Semantic trajectory-based high utility item recommendation system

Semantic trajectory-based high utility item recommendation system

Jia-Ching Ying

, Huan-Sheng Chen

, Kawuu W. Lin

, Eric Hsueh-Chan Lu

, Vincent S. Tseng

,

Huan-Wen Tsai

, Kuang Hung Cheng

, Shun-Chieh Lin

a

Department of Computer Science and Information Engineering, National Cheng Kung University, 1, University Road, Tainan City 701, Taiwan, ROC b

Department of Computer Science and Information Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung 80778, Taiwan, ROC c_{Cloud Service Technology Center, Industrial Technology Research Institute (ITRI South), Tainan, Taiwan, ROC}

a r t i c l e

i n f o

Keywords: Trajectory database Trajectory pattern High utility pattern Items recommendation Semantic prediction Data mining

a b s t r a c t

The topic on recommendation systems for mobile users has attracted a lot of attentions in recent years. However, most of the existing recommendation techniques were developed based only on geographic features of mobile users’ trajectories. In this paper, we propose a novel approach for recommending items for mobile users based on both the geographic and semantic features of users’ trajectories. The core idea of our recommendation system is based on a novel cluster-based location prediction strategy, namely TrajUtiRec, to improve items recommendation model. Our proposed cluster-based location prediction strategy evaluates the next location of a mobile user based on the frequent behaviors of similar users in the same cluster determined by analyzing users’ common behaviors in semantic trajectories. For each location, high utility itemset mining algorithm is performed for discovering high utility itemset. Accord-ingly, we can recommend the high utility itemset which is related to the location the user might visit. Through a comprehensive evaluation by experiments, our proposal is shown to deliver excellent performance.

Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction

Recommendation systems provide people convenient access to the products they might be interested in. However, the traditional recommendation systems only focus on the virtual world. The market of location-based services, including navigational services, traffic management and location-based advertisement, have grown rapidly in recent years. These services bring users from virtual world to physical word. Due to the needs of effective marketing and efficient system operations, it is beneficial for these LBSs to be able to forecast the activities a user may perform at the next location to visit. Intuitively, we can recommend some items related to the user’s next location. Here, the ‘‘item’’ is anything that is sold by stores located in the location where users might visit, such as merchandise, goods, services, etc. Thus, effective items recommen-dation and effective location prediction techniques for LBSs target-ing on mobile users are desirable. Accordtarget-ing to above-mentioned reasons, in this paper, we propose a recommender which can not

only predict users’ next location but also recommend items which are sold by stores located in users’ next locations.

In fact, such new breed of items recommendation methods, considering users’ current and next locations, is called location-based recommendation and has been discussed in many existing work (Bao, Zheng, & Mokbel, 2012; Eagle, Pentland, & Lazer, 2009; Lu, Lee, & Tseng, 2012; Lu, Tseng, & Yu, 2011; Taiwan Tourism Bureau, 0000). The location-based recommendation meth-ods usually use the frequent moving behaviors of users to predict the next move of a user and recommend the items which are related to that location. To make accurate location prediction, the location-based recommendation systems always not only record users’ GPS trajectories but also mine the frequent moving behaviors from the users’ GPS trajectories.Fig. 1shows some examples of the GPS trajec-tory, which typically consists of a sequence of spatio-temporal points (in form of latitude, longitude, and time). Among the trajec-tory-based recommendation methods, mobile sequential pattern mining techniques (Lu & Tseng, 2009; Monreale, Pinelli, Trasarti, & Giannotti, 2009) have been widely used for analyzing patterns in mobile user movement data sets. However, they tend to predict pop-ular locations where most people visited, leading to the imbalanced data problem (Yen, Lee, Lin, & Ying, 2006). Additionally, these pat-tern-based Prediction methods usually make a Prediction only if an anticipated movement has a full match with the prefix of a

http://dx.doi.org/10.1016/j.eswa.2014.01.042 0957-4174/Ó 2014 Elsevier Ltd. All rights reserved.

⇑Corresponding author. Fax: +886 6 2747076.

E-mail addresses:[email protected](J.-C. Ying),[email protected](H.-S. Chen), [email protected] (K.W. Lin), [email protected] (E.H.-C. Lu), [email protected](V.S. Tseng),[email protected](H.-W. Tsai),itriA00432@ itri.org.tw(K.H. Cheng),[email protected](S.-C. Lin).

Contents lists available atScienceDirect

Expert Systems with Applications

pattern, leading to loss of recall in predictions and recommenda-tions. As the result the traditional location-based recommendation methods are always to recommend the items which are related to popular locations.

Although the issues of discovering mobile users’ frequent pat-terns in their trajectories have been discussed in the literature, existing studies mostly consider only on the geographic features of user trajectories (Lu & Tseng, 2009; Monreale et al., 2009). No-tice that a geographic trajectory typically consists of a sequence of geographic points (represented as hlatitude, longitudei) tagged with timestamps. As the result, the frequent pattern of user move-ment behavior based on geographic trajectory is constrained by the geographic properties of the trajectory data. For example, asFig. 1

shows, the geographic distance and shape between Trajectory1and

Trajectory2is closer and more similar than that between

Trajec-tory1 and Trajectory3. Thus, some location prediction and items

recommendation techniques would predict the destination of Tra-jectory1based on its geographical similarity to Trajectory2.

Addi-tionally, such location prediction and items recommendation strategies only consider the previously visited locations and thus do not work well when previously unvisited locations are consid-ered. We argue that merely using geographic information to pre-dict the destination of a trajectory or a user’s next location is not sufficient.

The notion of semantic trajectory has been proposed byAlvares et al. (2007) and Bogorny, Kuijpers, and Alvares (2009). Basically, a semantic trajectory consists of a sequence of locations labeled with semantic tags (called semantic locations) to capture the landmarks passed byLiu, Wolfson, and Yin (2006), Ying, Lee, Weng, and Tseng (2011). These semantic tags of locations imply the activities being carried out in the trajectory. ConsiderFig. 1where trajectories are tagged with a number of semantic tags such as School, Park, etc. We observe that both Trajectory2and Trajectory3can be denoted

by the sequence hSchool, Bank, Hospitali, implying that the seman-tic behaviors of users in Trajectory2and Trajectory3are quite the

same. Thus, we exploit their similarity in visited semantic locations to predict the next locations of mobile users.

Besides, the existing location-based recommendation methods always recommend items which users or their most people fre-quently buy. It leads recommenders always recommend low profit item for users. As the result, recommendation might not benefit stores even the recommenders can precisely predict users’ next locations. Fortunately, high utility itemset mining has been pro-posed for discovering the itemsets with high benefits from transac-tion database. To improve the applicability of our recommendatransac-tion system, we can adopt this useful itemset mining algorithm for dis-covering high utility itemsets in each location. TakeFig. 1as exam-ple; if a user move along with Trajectory1, the system can

recommend some high utility foods for him before the user visit the Restaurant. Therefore, we have to predict not only where users

will go but also what users will want to do. Inherently, the seman-tics of location is related to what kind of activity user should per-form there. According to the semantics of predicted next location, we can recommend some items which are sold by the retailers in the location.

However, the traditional item recommendation methods do not fully consider both the users’ next locations and high utility item-sets. To support location prediction and items recommendation based on the semantic trajectories of mobile users, we propose a novel location prediction and items recommendation framework, called TrajUtiRec, to evaluate the next location of a user’s move-ment. The framework consists of two major modules: (i) offline mining module, and (ii) on-line location prediction and items rec-ommendation module. In the offline mining module, we adopt the notion of stay locations to represent the users’ movement behav-ior. To extract the semantic feature from individual user’s move-ment behavior, we mine the semantic trajectory patterns for each individual user. Moreover, we form user clusters based on the notion of semantic trajectory similarity we proposed. Further-more, we mine the frequent trajectory patterns of users in the same cluster based on their geographic features. Besides, we adopt high utility mining algorithm to discover high utility itemset of each location. In the on-line location prediction and items recom-mendation module, based on these semantic and geographic pat-terns, we develop a novel cluster-based location prediction and items recommendation technique to predict a mobile user’s next location. To our best knowledge, this is the first work on predicting a mobile user’s next location by exploiting both geographic and semantic features of trajectories. Through an experimental evalua-tion, we show that the proposed location prediction and items rec-ommendation approach delivers excellent performance.

The contributions of our research are fivefold.

 We propose the TrajUtiRec framework, a new approach for mobile users’ movement behavior mining and location predic-tion and items recommendapredic-tion. The problems and ideas in TrajUtiRec have not been explored previously in the research community.

 We employ the notion of semantic trajectory similarity we pro-posed to cluster similar users together.

 We employ the high utility itemset mining algorithm is per-formed for discovering high utility itemset of each location.  Based on the trajectory patterns and high utility itemsets, we

propose a novel location prediction and items recommendation strategy to predict a user’s next location.

 We use a simulation dataset in a series of experiments to eval-uate the performance of our proposal. The results show superior performance over other location prediction and items recom-mendation techniques in terms of precision and recall. The rest of this paper is organized as follows. We briefly review the related work in Section2and provide an overview of our loca-tion predicloca-tion and items recommendaloca-tion framework in Secloca-tion3. We detail the proposed semantic mining and geographic mining in Section4and describe our location prediction and items recom-mendation technique in Section5. Finally, we present the evalua-tion result of our empirical performance study in Secevalua-tion6 and discuss our conclusions and future work in Section7.

2. Related works

Many data mining studies have discussed the problems of pre-dicting the next location where a mobile user moves to. Personal-based prediction (Jeung, Liu, Shen, & Zhou, 2008; Yavas, Katsaros, Ulusoy, & Manolopoulos, 2005; Ye, Zheng, Chen, Feng, & Xie, 2009) and general-based prediction (Monreale et al., 2009; Morzy,

School Park Hospital Bank Restaurant

Trajectory

Trajectory

Trajectory

Geographic Point

School Park Hospital Bank Restaurant

Trajectory

Trajectory

Trajectory

Geographic Point

2006, 2007; Zheng, Zhang, Xie, & Ma, 2009a,b) are two approaches often adopted in this problem domain. The personal-based predic-tion approach considers movement behavior of each individual as independent and thus uses only the movements of an individual user to predict his/her next location. On the contrary, the gen-eral-based prediction makes a prediction based on the common movement behavior of general mobile users. In Jeung et al. (2008), propose an innovative approach which forecasts future locations of a user by combining predefined motion functions, i.e., linear or non-linear models that capture object movements as sophisticated mathematical formulas, with the movement pat-terns of the user, extracted by a modified version of the Apriori algorithm. InYavas et al. (2005), mine the movement patterns of an individual user to form association rules and use these rules to make location prediction. Additionally, they consider the sup-port and confidence in selecting the association rules for making predictions. InYe, Zheng, Chen, Feng, and Xie (2009), propose a no-vel pattern, called Individual Life Pattern, which is mined form individual trajectory data, and they uses such pattern to describe and model the mobile users’ periodic behaviors. InMorzy (2006), uses a modified version of Apriori algorithm to generate associa-tion rules, and inMorzy (2007), he uses a modified version of Pre-fixSpan algorithm to discover frequent patterns of users’ movements for generating the prediction rules. The matching func-tions employed in these previous works are based on the nofunc-tions of support and confidence. Although all of Morzy’s approaches have considered temporal information and location hierarchy, they do not take into account the semantic tags of locations. InMonreale et al. (2009), proposes a method aiming to predict with a certain level of accuracy the next location of a moving object. The move-ment patterns extracted for prediction covers three different movement behaviors, including order of locations, travel time, and frequency of user visits. InZheng et al. (2009a), uses a HITS-based model to mine users’ interesting location and detect users’ travel sequence to make locations prediction, and inZheng et al. (2009b), they consider the location correlation for generating the users’ interesting locations and travel sequence. Note that the above-mentioned prediction methods are based on geographic information only. On the contrary, our proposal predicts the next location of a user based on both geographic and semantic informa-tion in trajectories.

In recent years, a number of studies on semantic trajectory data mining have appeared in the literature (Alvares et al., 2007; Bogorny et al., 2009). InAlvares et al. (2007), propose to explore the geo-graphic semantic information to mine semantic trajectory patterns from mobile users’ movement histories. First, they discover the stops of each trajectory and map these stops to semantic landmarks to transform geographic trajectories into semantic trajectories. By applying a sequential pattern mining algorithm on semantic trajec-tories, they obtain frequent patterns, namely, semantic trajectory patterns, to represent the frequent semantic behaviors of mobile users. InBogorny et al. (2009), use a hierarchy of geographic seman-tic information to discover more interesting patterns. Noseman-tice that the notion of stops in the above-mentioned works only considers the as-pect of ‘stay’ in stops but not the ‘positions’ of these stops in geo-graphic space. As a result, many unknown stops are generated. For example, as shown inFig. 2, stop1c, stop2c, and stop3bare not

associ-ated with any semantic landmark and thus marked as Unknown. Hence, Trajectory1is transformed as the sequence hSchool, Park,

Un-known, Restauranti. From the figure, it is clear that stop1cis near the

Restaurant. Thus, in our work, by taking into account the geometric distribution of these stops, stop1cand stop1dare grouped together

such that the Trajectory1is transformed as the sequence hSchool,

Park, Restauranti instead.

Besides, a feature vector is proposed by Zheng to describe the semantics of each location. Based on the feature vector, the

seman-tic similarity between two mobile users could be calculated. In addition to the GPS trajectory, Ying, Lu, Lee, Weng, and Tseng (2010)also exploit the cell trajectory to derive the semantic simi-larity between two mobile users. The cell trajectory consists of a sequence of spatio-temporal points in form of cell station ID, arrive time, and leave time as shown inFig. 3. They propose a novel sim-ilarity measurement, namely, Maximal Semantic Trajectory Pat-tern Similarity (MSTP-Similarity) to evaluate the user similarity. As such, the similarity of two mobile users, even if they live in dif-ferent cities, may be evaluated based on their similar semantic tra-jectory patterns.

3. Overview of TrajUtiRec

With the notion of semantic trajectory, we propose a novel loca-tion predicloca-tion and items recommendaloca-tion framework, namely, TrajUtiRec, based on both the geographic and semantic features in trajectories. The proposed approach works for locations where the users may have never visited, e.g., a location in other cities. The TrajUtiRec framework consists of (1) an offline training module, and (2) an online location prediction and items recommendation module.

Fig. 4shows the framework and its flow of data processing. The idea is to explore the activities of mobile users, captured in seman-tic trajectories, to improve accuracy of location prediction and items recommendation. As shown, the training module includes three steps. The first step, called data preprocessing, transforms each user’s trajectories as stay location sequences. The second step, called semantic mining, extracts users’ semantic behaviors (as ‘semantic trajectory patterns’ which will be detailed later). It also obtains user clusters based on the semantic behavior similarity of users. The third step, called geographic mining, extracts the geo-graphic behaviors of users in each cluster (as ‘stay location pat-terns’ which will be detailed later). In the online module, we propose a scoring function to evaluate the probability for a location to be the next location. Here, we consider not only geographic

Trajectory

Stop

stop

stop

_stop

stop

stop

stop

stop

stop

stop

stop

stop

stop

Trajectory

Trajectory

Trajectory

Stop

stop

stop

_stop

stop

stop

stop

stop

stop

stop

stop

stop

stop

Trajectory

Trajectory

Fig. 2. An example of semantic trajectory.

Trajectory

Trajectory

Cell

Cell

Cell

Cell

Cell

Cell

Cell

Trajectory

Trajectory

Cell

Cell

Cell

Cell

Cell

Cell

Cell

C

C

e

e

l

l

l

l

Cell

C

C

e

e

l

l

l

l

l

l

Cell

C

C

C

C

e

e

l

l

l

l

Cell

C

C

e

e

l

l

l

l

Cell

C

C

e

e

l

l

l

l

Cell

C

C

e

e

l

l

l

l

Cell

C

C

e

e

l

l

l

l

Cell

information but also semantic information. First, we calculate the geographic score and derive several candidate paths. Then, the semantic score of each candidate path is evaluated. Finally, we compute a weighted average of geographic score and semantic score for each candidate path to select the most probable path for predicting the next location in a user’s move.

4. Offline training module

In this section, we propose an approach to extract the users’ fre-quent movement behaviors which includes the semantic behavior information for individual users and the geographic behavior infor-mation for clusters of similar users. We mine a kind of frequent patterns, called semantic trajectory patterns (Alvares et al., 2007; Ying et al., 2010), from trajectories of individual users and adopt a prefix tree, called semantic trajectory pattern tree, to compactly represent a collection of semantic trajectory patterns. Based on individual semantic information (i.e., the semantic trajectory pat-terns and their support values), we cluster mobile users. For each cluster, the sequential pattern mining is used to extract cluster geographic information, called stay location patterns. Similarly, we also adopt a prefix tree to compactly represent a collection of stay location patterns. As mentioned earlier, this mining module consists of (1) Data Preprocessing step, (2) Semantic Mining step, and (3) Geographic Mining step.

4.1. Data preprocessing

To the data preprocessing step transforms each user’s GPS tra-jectories into stay location sequences. We argue that most activi-ties of a mobile user are usually performed at where the user stays. For example, a user may stay with a café to have a drink. Thus, we have to first capture the stay locations where a user stops for a while.

Our framework is able to deal with both the GPS trajectories and cell trajectories (Ye et al., 2009). For GPS trajectory, we follow Zheng et al.‘s work (Zheng, Zhang, & Xie, 2010) to discover stay points from users’ GPS trajectories. Then, a density-based cluster-ing algorithm is performed on these stay points to obtain stay loca-tions. For cell trajectories, we follow Ying et al.‘s approach (Ying et al., 2010), which treats a cell as a geographic location. The stay time in a cell is derived by calculating the difference between the time a user arrives in and leaves from the cell. A user-specified

time threshold is used to filter the cells with stay time shorter than the threshold. The remaining cells are further filtered by the num-ber of users passed through (i.e., a crowd threshold). Finally, the stay locations (i.e., the cells with stay time equal or greater than the time threshold and the number of visitors equal or greater than the crowd threshold) are obtained and each trajectory is trans-formed into a stay locations sequence. TakeFig. 6as an example. Trajectory1, Trajectory2 and Trajectory3 are transformed into the

sequences hStay Location1, Stay Location5, Stay Location6i, hStay

Location0, Stay Location1, Stay Location4, Stay Location3i, and hStay

Location0, Stay Location1, Stay Location2, Stay Location3i,

respectively. 4.2. Semantic mining

In this section we describe how to extract semantic trajectory patterns from a user’s stay location sequences and build semantic trajectory pattern tree based on the discovered patterns. Fig. 5

shows the flow of semantic information extraction. We can ob-serve that there are two main steps in the flow. First, we mine semantic trajectory pattern form each user’s stay location se-quence set. Then, we perform a hierarchical clustering method to cluster users, where the user’s similarity is based on MSTP-Similar-ity (Ying et al., 2010).

4.2.1. Semantic trajectory pattern mining

We follow Ying et al.’s approach (Ying et al., 2010) to mine semantic trajectory pattern from each user’s stay location se-quences. A geographic semantic information database (GSID) is used to assign semantic labels to the discovered stay locations. The GSID is a customized spatial database that stores the semantic information of landmarks that we collect via Google Map (alterna-tively, a gazetteer can be used as a general-purpose GSID for this operation.) In our GSID, we store landmarks, their geographic scopes, and the associated semantic labels. In this paper, we use some general categories of the landmarks as their semantic labels. If a stay location overlaps one or several landmarks stored in the GSID, the semantic labels of these landmarks are assigned to this stay location. TakeFig. 6as an example; we will assign the seman-tic label ‘‘School’’ to Stay Location1. Similarly, the semantic label of

the landmark ParkB is ‘‘Park’’. Since Stay Location5 overlaps the

landmark ParkBand Bank, the semantic labels ‘‘Park’’ and ‘‘Bank’’

are assigned to Stay Location5. Accordingly, the Stay Location5is Fig. 4. The TrajUtiRec framework for items recommendation.

transformed as {Bank, Park}. Since the area covered by single cell tower might be very large. It leads many landmarks are overlapped by a single stay location. In other words, we can not aware the ex-actly certain landmark which the users visit when he stay in Stay Location5. Accordingly, {Bank, Park} means that users probably

vis-it Bank or Park. On the contrary, vis-it is possible that a stay location overlaps none of landmark. For example, inFig. 6, there is no land-mark overlapped with Stay Location0. In this case, we assign the

semantic label ‘‘Unknown’’ to the stay location.

After assigning semantic labels to the stay location, a stay loca-tion sequence can be transformed into a semantic trajectory. Gen-eral speaking, the semantic trajectory is to present each stay location by its semantic label. For example, the stay location se-quence hStay Location0, Stay Location1, Stay Location4, Stay

Loca-tion3i is transformed as hUnknown, School, Park, Hospitali.

Unlike stay location sequence representing users’ specific, sematic trajectory can represent users’ general moving behavior. The gen-eral moving behaviors usually reflect users’ interests or life style. For example, users who like seeing movies might usually visit cin-emas. However, they might visit different cinemas because they live in different area. In other word, they may have the same semantic trajectory but totally different stay location sequences. TakeFig. 6as an example; we can observe that Trajectory2and

Tra-jectory3have very similar stay location sequences, hStay Location0,

Stay Location1, Stay Location4, Stay Location3i and hStay Location0,

Stay Location1, Stay Location2, Stay Location3i. In other word, there

is only one different location (i.e., Stay Location4 V.S. Stay

Loca-tion2). On the contrary, Trajectory1and Trajectory3have not

simi-lar stay location sequences (i.e., only one same location). Suppose three different users provide the three trajectories. We would con-sider the two users who provide Trajectory2and Trajectory3are

similar users if we focus only on similarity among stay location se-quences. As mentioned earlier, our prediction model predicting users’ next location by consulting their similar users’ trajectory. If we focus only on stay location sequences, the prediction result would be limited in the area which the user always visits. Since the prediction result is used to support item recommendation, if the predictor always predict the location which the user always visit, the recommender might recommend the item which the user always buy. Such recommender can NOT work well in real life. On the other hand, if the semantic trajectory is used to measure sim-ilarity among users while we predict users’ next location, some locations which users visit seldom might be yield. Such location can lead the recommender suggest users some items which user seldom buy but interest.

After transforming each stay location sequence into a semantic trajectory, each user’s stay location sequences are transformed into a semantic trajectory dataset. The semantic trajectories of a user may be quite diverse since the user movements may change time to time. However, the main behaviors of a user may exhibit some patterns and thus can be discovered. For example, a user goes to her school regularly and sometimes passes by a gas station. Hence, to identify the user frequent movement behaviors, we apply the sequential pattern mining algorithm Prefix-Span (Pei et al., 2001) on each user’s semantic trajectory dataset to mine the frequent semantic trajectories. TakeFig. 6as an example. Given Trajectory1

and Trajectory2of a mobile user, her trajectory log is transformed

into the semantic trajectory dataset as shown inTable 1. Suppose that we set the minimum support of Prefix-Span algorithm as 50%, the patterns hUnknown, School, Park, Hospitali, hSchool, {Bank, Park}i and all of its subsequences are discovered as frequent patterns.

Such patterns, called semantic trajectory patterns, could pro-vide several decision rules for location prediction. For example, if a pattern hUnknown, School, Park, Hospitali is discovered from a mobile user’s semantic trajectory set, we can predict that he/she Stay Location

Sequence Set

of User 1 …

Semantic Trajectory Pattern Mining

User 1’s Semantic Trajectory Pattern Tree User 2’s Semantic Trajectory Pattern Tree User 3’s Semantic Trajectory Pattern Tree User k’s Semantic Trajectory Pattern Tree

User Clustering based on MSTP-similarity

User Clusters … Stay Location Sequence Set of User 2 Stay Location Sequence Set of User 3 Stay Location Sequence Set of User k Geographic semantic information database Stay Location Sequence Set of User 1 …

Semantic Trajectory Pattern Mining

User 1’s Semantic Trajectory Pattern Tree User 2’s Semantic Trajectory Pattern Tree User 3’s Semantic Trajectory Pattern Tree User k’s Semantic Trajectory Pattern Tree

User Clustering based on MSTP-similarity

User … Stay Location Sequence Set of User 2 Stay Location Sequence Set of User 3 Stay Location Sequence Set of User k semantic information database

Fig. 5. A work flow of semantic mining.

Stay Location₆ Stay Location₅ School Park_A Hospital Bank Park_B Restaurant_A Restaurant_B

Trajectory

Trajectory

Trajectory

Trajectory

Trajectory

Bannnkk S SSttt SS SStttttaaaaayy L Looccaatitioonn₄₄ L LL Hospita S Stt SS Sttttaaaayyy L LLooocccaaatititiooonnn₃₃₃

L LL S Stt SS Sttttaaaayy L Looccaatitioonn₅₅ Parkkk_B S Stt SS Sttttaaaayy L Looccaatitioonn₆₆ R R Reeesssttttaaaaauuuurrraanntt_BB 4 44

Trajectory

6.4. Items recommendation

Items recommendation model is more effective regularization way, the next we discuss the effect of the different benefit function in different regularization premise. We have changes in the mini-mum utility to render hit rate results, and the horizontal axis for the minimum utility, the vertical axis is the hit rate, to prove that we look at different effective threshold. As shown in theFig. 18, when the user benefits as well as businesses benefit stores based on regularization, platform-effective no matter what way and reg-ularization are worse than do the regreg-ularization. We can observe that hit ratio of all regularizations are shown in straight line after minimum utility 42. The reason is that our recommender is to rec-ommend high utility itemsets to users. It leads the recrec-ommending list would not be changed while minimum utility is set high en-ough because only the highest utility itemset is filtered from data-base. Thus, we can conclude the critical value of minimum support in this dataset is 42.

7. Conclusions and future works

In this paper, we propose a novel framework, TrajUtiRec, which not only predict users’ next locations but also recommend items for the users. To address issues caused by that the traditional rec-ommendation systems focus only on users’ behaviors of virtual world, our system involve their physical moving behavior to make the location-based recommendation. In other word, we can recom-mend the items which are sold by the retailers in the location which users’ syay in or will go to. The users’ next prediction mod-ule explores the semantic trajectories of mobile users, to predict the next location of a mobile user in order to supporting tion-based recommendation. To improve the accuracy of our loca-tion predicloca-tion, we propose a novel cluster-based localoca-tion prediction technique to predict the next location of a mobile user which is based on our previous work, SemanTraj. The core idea of the cluster-based location prediction technique is to group users according to their similarity of semantic trajectory. To address the problems caused by low benefits of traditional recommenda-tion system, we utilize high utility itemset for improving the accu-racy of our location-based recommendation system. By adopting the high utility itemset mining, we can find high utility products of each location. As the result, we can provide more beneficial rec-ommending list for mobile users. Accordingly, our recommenda-tion system can recommend the high utility itemset that is related to the location the user might visit. To our best knowledge, this is the first work that exploits both semantic trajectory and high utility itemset for location prediction and items recommenda-tion. Through a series of experiments, we validate our proposal and

show that the proposed location-based recommendation frame-work has excellent performance under various conditions.

Although our recommendation system has excellent perfor-mance under various conditions, some research issues still have not been addressed in this paper. For example, such location-based recommender must be performed on the data which consists of users’ moving behavior and transaction logs. However, the real world system might not completely collect each users’ both mov-ing behavior and transaction logs. It leads a serious cold start prob-lem. Furthermore, our system adopts an existing high utility itemset mining algorithm to discover the high utility products sold by the retailers in each location. The different retailers would sell the same product. It leads the utility table is hard to be defined. Therefore, the traditional high utility itemset mining algorithm cannot work well for our recommendation system. We leave these issues as future works and plan to design more advanced recom-mendation strategies to address these issues in location-based ser-vices. For example, we will develop a location-based recommender which can recommend high utility items for new users. Moreover, we will address utility function modeling which can automatically define utility value of each product for dealing with the problem caused multiple retailers in single location.

References

Alvares, L. O., Bogorny, V., Palma, A., Kuijpers, B., Moelans, B., & Macedo, J. A. F. (2007). Towards Semantic trajectory knowledge discovery. Technical Report, Hasselt University, Belgium, October 2007.

Bao, J., Zheng, Y., & Mokbel, M. F. (2012). Location-based and preference-aware recommendation using sparse geo-social networking data. In Proceedings of the 20th international conference on advances in geographic information systems, November 06–09 2012. Redondo Beach, California.

Bogorny, V., Kuijpers, B., & Alvares, L. O. (2009). ST-DMQL: A semantic trajectory data mining query language. International Journal of Geographical Information Science, 23(10), 1245–1276.

Eagle, N., Pentland, A., & Lazer, D. (2009). Inferring social network structure using mobile phone data. In Proceedings of the national academy of sciences (PNAS) (Vol. 106, no. 36, pp. 15274–15278).

Jeung, H., Liu, Q., Shen, H. T., & Zhou, X. (2008). A hybrid prediction model for moving objects. ICDE, 70–79.

Liu, J., Wolfson, O., & Yin, H. (2006). Extracting semantic location from outdoor positioning systems. In MDM.

Lu, E. H.-C., & Tseng, V. S. (2009). Mining cluster-based mobile sequential patterns in location-based service environments. In MDM.

Lu, E. H. C., Lee, W. C., & Tseng, V. S. (2012). A framework for personal mobile commerce pattern mining and prediction. IEEE Transactions on Knowledge and Data Engineering, 24(5), 769–782.

Lu, E. H.-C., Tseng, V. S., & Yu, P. S. (2011). Mining cluster-based temporal mobile sequential patterns in location-based service environments. IEEE TKDE, 23(6).

Monreale, A., Pinelli, F., Trasarti, R., & Giannotti, F. (2009). WhereNext: A location predictor on trajectory pattern mining. KDD, 637–646.

Morzy, M. (2006). Prediction of moving object location based on frequent trajectories. In ISCIS. LNCS (Vol. 4263, pp. 583–592). Springer.

Morzy, M. (2007). Mining frequent trajectories of moving objects for location prediction. In MLDM. LNCS (Vol. 4571, pp. 667–680). Springer.

Pei, J., Han, J., Mortazavi-Asl, B., Pinto, H., Chen, Q., Dayal, U., et al. (2001). PrefixSpan: Mining sequential patterns efficiently by prefix-projected pattern growth. In Proceedings of the 17th international conference on data engineering (ICDE) (pp. 215–224).

Taiwan Tourism Bureau, <http://eng.taiwan.net.tw/>.

Yavas, G., Katsaros, D., Ulusoy, Ö., & Manolopoulos, Y. (2005). A data mining approach for location prediction in mobile environments. Data & Knowledge Engineering, 54(2), 121–146.

Ye, Y., Zheng, Y., Chen, Y., Feng, J., & Xie, X. (2009). Mining individual life pattern based on location history. In MDM.

Ye, Y., Zheng, Y., Chen, Y., Feng, J., & Xie, X. (2009). Mining individual life pattern based on location history. In Proceedings of the international conference on mobile data management 2009 (MDM 2009) (pp. 1–10). IEEE.

Yen, S.-J., Lee, Y.-S., Lin, C.-H., & Ying, J.-C. (2006). Investigating the effect of sampling methods for imbalanced data distributions. In Proceedings of IEEE international conference on systems, man, and cybernetics (SMC’2006) (pp. 4163– 1468).

Ying, J. J.-C., Lu, E. H.-C., Lee, W.-C., & Weng, T.-C., Tseng, V. S. (2010). Mining user similarity from semantic trajectories. In Proceedings of ACM SIGSPATIAL international workshop on location based social networks (LBSN’ 10), November 2. San Jose, CA, USA.

Ying, J. J.-C., Lee, W.-C., Weng, T.-C., Tseng, V. S. (2011). Semantic trajectory mining for location prediction. In Proceedings of the 19th ACM SIGSPATIAL international conference on advances in geographic information systems (ACM GIS’ 11). Chicago, IL.

Zheng, Y., Zhang, L., Xie, X., & Ma, W.-Y. (2009a). Mining interesting locations and travel sequences from GPS trajectories. In WWW.

Zheng, Y., Zhang, L., Xie, X., & Ma, W.-Y. (2009b). Mining correlation between locations using human location history. In ACM SIGSPATIAL GIS.

Zheng, Y., Zhang, L., Ma, Z., Xie, X., & Ma, W. Y. (2011). Recommending friends and locations based on individual location history. ACM Transaction on the Web, 5, 1–44.