersonalization is a basic strategy to promote user experience

With the deepening of the era of big data the applicationof deep learning in recommendation systems has been paidmore and more attention by academics and industry In amobile wireless network the relationship between users isdifferent The users who establish the relationship may berelatives classmates colleagues friends in the real world orvirtual friends in the network such as members of social net-works with common interests This information constitutesa huge dataset So far the combination of deep learning andsocial network-based recommendation systems has triggereda series of research results and the recommendation oflocation-based social network sequence modeling based ondeep learning is in the ascendant Deep learning has shownoutstanding performance in many research fields such ascomputer vision natural language processing and so onwhich has aroused great interest At present how to securelyimplement multisource data efficient recommendation rec-ommendation service based on big data environment has alsoattracted the attention of many scholars [3ndash8] Obviously thefield of deep learning in recommendation system is booming

In this paper we deeply studied the personalized recom-mendation of social network based on the trajectory data ofmobile wireless network users This paper proposes a newlocation-based social network recommendation frameworkwhich combines the userrsquos location trajectory sequence user-shared images and text information together to recommendthe location of mobile social network users more accuratelyThe main contributions of this paper are as follows

(i) Firstly the location information of mobile networkusers in different time periods is analyzed then theuser trajectory in a certain time period is constructedthen the space-time similarity calculation method of

the location trajectory of the mobile network user isselected and finally the location-based neighbor ispicked out for the user

(ii) Study the problem of how to label user according tothe extract features of pictures and text We use thedual CNN network to extract the characteristics andsemantics of the images posted by the users to judgethe userrsquos interest points so as to find the neighborsin the mobile social network

(iii) Conduct experiments on a dataset collected fromGowalla to demonstrate the effectiveness of the pro-posed framework

The rest of this paper is organized as follows Section 2describes the progress of the deep learning algorithm and therecommended algorithm Section 3 specifies the preparationsthat the algorithm model needs to perform Section 4 detailsthe three main components of the proposed model InSection 5 we conducted extensive experiments and casestudies Finally in Section 6 we summarize and look forwardto the related work

2 Related Work

Recently a lot of research has been done in the field of deeplearning recommendation Recommender system estimatesuser preferences for projects and recommends items thatusers may like [9ndash11] Recommendation models are usuallyclassified into three types content based collaborative fil-tering and hybrid recommendation system [12] Contentbased recommendation is mainly based on the comparisonof project and user assistance information Collaborativefiltering provides recommendations by learning from userproject history interactions whether explicit (for examplethe userrsquos previous rating) or implicit feedback (for examplebrowsing history) Various auxiliary information (such astext image and video) can be considered A hybrid modelis a recommendation system that integrates two or morerecommendation strategies [12]

Zheng et al [13] based on deep collaborative neuralnetworks use comment information to jointly learn projectattributes and user behavior The model uses the shared layerto couple the project characteristics with user behavior Themodel was compared with five lines based on matrix decom-position probability matrix decomposition LDA coopera-tive topic regression hidden factor and cooperative in-depthlearning using three real-world datasets yelp review Amazonreview and Beer reviewThemodel outperforms all baselineson all benchmark datasets [14]

In terms of collaborative filtering based e-commercerecommendation system Li et al [15] first proposed theframework of combining in-depth learning features with CFmodels (such as matrix decomposition) in 2015 Kriegeskorteet al [16] developed a probabilistic rating autoencoder forunsupervised feature learning The autoencoder generatesuser profiles based on user item rating data effectivelyenhancing collaborative filtering methods Hidasi first usedRNN to recommend data based on short sessions rather







3 Preliminaries





3340

47

3022

55

159

41

176

0

20

40

60

80

100

120

140

160

180

200

467910

1214151620


-7399542582

-1223861504-1201283097

-1059385872

-9874755001



29 31 33 35 37 41 43

162015

(b) Motion trail











119869 = 119896sum119896=1

119899sum119894=1

1003817100381710038171003817119909119894 minus 11990611989610038171003817100381710038172

(1)







119891119894 = 119865 (119872 sdot 119882119894119894+119898minus1 + 119887) (3)




119883 = 1199091 1199092 119909119899minus119889+1 (4)


119911 = [1198621 1198622 119862119898] (5)


119910 = 119882 sdot 119911 + 119887 (6)













Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia








3 Preliminaries





3340

47

3022

55

159

41

176

0

20

40

60

80

100

120

140

160

180

200

467910

1214151620


-7399542582

-1223861504-1201283097

-1059385872

-9874755001



29 31 33 35 37 41 43

162015

(b) Motion trail











119869 = 119896sum119896=1

119899sum119894=1

1003817100381710038171003817119909119894 minus 11990611989610038171003817100381710038172

(1)







119891119894 = 119865 (119872 sdot 119882119894119894+119898minus1 + 119887) (3)




119883 = 1199091 1199092 119909119899minus119889+1 (4)


119911 = [1198621 1198622 119862119898] (5)


119910 = 119882 sdot 119911 + 119887 (6)













Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



3340

47

3022

55

159

41

176

0

20

40

60

80

100

120

140

160

180

200

467910

1214151620


-7399542582

-1223861504-1201283097

-1059385872

-9874755001



29 31 33 35 37 41 43

162015

(b) Motion trail











119869 = 119896sum119896=1

119899sum119894=1

1003817100381710038171003817119909119894 minus 11990611989610038171003817100381710038172

(1)







119891119894 = 119865 (119872 sdot 119882119894119894+119898minus1 + 119887) (3)




119883 = 1199091 1199092 119909119899minus119889+1 (4)


119911 = [1198621 1198622 119862119898] (5)


119910 = 119882 sdot 119911 + 119887 (6)













Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









119869 = 119896sum119896=1

119899sum119894=1

1003817100381710038171003817119909119894 minus 11990611989610038171003817100381710038172

(1)







119891119894 = 119865 (119872 sdot 119882119894119894+119898minus1 + 119887) (3)




119883 = 1199091 1199092 119909119899minus119889+1 (4)


119911 = [1198621 1198622 119862119898] (5)


119910 = 119882 sdot 119911 + 119887 (6)













Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia




119883 = 1199091 1199092 119909119899minus119889+1 (4)


119911 = [1198621 1198622 119862119898] (5)


119910 = 119882 sdot 119911 + 119887 (6)













Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



Rough Layer


uselessinformation



for mobilesocial

networkusers




Pick out kneighbor


Enhanced Layer


picturesposted by

k neighborusers

(mlt=k)

CNNconvolutional

neuralnetwork



Accurate Layer

CNN-rand


map

Max Pooling

Fullconnected layer







119877119894119895 = 119901119894119895sum119899119895=1 119901119894119895 (7)

119878119894119895 = radicsum(119877119894119895 minus 119877119894)2

119899 minus 1 (8)




0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia



0 7 31 52 66

ratioprobabilitySD

0

005

01

015

02

025

03

035

04

045

05














UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia







UA 2 4 1 1 1 0 0 0 0UB 3 0 1 0 1 4 2 1 0UC 0 0 0 0 1 3 3 2 1




119875119894119895 = (119897119894119895)119872times119873 119894 = 1 2 119872 119895 = 1 2 119873 (9)




(10)








UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





UA 0 02949 00426UB 02949 0 07578UC 00426 07578 0

0

1

2

3

4

5

6

7

8

0 1 2 3 4 5 6 7 8 9

Tim

e sp

an(d

ay)

Spot Tag

U1U2U3


Gym

Hospital

Bank

Museum Library

Starbucks

Bowlingalley

Restaurant

Natatorium

pool

５

５B５C






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia






CNN

CNN

CNN

User 1

User 2

User 3




5 Experiments


















Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia













Pr119890119888119894119904119894119900119899 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 1198711198731198771003816100381610038161003816119872 (11)

Re119888119886119897119897 = 1003816100381610038161003816119871V119894119904119894119905119890119889 cap 11987111987311987710038161003816100381610038161003816100381610038161003816119871V1198941199041198941199051198901198891003816100381610038161003816 (12)








0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia





0

01

02

03

04

05

06

1 2 3 4 5

Precision

HRDLSEERVPOIPACE

(a)

0

002

004

006

008

01

012

014

016

018

1 2 3 4 5

Recall

PACEVPOISEERHRDL

(b)

00000

00200

00400

00600

00800

01000

01200

1 2 3 4 5

F1-Score

PACEVPOISEERHRDL

(c)









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia









Data Availability




Acknowledgments


References





















































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia











































RoboticsJournal of




VLSI Design



Shock and Vibration







Journal of



Volume 2018



Volume 2018


Journal of




SensorsJournal of



RotatingMachinery





Propagation






Hindawi


Advances in

Multimedia


deep cnn-assisted personalized recommendation over big...

Documents