 Research
 Open Access
 Published:
Providing privacy preserving in next POI recommendation for Mobile edge computing
Journal of Cloud Computing volume 9, Article number: 10 (2020)
Abstract
Point of interest (POI) recommendation can benefit users and merchants. It is a very important and popular service in modern life. In this paper, we aim to study the next new POI recommendation problem with the consideration of privacy preserving in edge computing. The challenge lies in capturing the transition patterns between POIs precisely and meanwhile protecting users’ location. In this paper, first, we propose to model users’ checkin sequences with their latent states based on HMM, and EM algorithm is used to estimate the parameters of the model. Second, we propose to protect users’ location information by a weighted noise injection method. Third, we predict users’ next movement according to his current location based on Forward algorithm. Experimental results on two largescale LBSNs datasets show that our proposed model without noise injection can achieve better recommendation accuracy than several stateoftheart techniques, and the proposed weighted noise injection approach can achieve better performance on privacy preserving than traditional one with a little cost on accuracy.
Introduction
Mobile edge computing (MEC) is a “hardware + software” system that reduces network operations and service delivery delays by providing IT service environments and cloud computing capabilities on the edge of the mobile network. Its technical features include “proximity, location awareness, high bandwidth and low latency”. MEC can provide users with local video, AR/VR, user positioning, video QoS optimization, video monitoring, traffic analysis and other services [1,2,3]. A typical application of MEC is locationbased social media networks, where users can checkin and discuss their experiences. The checkin data of users can be used to study the life patterns for improving the performance of POI recommendation. POI recommendation provides recommendations of places to users, which is a popular service in LBSN.
Traditional POI recommendation usually recommends POIs to users according to the location, time, others’ comments or their friends’ interest. Being different from traditional POI recommendation, the next POI recommendation additionally considers users’ checkin sequential transitions. Different kinds of people may exhibit different sequential transition patterns. For example, for students, after studying in the classroom, they may prefer to participating in intense outdoor activities or having meals in the canteen. While for workers in workdays, after working at offices, they may prefer to going shopping or going home. Meanwhile, we observe that users always visit new POIs that they have never visited before, so it is meaningful to mine the sequential transition patterns for suggesting new next POIs to users.
However, the computing power of cloud data centers has been centralized at the edge of the network by MEC. On one hand, MEC infrastructure is usually deployed at the edge of the network, such as wireless base stations, making it more vulnerable to unsafe environments. On the other hand, MEC will adopt open application programming interface (API), open network function virtualization (NFV) and other technologies, and the introduction of openness will easily expose MEC to external attackers [4, 5]. MEC services are faced with certain security risks due to limited resources and capabilities of edge facilities. Therefore, when users check in on LBSNs, specific location information is needed to submit to the edge server, the trajectory involved in users’ checkin data may reveal users’ privacy, such as home address, living habits, and social relations. If such information is leaked to malicious attackers, the user will be exposed to a serious threat. Therefore, it is necessary to protect users’ current location when recommending POIs to users.
On one side, in order to recommend POIs with high accuracy, many recent studies have been conducted, including temporal influence enhanced POI recommendation, geographical influence enhanced POI recommendation, contentaware POI recommendation, and social influence enhanced POI recommendation. However, the sequential transition patterns between successive checkins has not yet wellstudied. The next POI that a user will visit may be correlated with the former POI where he has been, and a large number of sequential transitions between POIs may exhibit a certain transition patterns for specific people, while such patterns may change with the time, therefore, it is a challenge to improve the performance of next POI recommendation by capturing such sequential transition patterns precisely.
On the other side, researchers have proposed many privacy preserving methods to solve the problem of privacy disclosure, including Kanonymity, cryptography and obfuscation. Kanonymity is easy to implement; however, the original location is still easy to be recognized. Cryptography has a strong protection ability; however, it needs a large amount of computational cost so that it is hard to apply. Obfuscation is achieved by injecting noise into the original data; however, many existing obfuscation methods always result in low data availability. Therefore, it is also a challenge to find a tradeoff between recommendation accuracy and privacy protection by designing a sound way of noise perturbation.
In this paper, we aim to achieve the satisfying performance of next POI recommendation and meanwhile prevent user’s location information from being leaked in edge computing. First, we propose to use Hidden Markov Model (HMM) to capture user’s latent state and transition patterns from the successive checkin data, and the model parameters are estimated by EM algorithm. Second, we propose to provide weighted noise perturbation of users’ current locations based on the distance between user’s current location and the nearest center. Finally, we propose to use the Forward algorithm for predicting the probability of users’ next movement. We evaluate our proposed approach by extensive experiments on largescale LBSN datasets and the results demonstrate that our proposed approach can not only effectively prevent the leakage of user’s location privacy, but also have a satisfying accuracy of POI recommendation.
The remainder of the paper is organized as follows. Related work is discussed in Section “Related Work”. The proposed approach is illustrated in Section “Proposed Method”. The experimental results are presented in Section “Experiment”. And finally, the conclusions and future work are given in Section “Conclusion and Future Work”.
Related work
Many researchers pay attention to the development of POI recommendation because of the variety of applications in real life. POI recommendation was studied on the checkin behavior of many users [6]. The research of POI recommendations can be divided into four kinds: 1) Temporal influence enhanced POI recommendation, which mainly investigates the influence of time on checkin data. Yuan et al. establish a new model based on the fact that users visit different places at different times [7]. Yin et al. study successive checkins based on users’ preference as well as the temporal context [8]. Although these works considered the temporal information, the temporal relationship between successive checkins has not been well studied. 2) Geographical influence enhanced POI recommendation, which studies “the phenomenon of geographical clustering” of checkins for improving POI recommendation accuracy. Liu et al. propose a framework by combining Bayesian nonnegative matrix factorization with geographical influence to achieve a good performance of POI recommendation [9]. Ye et al. explore the geographical influence of checkins and propose a model that incorporates three factors: geographical impact, social impact and user preference [10]. 3) Contentaware POI recommendation, which detect users’ current locations by analyzing their published tweets and rank POIs by analyzing user’s comments on them. Chen et al. propose a method to detect user preference from text content and establish a relationship between user preference and POI [11]. Gao et al. integrate user sentiment information and POIrelated contents into POI recommendation system [12]. Due to the fact that many text contents in LBSN are vague, semantic analysis is a difficult research problem for POI recommendation. 4) Social influence enhanced POI recommendation, researchers study social impact because many people believe that friends often have many common preferences in locations. By studying social impact, the quality of recommendation could be enhanced. However, there are also studies showing that many friends have little common interests in terms of POI visits [13] . And E. Cho et al. report their findings that social relationships have a greater impact on long distance journey [14].
The next new POI recommendation is a newly emerging issue and even challenging. There exist only a few works on this research issue in the literature. Feng et al. propose a personalized ranking metric embedding method to study users’ checkins for next new POI recommendation [15]. Cheng et al. establish a model for exploring the order of location visits [16]. Feng et al. propose a POI2vec algorithm to jointly learn the latent representation for users and POIs, and then capture users’ preference and POI sequence to improve the accuracy [17]. Oppokhonov et al. find the user’s current location and then recommends new POIs based on collaborative filtering [18]. Since users’ preferences are changing with the time going on and the checkin data exhibits users’ latent states, although the above studies have considered the sequential influence, these studies have not yet well studied the latent states and users’ transition patterns hidden in the checkin data.
In order to relieve the privacy disclosure in POI recommendation, researchers have proposed many traditional privacy protecting methods, which can roughly divided into three kinds: 1) Kanonymity is a general privacy preserving method, which makes it possible for attackers to identify the target user with 1/k probability by ensuring the targeted user is indistinguishable from k set of users [19]. Gedik et al. study a customized framework to support kanonymity with variable k to meet the personalized requirements of location privacy preserving [20]. However, the location of the target user can still be easily recognized. 2) Cryptography provides a strong privacy preserving based on encryption and decryption mechanisms [21]. The data is firstly encrypted using some algorithms at the client and then transferred over the network. Liu et al. use partially homomorphic encryption technology to prevent users’ location information from leaking [22]. However, due to the large calculation and the complexity of security protocols, this method is difficult to apply widely. 3) Obfuscation uses a random value to replace a certain percentage of the user’s information for achieving the purpose of privacy protection [23]. Ardagna et al. use different obfuscation operators for protecting users’ location information, which can be used individually or in combination [24]. Polat et al. propose a randomized perturbation, where randomness noises following a specific distribution are injected into the original data [25].
Although the existing methods can provide privacy protection to a certain extent, the accuracy of POI recommendation has been decreased due to the decrease of data availability, in addition, it is hard to determine the magnitude of noise to be injected and the proportion of information to be replaced. So achieving the balance between location privacy protection and data availability is a difficult task.
Proposed method
Problem definition
U = {u_{1}, u_{2}, …, u_{s}} denotes a set of LBSN users, and I denotes POI, which is geocoded by {longitude, latitude, category}. The track of user u_{i} before time t is represented by \( {O}_{u_i}=<{\mathrm{I}}_{u_i1},{\mathrm{I}}_{u_i2},\dots, {\mathrm{I}}_{u_i\mathrm{t}}> \) .Given the tracks of all users \( \left\{{O}_{u_1},{O}_{u_2},\dots, {O}_{u_s}\right\} \), our goal is to precisely recommend topk POIs for users’ next movement at time t + 1 and meanwhile effectively protect the location privacy of users.
Approach overview
Figure 1 shows that our proposed method mainly includes three steps: constructing sequential model, injecting noise and constructing prediction model. 1) In the first step, we adopt HMM to model users’ checkin data as POI sequential transitions. It is meaningful to capture POI sequential transition since a user’s subsequent movement behavior is highly influenced by the previously visited locations. 2) In the second step, we extract checkin centers from the checkin data first, and then we perturb users’ current location by injecting Laplace noise according to the distance between users’ location and the nearest checkin center. The improved noise injection way prevents user’s location information from leaking to malicious attackers. 3) In the third step, we predict POIs for users’ next movement according to the perturbed user’s location by the forward algorithm.
Constructing sequential model
The base of sequential model is motivated by the following behavior in modern life:
(1) The checkin location is closely related to people’s current state. For instance, people may keep the state of “eating breakfast in the morning” or “writing code in the afternoon”. In the former state, people may checkin at coffeehouse, but in the latter state, people may checkin at the office.
(2) The change of state highly depends on the previous state. The state is a description of a series of behaviors over a period of time. People’s state will not change suddenly because many people have a regular lifestyle. In other words, over a period of time, the state changing rules for regular users will remain stable. For example, after eating breakfast, people is most likely to go to work. Whether she is going to do architectural design on a business trip or at the office, she is going to do her work. Therefore, it is appropriate to apply a deterministic finite automaton for describing people’s state changing rules [26].
The above phenomena are the base of the proposed sequential model. Hidden Markov Model (HMM) is a doubly stochastic process with an underlying stochastic process that is hidden, but can only be observed through another set of stochastic processes that produce the sequence of observed symbols, it manifests a rather sequentially changing process, the property of the process is usually held pretty steadily, except for minor fluctuations, for a certain period of time, and then, at certain instances, change to another property [27]. This property in HMM is very similar to users’ states in real life. Since the lifestyle of many people is regular, and the changing rule of user’s states will be stable over a period of time. Therefore, it is appropriate to adopt HMM to model users’ checkin data for POI sequential transitions.
Modeling the checkin data by HMM
Firstly, to simplify the expression, we provide the parameter definitions as shown in Table 1.
Estimating parameters by EM algorithm
The main task of this section is to estimate three important parameters: the initial state distribution π, the state transition probability distribution T, and the preference probability distribution P.
Generally, it is considered that the POI sequential record O is the most likely time series when given the three important parameters, so people can maximize the probability Pr(O_{u}/u) to get the parameters. The Expectation Maximization (EM) algorithm is always used to solve optimization problems. We represent α_{t}(i) as a forward vector, this vector denotes the probability of generating a POI sequence <I_{1}, I_{2}, …, I_{t}> in the state q_{i} at time t. Similarly,we represent β_{t}(i) as a backward vector, this vector denotes the probability of generating a POI sequence <I_{t + 1}, …, I_{L}> when current user state is q_{i} at time t. So we can formulate forward vector and backward vector recursively:
ξ_{t}(i, j) is the probability that user state q_{i} transforms to q_{j}, which can be denoted with α and β:
And γ_{t}(i) represent that when user in the t_{th} step, the user state is q_{i}:
The three core parameters in POI recommendation can be denoted by ξ_{t}(i, j) and γ_{t}(i) as follows:
The parameters of the next iteration are obtained by maximizing the expectations of μ = {π, T, P}. The iteration ends when μ converges. Hence, μ can be calculated by EM algorithm.
Perturbing user’s location with noise injection
If users send their specific information to the LBSN, their private information may be revealed. We propose a checkin centers distance weighted noise injection algorithm to protect user’s locations information in this paper. In order to prevent users’ location information from being leaked, we inject different quantity of noise into the latitude and longitude of the user’s location based on the distance between user’s current location and their nearest checkin center. Then, we use the perturbed data for the next POI recommendation.
Noise injection is an important method of differential privacy mechanism. This method is to inject noise into the initial data and achieve the goal of privacy preserving by perturbing the object data. Laplace mechanism is a method for differential privacy protection [28]. The Laplace mechanism can be used for numerical datasets, it ensures the validity of εdifferential privacy by injecting noise into the original dataset D with Laplace distribution:
, where Y ~ Lap(∆f/ε). However, such method of noise injection injects the same level of noise into the location data, resulting in low data availability and large impact of noise on the results.
An important feature of checkin data is that it obeys the Gaussian distribution around several checkin centers [29], as Fig. 2 shows. Another important feature is that most checkins occur in nearby areas, and the probability that a user will visit a POI is in inverse proportion to the distance between users’ location and the nearest checkin center [15]. Based on the two important features of checkin record, we propose a checkin center distance weighted noise injection method to protect users’ location. Firstly, we find out all the checkin centers according to the greedy clustering algorithm proposed by Chen, which scans from the most visited POIs, and then merges POIs with a distance no larger than d kilometers into one region, if the ratio of the total number of checkin behavior in the area to the total number of checkin behavior of the user is greater than the threshold θ, then the checkin positions is set to a region and its center is determined [29]. Secondly, we get the distance between user’s current location and their nearest checkin center, then we normalize the distance and convert it to a weight value. Thirdly, we inject different magnitudes of noise into the location information based on the weight value. The closer to the checkin center, the lower level the noise injected, since the closer to the checkin center, the more prosperous the place is, and the more users may checkin, the higher the population density is. Even if a small level of noise is added, it is difficult for an attacker to distinguish the target user. Motivated by Dou [30], the weighted noise injection method is shown in formula (9):
, where I_{ui} represents the realtime location of user u_{i}, and W_{ui}ϵ[0, 1] denotes the distance between the user’s current location and his nearest checkin center.
Then, we provide the verification that formula (9) can keep differential privacy. To ensure that the algorithm’s privacy budget ε is within the threshold while using the differential privacy technique multiple times, this paper needs to use the sequence combination property and parallel combination property to prove that the proposed algorithm satisfies the requirements of differential privacy. We inject the weighted Laplace noise into the query function \( \mathrm{f}\left({\mathrm{I}}_{u_i}\right)=\mathrm{f}\left({\mathrm{I}}_{u_i1}\right)+\mathrm{f}\left({\mathrm{I}}_{u_i2}\right)+\dots +\mathrm{f}{\left({\mathrm{I}}_{u_it}\right)}^T \), and the output function is:
, where \( \Delta \mathrm{f}=\mid \mathrm{f}\left({I}_{u_i}\right)\mathrm{f}\left({\mathrm{I}}_{ui}^{\prime}\right)\mid \),and the perturbed data is denoted as \( {\mathrm{I}}_{ui}^{\prime } \). According to relevant definitions, the following formula is the condition that algorithm M satisfies differential privacy:
And then, we will prove that formula (10) satisfies formula (11).
Assuming that \( \mathrm{f}\left({\mathrm{I}}_{u_i}\right)={\left({\mathrm{I}}_{u_i1},{\mathrm{I}}_{u_i2},\dots, {\mathrm{I}}_{u_it}\right)}^T \)
The output variable is \( \mathrm{O}={\left({\mathrm{y}}_{u_i1},{\mathrm{y}}_{u_i2},\dots, {\mathrm{y}}_{u_it}\right)}^T \), and then,
For \( \left{\mathrm{y}}_{u_ij}\Delta {I}_{ui}\right\mid {\mathrm{y}}_{u_ij}\mid \), by the definition of absolute value inequality, we can get the following formula:
Substituting the above inequality by
Then we can get the following formula,
Therefore, each noise injection process satisfies the εdifferential privacy. Since the checkin information contains locations of many users, according to the parallel combination of differential privacy, the algorithm satisfies max(w ∗ ε_{i})differential privacy. Since max(w) = 1, the method satisfies εdifferential privacy.
Recommending the next new POIs
As users would prefer to visiting the near POIs rather than the POIs far way, we can see that geographical factor affects users’ visiting behavior. To incorporate the geographical influence into POI recommendation, we construct a circular region, the center of which is the user’s current location at time t, and all the POIs in the circle are the candidates for the user’s next movement. The user’s previous sequential checkin POI record is represented as <I_{1}, I_{2}, …, I_{t,}I_{t + 1}>, and the candidate for user’s next movement is represented as I_{et + 1}, and then we can calculate the probability of O = < I_{1}, I_{2}, …, I_{t,}I_{t + 1}> for each POI in the region based on Forward algorithm. By ranking the probability of O = < I_{1}, I_{2}, …, I_{t,}I_{t + 1}> for each POI in this region, we can get Top k POIs for users’ recommendation list. Given the parameter μ = {π, T, P} and the POI observation sequence O_{u} = < I_{1}, I_{2}, …, I_{t,}I_{t + 1}>, then the probability of Pr(O_{u} μ) can be calculated by Forward algorithm. The Forward algorithm uses the forward vector α_{t}(i) that mentioned above to reduce the computational cost. Forward algorithm is to use the recursive relationship of formula (1) to get the forward vector in next step. And the required probability is given by
, where
The specific process of the POI prediction algorithm is given as follows:
Experiment
In this part, we evaluate the following: (1) how is the proposed approach of POI recommendation in comparison with the baseline and other stateoftheart recommendation techniques? (2) how do the scale of area and the number of latent states affect the model accuracy? (3) how is the performance of proposed model when providing privacy preserving?
Checkin data characteristics
We choose two largescale datasets for experiments, Gowalla and Foursquare. Gowalla provides a public API for crawling information about all users, including checkin history record. The Foursquare dataset is provided by Yin [31]. Each checkin record contains three columns of information: user, POI, time. The specific location of each POI is related to longitude and latitude. Table 2 shows the basic statistics of two datasets:
Evaluation metrics
In order to investigate the quality of the next new POI recommendation, we adapt the standard metric: precision@k. It shows the percentage of places in the next POI candidate list has been visited for each user. Formally, R_{u}(k) denotes the k recommended new POIs in candidate list and \( {\mathrm{V}}_u^{new} \) denotes the visited new places of user u.
Evaluated methods
We compare our proposed approach with the following methods:
(1) CLB: Current locationbased recommendation proposes a model that captures sequential influence and geographical influence. This model finds the user’s current location and then recommends new POIs based on collaborative filtering [18].
(2) PMC: Firstorder Markov Chain considers sequential influence, it uses the time series of POIs visited by all users and the last place visited by the target user to make POI recommendation predictions [16].
(3) PRME: this approach combines two kind of distances for POI prediction, it models users’ checkin sequences in hidden space, moreover, it considers user’s preference and geographical influence [15].
In the experiments, for the purpose of evaluating the quality of different algorithms, we split the two datasets into two nonoverlapping sets: for each user, the first 10 months checkin data is train set and the remaining 2 months checkin data is test set. The settings of the number of user’s latent state are 5 and 8 for Foursquare and Gowalla, respectively. The circular region scale are 13 and 15 for Foursquare and Gowalla, respectively. We set the threshold θ = 0.02 and the region radius d = 15 in greedy clustering algorithm to find all the checkin centers.
Comparison on accuracy
Table 3 shows the comparison results between our model and the baseline methods, the result shows that:
(1) Our proposed model has a better performance than CLB and FMC. In order to find the users with similar POI sequence record, CLB establishes a directed graph with users’ POI checkin sequences, however, it is hard to solve the problem of data sparsity. One possible explanation for FMC’s poor performance could be that it only considers the sequential influence of the last visited POI, however, FMC ignores the sequential influence of other recently visited POIs.
(2) PRME gets improved performance compared with CLB and FMC, since PRME develops a Metric Embedding algorithm to model the sequential transition of all visited POIs, moreover, it also captures geographical influence.
(3) Our proposed method consistently outperforms PRME, improving at least 7% and 5% over PRME for Gowalla and Foursquare respectively. According to our analysis, experimental results shows that the user’s latent state in HMM is effective to capture users’ actions and calculate users’ POI transition probability, moreover, the geographical influence is beneficial for the next new POI recommendation.
Effect of parameters
Effect of the scale
In this section, we aim to investigate the parameter scale that affects the quality of the next new POI recommendation. The scale is the radius of the circular region. It determines the amount of surrounding buildings and POIs near the user. Figure 3 shows the precision under different values of scale on two data sets. The result indicates that our method gets its maximum precision when the scale is 13 for Foursquare dataset and 15 for Gowalla dataset. As the value of scale increases from 11, the accuracy increases at first; however, when the scale reaches a certain value, the precision of the recommendation system begins to decline. This is because when the scale is too small, the POIs in the area are too little to meet the needs of users; when the scale is too large, there are too many POIs in the area for users to choose, which also leads to lower accuracy.
Effect of the number of user state
In this section, we aim to investigate the number of user states that affects the quality of the next new POI recommendation. We describe the number of groups of user behavior over a period of time as the number of user states. In Fig. 4, the result indicates that when the number of user states is 5 for Foursquare and 8 for Gowalla, the approach gets the maximum precision. The number of a regular user’s state is relatively small in real life, and a regular user state number provides generalization ability for this method.
Experiments on POI recommendation with privacy preserving
Comparison on perturbed data
In this section, we aim to compare our proposed privacy preserving algorithm based on the weight of distance(WD) with the traditional noise injection algorithm in terms of recommendation precision.
When the privacy parameter ε takes different values, the POI recommendation precision of the two methods is shown in Fig. 5. The privacy budget ε is an important parameter for determining the level of privacy preserving in differential privacy. By varying the privacy parameter ε from 0.1 to 1.0, we investigated the quality of our proposed algorithm at various privacy protection levels.
As the privacy budget ε increases, the precision of the POI recommendation increases. Figure 5 shows that our proposed algorithm performs better in POI recommendation than traditional noise perturbation algorithm. In our proposed algorithm, different quantities of noise were injected into user’s location based on the different distance between the user and the nearest checkin center. Malicious attackers can easily identify the current location of target user from the area with less POIs, therefore, the location with a long distance between the target user and the checkin center need to be added with a higher level of noise. As the quantity of noise injection increases, the data availability decreases. Our method has a better performance because our proposed noise injection method improves the availability of data.
Cost of privacy preserving
In this section, we aim to analyze what percentage of precision is lost as the cost of privacy preserving based on the weight of distance(WD). The precision of our proposed model without noise injection is 0.038 for Gowalla dataset and 0.035 for Foursquare dataset when k = 10. In our proposed POI recommendation with perturbed data, the precision changes as the privacy budget ε changes. Therefore, we analyze the cost of privacy preserving when the privacy budget ε takes different values.
Figure 6 shows how much percentage of precision is lost as the cost of privacy preserving by varying the privacy budget ε from 0.1 to 1.0. As the privacy budget ε increases, the cost of privacy preserving decreases. This is because when the privacy budget ε is larger, the level of privacy protection is lower, and the less noise is injected, and then the cost of privacy preserving is lower. When ε = 0.1, the cost of privacy preserving is 28.68%; when ε = 1, the cost of privacy preserving is 8.68%. Therefore, our proposed privacy preserving method protects users’ location privacy with a little cost on precision.
Conclusion and future work
To address the problem of next new POI recommendation that supports privacy preserving, in this paper we propose a HMMbased POI sequential model to capture the successive checkin behavior by exploring the user’s latent state for each user, and then we use EM algorithm to estimate the parameters of the model; we propose a checkin multicenter distance weighted noise injection method to protect users’ location; finally, we combine geographical information and sequential information to provide new POI recommendation services that support privacy preserving for users. Performance evaluation conducted on two datasets shows that our proposed approach without noise injection improves the recommendation precision compared with other stateofthe art methods and meanwhile protects user’s location information with little accuracy decrease.
For future work, we will try to consider different user states and periodic preference for users who have different lifestyles to further explore the capabilities of the next new POI recommendation.
Availability of data and materials
The datasets supporting the conclusions of this article are available in http://net.pku.edu.cn/daim/hongzhi.yin/
Abbreviations
 EM:

Expectation Maximization
 HMM:

Hidden Markov Model
 LBSNs:

LocationBased Social Networks
 MEC:

Mobile edge computing
 NFV:

Network function virtualization
 POI:

Pointofinterest
 WD:

Weight of distance
References
 1.
Chen Y, Deng S, Ma H, Yin J (2019) Deploying dataintensive applications with multiple services components on edge. Mob networks Appl:1–16
 2.
Gao H, Duan Y, Shao L, Sun X Transformationbased processing of typed resources for multimedia sources in the IoT environment. Wirel Networks:1–17
 3.
Kuang L, Yan X, Tan X et al (2019) Predicting taxi demand based on 3D convolutional neural network and multitask learning. Remote Sens 11:1265
 4.
Liao Z, Zhao B, Liu S et al (2019) A prediction model of the project lifespan in open source software ecosystem. Mob Networks Appl 24:1382–1391
 5.
Liao Z, Deng L, Fan X et al (2018) Empirical research on the evaluation model and method of sustainability of the open source ecosystem. Symmetry (Basel) 10:747
 6.
Zheng Y, Zhang L, Xie X, Ma WY (2009) Mining interesting locations and travel sequences from GPS trajectories. In: proceedings of the 18th international conference on world wide web. ACM, pp 791–800
 7.
Yuan Q, Cong G, Ma Z, et al (2013) Timeaware pointofinterest recommendation. In: proceedings of the 36th international ACM SIGIR conference on research and development in information retrieval. ACM, pp 363–372
 8.
Chen C, Yin H, Yao J, Cui B (2013) Terec: a temporal recommender system over tweet stream. Proc VLDB Endow 6:1254–1257
 9.
Liu B, Fu Y, Yao Z, Xiong H (2013) Learning geographical preferences for pointofinterest recommendation. In: Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pp 1043–1051
 10.
Ye M, Yin P, Lee WC, Lee DL (2011) Exploiting geographical influence for collaborative pointofinterest recommendation. In: Proceedings of the 34th international ACM SIGIR conference on Research and development in Information Retrieval. ACM, pp 325–334
 11.
Chen Y, Zhao J, Hu X, et al (2013) From interest to function: location estimation in social media. In: TwentySeventh AAAI Conference on Artificial Intelligence
 12.
Gao H, Tang J, Hu X, Liu H (2015) Contentaware point of interest recommendation on locationbased social networks. In: TwentyNinth AAAI Conference on Artificial Intelligence
 13.
Ye M, Yin P, Lee WC (2010) Location recommendation for locationbased social networks. In: Proceedings of the 18th SIGSPATIAL international conference on advances in geographic information systems. ACM, pp 458–461
 14.
Cho E, Myers SA, Leskovec J (2011) Friendship and mobility: user movement in locationbased social networks. In: proceedings of the 17th ACM SIGKDD international conference on knowledge discovery and data mining. ACM, pp 1082–1090
 15.
Feng S, Li X, Zeng Y, et al (2015) Personalized ranking metric embedding for next new POI recommendation
 16.
Cheng C, Yang H, Lyu MR, King I (2013) Where you like to go next: successive pointofinterest recommendation. In: TwentyThird international joint conference on Artificial Intelligence
 17.
Feng S, Cong G, An B, Chee YM (2017) Poi2vec: geographical latent representation for predicting future visitors. In: ThirtyFirst AAAI Conference on Artificial Intelligence
 18.
Oppokhonov S, Park S, Ampomah IKE (2017) Current locationbased next POI recommendation. In: Proceedings of the International Conference on Web Intelligence. ACM, pp 831–836
 19.
Sweeney L (2002) kanonymity: a model for protecting privacy’International Journal on Uncertainty, Fuzziness and Knowledgebased Systems 10, 5 (2002) 557–570
 20.
Gedik B, Liu L (2004) A customizable kanonymity model for protecting location privacy. Georgia Institute of Technology
 21.
Gentry C, Boneh D (2009) A fully homomorphic encryption scheme. Stanford University Stanford
 22.
Liu A, Wang W, Li Z et al (2017) A privacypreserving framework for trustoriented pointofinterest recommendation. IEEE Access 6:393–404
 23.
Duckham M, Kulik L (2005) A formal model of obfuscation and negotiation for location privacy. In: International conference on pervasive computing. Springer, pp 152–170
 24.
Ardagna CA, Cremonini M, di Vimercati SDC, Samarati P (2009) An obfuscationbased approach for protecting location privacy. IEEE Trans Dependable Secur Comput 8:13–27
 25.
Polat H, Du W (2003) Privacypreserving collaborative filtering using randomized perturbation techniques. In: Third IEEE International Conference on Data Mining. IEEE, pp 625–628
 26.
Xiang Z, Deng S, Liu S et al (2016) Camer: a contextaware mobile service recommendation system. In: 2016 IEEE international conference on web services (ICWS). IEEE:292–299
 27.
Rabiner LR, Juang BH (1986) An introduction to hidden Markov models. Ieee Assp Mag 3:4–16
 28.
Dwork C, Roth A (2014) The algorithmic foundations of differential privacy. Found trends®. Theor Comput Sci 9:211–407
 29.
Cheng C, Yang H, King I, Lyu MR (2012) Fused matrix factorization with geographical and social influence in locationbased social networks. In: TwentySixth AAAI Conference on Artificial Intelligence
 30.
Dou K, Guo B, Kuang L (2019) A privacypreserving multimedia recommendation in the context of social network based on weighted noise injection. Multimed Tools Appl 78:26907–26926
 31.
Yin H, Wang W, Wang H et al (2017) Spatialaware hierarchical collaborative deep learning for POI recommendation. IEEE Trans Knowl Data Eng 29:2537–2551
Acknowledgements
We want to thank the authors of the literature cited in this paper for contributing useful ideas to this study.
Funding
This work was funded by National Natural Science Foundation of China (No. 61772560), National Key R&D Program of China (No. 2018YFB1003800), National Natural Science Foundation of China (No. 61902236).
Author information
Affiliations
Contributions
Li Kuang, Shenmei Tu and Yangqi Zhang have written this paper and have done the research which supports it. Li Kuang, Shenmei Tu and Xiaoxian Yang has collaborated in the conception, research and design of the paper. All authors read and approved the final manuscript.
Author information
Li Kuang received her B.S. and Ph.D. degrees in computer science from Zhejiang University, Hangzhou, China, in 2004 and 2009 respectively. She is currently a professor at the school of Computer Science and Engineering, Central South University. Her research interests include Service Computing, Data Mining, Crowdsourcing software Ecosystem. She has authored or coauthored over 40 journal and conference publications.
Shenmei Tu received her bachelor’s degrees in software engineering from Jiangxi Agricultural University, Nanchang, China, in 2017. She is currently pursuing a graduate student in software engineering at the school of Computer Science and Engineering, Central South University. Her research interests include service computing, data mining, recommendation system and privacy preserving.
Yangqi Zhang received his bachelor’s degrees in software engineering from Central South University, Changsha, China, in 2019. He is currently pursuing a graduate student in software engineering at the school of Computer Science and Engineering, Central South University. His research interests include service computing, data mining and privacy preserving.
Xiaoxian Yang, received the Ph.D. degree in Management Science and Engineering from Shanghai University, Shanghai, China, in 2017. She is currently an assistant professor at Shanghai Polytechnic University, China. Her research interests include business process management, and formal method.
Corresponding author
Ethics declarations
Competing interests
The authors declare that there is no conflict of interest regarding the publication of this manuscript.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Kuang, L., Tu, S., Zhang, Y. et al. Providing privacy preserving in next POI recommendation for Mobile edge computing. J Cloud Comp 9, 10 (2020). https://doi.org/10.1186/s1367702001583
Received:
Accepted:
Published:
Keywords
 HMM
 Sequential transition patterns
 Latent state
 Privacy preserving
 POI recommendation