712 ieee transactions on knowledge and data …home.ustc.edu.cn/~yingzi/yingzi/func_zone.pdf · ple...

Discovering Urban Functional ZonesUsing Latent Activity TrajectoriesNicholas Jing Yuan,Member, IEEE, Yu Zheng, Senior Member, IEEE,

Xing Xie, Senior Member, IEEE, Yingzi Wang, Kai Zheng, and Hui Xiong, Senior Member, IEEE

Abstract—The step of urbanization and modern civilization fosters different functional zones in a city, such as residential areas,

business districts, and educational areas. In a metropolis, people commute between these functional zones every day to engage in

different socioeconomic activities, e.g., working, shopping, and entertaining. In this paper, we propose a data-driven framework to

discover functional zones in a city. Specifically, we introduce the concept of latent activity trajectory (LAT), which captures

socioeconomic activities conducted by citizens at different locations in a chronological order. Later, we segment an urban area into

disjointed regions according to major roads, such as highways and urban expressways. We have developed a topic-modeling-based

approach to cluster the segmented regions into functional zones leveraging mobility and location semantics mined from LAT.

Furthermore, we identify the intensity of each functional zone using Kernel Density Estimation. Extensive experiments are conducted

with several urban scale datasets to show that the proposed framework offers a powerful ability to capture city dynamics and provides

valuable calibrations to urban planners in terms of functional zones.

Index Terms—Functional zones, latent activity trajectories, human mobility, points of interest

Ç

1 INTRODUCTION

MODERN cities develop with the gestation, formationand maturity of different functional zones. These func-

tional zones provide people with various urban functions tomeet their different needs of socioeconomic activities (herein-after interchangeably referred to as “activities”), e.g., WallStreet is a well-known financial district in New York City,and Silicon Valley is a high-technology business region ofthe San Francisco Bay Area. These functional zones caneither be artificially designed by urban planners (termed aszoning [2]), or naturally formulated according to people’sactual lifestyles. Meanwhile, both the territories and func-tions of these zones can be reshaped during the evolution ofa city. Discovering functional zones is crucial for uncover-ing the physical and social characters of a city, and canenable a variety of valuable applications, such as tourismrecommendation, business site selection, and calibration forurban planning.

The recent proliferation of ubiquitous sensing technolo-gies, intelligent transportation systems, and location basedservices increases the availability of human trajectories. Forexample, in big cities like New York, Munich and Beijing,most taxis are equipped with GPS devices for dispatching

and security management. These taxis regularly report theirlocations to the data center at a certain frequency. Hence, alarge number of taxi trajectories are cumulated every day.Another good example is smart cards and integrated ticket-ing, which are provided by public transit operators in manycities. Customers can swipe the purchased cards to check-inand check-out when using public transport like subways orbuses. Examples include London’s Oyster Card, Dublin’sLeap Card, Hong Kong’s Octopus Card, and Beijing’sBMAC Card.

In addition to revealing human mobility, these trajecto-ries imply the socioeconomic activities of people at differentlocations at different times, since the activities are actuallythe essential reason that mobilizes people to commutebetween different places. We term such a trajectory as alatent activity trajectory (LAT), where sequential locationsvisited by the users are observable while socioeconomicactivities implied by the sequence of locations are latent.For instance, a taxi trajectory can be segmented into multi-ple trips pertaining to different customers, where each cus-tomer travels from an origin to a destination for a certainactivity, e.g., going to work from home on a weekday morn-ing, or going shopping on a weekend evening.

In this paper, we aim to discover functional zones inurban areas leveraging latent activity trajectories. Typically,a city is naturally partitioned into individual regions bymajor roads, like expressways and ring roads (refer to thewhite lines in Fig. 1a). A functional zone is comprised of anumber of regions (not necessarily connected) with similarurban functions, where the function of a region is repre-sented by the distribution of socioeconomic activities. For exam-ple, Fig. 1a shows the functional zones we have identified inthe urban area of Beijing, where different colors indicate dif-ferent functional zones. Furthermore, we analyze the func-tionality intensity in different locations of a functional zone.

� N.J. Yuan, Y. Zheng, and X. Xie are with Microsoft Research Asia,Beijing, China. E-mail: {nicholas.yuan, yuzheng, xingx}@microsoft.com.

� Y. Wang is with the School of Computer Science and Technology, Univer-sity of Science and Technology of China, Hefei, Anhui, China.E-mail: [email protected].

� K. Zheng is with the ITEE,University ofQueensland,GPSBuilding, St LuciaCampus, Brisbane , Qld. 4072, Australia. E-mail: [email protected].

� H.Xiong iswith State University of New Jersey. E-mail: [email protected].

Manuscript received 3 June 2013; revised 19 June 2014; accepted 24 July 2014.Date of publication 4 Aug. 2014; date of current version 28 Jan. 2015.Recommended for acceptance by N. Mamoulis.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference the Digital Object Identifier below.Digital Object Identifier no. 10.1109/TKDE.2014.2345405

712 IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 27, NO. 3, MARCH 2015

1041-4347� 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

For instance, Figs. 1b and 1c show the functionality intensityof developed commercial/entertainment areas and residen-tial areas respectively, where the higher hills suggest ahigher intensity.

To identify the function of a region (the unit of a func-tional zone), we need to take into account two underlyingsignals from LAT, which reveal the socio-economic activi-ties of citizens, thus reflecting urban functions:

1) Mobility semantics. The activities conducted in aregion are strongly associated with the spatiotempo-ral patterns of the people who visit that region. Theknowledge that human mobility contributes toreveal the urban function of a region mainly is twofold. One is when people arrive at and leave a region,and the other is where people come from and leavefor. Intuitively, in a workday people usually leave aresidential area in the morning and return in theevening. The major time when people visit an enter-tainment area, however, is the evening of workdaysor the whole day of non-workdays. Furthermore, dif-ferent functional zones are correlated in the contextof human mobility. For instance, there is a high prob-ability that people reaching an entertainment areaare originating from a working area (on a workday)and a residential area (on non-workdays). As aresult, two zones are more likely to have similarfunctions, if people traveling to the two zones comefrom similar functional zones or leave for similarones.

2) Location semantics. The urban road network is lever-aged as a kind of location semantics for segmentingthe urban area into regions, since different regionsare geo-spatially connected with each other throughthe road network. Another form of location seman-tics, the allocation of points of interest (POI), whichare typically associated with a coordinate and a cate-gory like restaurants or shopping malls, uncoversthe potential socioeconomic activities of a region. Forexample, a region containing a number of universi-ties and schools has a high probability of being aneducational area. A region that usually contains avariety of POIs is probably serving multiple socio-economic activities instead of a single one. Someregions may serve as both business districts andentertainment areas in a city. In addition, the

information from POI data cannot differentiate thequality of different venues and reflect the interac-tions between functional zones. For instance, restau-rants are everywhere in a city, but they could denotedifferent functions. Some small restaurants werebuilt just for satisfying local residences’ daily needs,while a few famous restaurants attracting manypeople might be regarded as a feature location of anentertainment area. As a result, sometimes tworegions sharing a similar distribution of POIs couldstill have different functions.

This paper is an extension of our previous paper [1], inwhich we presented a topic-modeling-based approach fordiscovering region functions and intensity of functionalityusing POIs and human mobility. In this paper, we furtheroffer the following contributions:

� We have introduced the concept of latent activity tra-jectory, and generalized the problem of identifyingfunctional zones using both location and mobilitysemantics mined from latent activity trajectories.

� We have developed and detailed a morphologicalapproach to segment a city into individualregions and presented a collaborative-filtering-basedapproach to learn the location semantics from POIconfigurations of a region, which outperforms theTF-IDF vectors (as metadata for topic modeling)used in our previous paper based on experimentalresults.

� We performed exploratory study with extensiveexperiments using large-scale and real-world data-sets with regard to Beijing. In addition to the taxi tra-jectory data used in our previous work, we utilizedpubic transit records of 1.5 M trips from 0.3 M cardholders. The results suggest that the performance ofour model is improved by integrating heterogeneousmobility datasets and considering collaborative loca-tion semantics of different regions.

2 MAP SEGMENTATION

A road network is usually comprised of somemajor roads likehighways and ring roads, which naturally partition a city intoregions. For example, as shown in Fig. 2, the red segmentsdenote freeways and city expressways in Beijing, and bluesegments represent urban arterial roads. The three kinds ofroads are associated with a road level 0, 1, and 2 respectively(in a road network database), forming a natural segmentationof the urban area of Beijing. Intuitively, we consider each

Fig. 1. Territory and intensity of functional zones. (a) functional zonesidentified in Beijing (indicated by different colors); (b) intensity of thedeveloped commercial functional zone (labeled 5 in (a)); (c) intensity ofthe developed residential functional zone (labeled 6 in (a)).

Fig. 2. Beijing road network. red: level-0/1; blue: level-2.

YUAN ET AL.: DISCOVERING URBAN FUNCTIONAL ZONES USING LATENT ACTIVITY TRAJECTORIES 713

segmented region a basic unit carrying urban functions sincePOIs often fall inside regions and people perform socioeco-nomic activities (such as staying home and working) insideregions, which is also the root cause of humanmobility.

Typically, in a geographical information system (GIS),there are two models to represent spatial data: a vector-based model and a raster-based model. The vector-basedmodel uses geometric primitives such as points, lines andpolygons to represent spatial objects referenced by Carte-sian coordinates, while the raster-based model quantizes anarea into small discrete grid-cells. Both models have advan-tages and disadvantages depending on the specific applica-tions. For instance, on one hand, the vector- basedmethod is more powerful for precisely finding the shortest-paths, but requires intensive computation when performingtopological analysis, such as map simplification [3], whichis proven to be NP-complete [3]. On the other hand, the ras-ter-based model is more computationally efficient and suc-cinct for territorial analysis, but the accuracy is limited bythe number of cells used for discretizing road networks.

We display the vector-based road network on a plane byperforming map projection [4], which transforms the sur-face of a shpere (i.e., the Earth) into a 2D plane (we use Mer-cator projection in our implementation). Then we convertthe vector-based road network into the raster model bygridding the projected map.1 Intuitively, each pixel of theprojected map image can be regarded as a grid-cell of theraster map. Consequently, the road network is converted toa binary image, e.g., 1 stands for the road segments (termedas foreground) and 0 stands for the blank areas (termed asbackground).

This section introduces an image processing approachfor segmenting the raster-based road network into regionsthrough morphological operators.2

2.1 Dilation

In general, a morphological operator calculates the outputimage given the input binary image and a structure element,whose size and shape are pre-defined. Dilation is a basicmorphological operator. Let A be a binary image and B bethe structure element, the dilation of A by B is defined as:

A�B ¼[b2B

Ab; (1)

where Ab ¼ faþ b j a 2 Ag, i.e., the translation of A by vec-tor b. The dilation operator is commutative.

For any p in A, after the dilation, p ¼ 1 iff the intersectionbetween A and B, centred at p, is not empty.

The purpose of the dilation operation is to remove theunnecessary details for map segmentation, avoiding thesmall connected areas induced by these unnecessarydetails such as bridges and lanes. Fig. 3a plots a portion ofroad network before the dilation operator. As is shown,the small holes between the lanes and viaducts are filled,where we use a 3� 3 matrix with all values set to 1 as thestructure element B.

2.2 Thinning

As a consequence of the previous dilation operator, the roadsegments are turgidly thickened. In this step, we aim toextract the skeleton of the road segments while keeping thetopology structure (such as the Euler number) of the origi-nal binary image. The thinning operator is performed toremove certain foreground pixels from the input binaryimage. For a given pixel in the input image, whether itshould be removed depends on its neighbouring pixels. Fora given pixel x, the neighbouring 4 pixels shown in Fig. 4aare called the 4-neighbours of x. Similarly, the eight neigh-bouring pixels shown in Fig. 4b are called the 8-neighboursof x. Here, we employ the subfields-based parallel thinningalgorithm proposed in [5]. This method first divides thebinary image space into two disjointed subfields in a check-erboard pattern, then iterations are performed to removeforeground pixels. Each iteration consists of two sub-itera-tions in these two subfields:

� In the first sub-iteration, we check every pixel p inthe first subfield, delete p iff Conditions 1, 2 and 3are all satisfied.

� In the second sub-iteration, we check every pixel p inthe second subfield, delete p iff Conditions 1, 2 and 4are all satisfied.

Condition 1.XHðpÞ ¼ 1, where

XHðpÞ ¼X4i¼1

bi;

bi ¼ 1 if x2i�1 ¼ 0 and ðx2i ¼ 1 or x2iþ1 ¼ 1Þ;0 otherwise:

�

Fig. 3. Dilation operator.

Fig. 4. The 4-neighbours and 8-neighbours of x.

1. We used a 2,400� 2,400 grid to rasterize the map of Beijing withleft-top geo-coordinates (40.09, 116.17), right-bottom geo-coordinates(39.77, 116.56), which covers the main area of downtown Beijing.

2. Sample dataset and code can be downloaded at http://1drv.ms/1lhQ4xn.


Condition 2. 2 � minfn1ðpÞ; n2ðpÞg � 3, where

n1ðpÞ ¼X4k¼1

x2k�1 _ x2k;

n2ðpÞ ¼X4k¼1

x2k _ x2kþ1:

Condition 3. ðx2 _ x3 _ x8Þ ^ x1 ¼ 0

Condition 4. ðx6 _ x7 _ x4Þ ^ x5 ¼ 0:

The above conditions ensure that the connectivity of thepixels is preserved when a certain pixel is deleted. Note thatin this operation, connectivity paradox may be induced ifwe keep the same type of connectivity (4-connected or8-connected) for both the foreground and the background[6]. Since it is desired for the road segments (foreground) tohave unit width, typically, we preserve the 8-connectivity ofthe foreground (i.e., the 8-connectivity does not changebefore and after the thinning process for the road segments)and the 4-connectivity of the background [7]. Figs. 5a andFig. 5b are the results after seven iterations and until conver-gence (no pixel will be deleted any more) respectively.

2.3 Connected Component Labeling

The connected component labeling operation finds the con-nected 0 pixels (the blank area) in the binary image, afterthe thinning operation. We call the sequence y1; y2; . . . ; yn an8-path (4-path), if 8i ¼ 1; 2; . . . ; n� 1, yiþ1 is an 8-neighbour(4-neighbour) of yi. We say a region Q in a binary image is8-connected (4-connected) iff all the pixels in Q have the samevalue and for any two pixels in Q, there exists an 8-path(4-path) connecting the two pixels. There exist many algo-rithms for connected component labeling. Here, we applythe classical two-pass algorithm introduced in [8] to thebinary image 5b, and obtain the segmented regions asshown in Fig. 6a. Fig. 6b presents the result for Beijing’sentire road network. We note that the computational com-plexity of all the morphological operations in the map seg-mentation method is linear in terms of number of pixels.

3 DISCOVERY OF ACTIVITIES IN A REGION

In this section, we infer the distribution of activities in eachregion unit using a topic-model-based method.

3.1 Preliminary

Definition 1 (Transition). A transition Tr is a quadruple con-taining the following four items: origin region (Tr:rO), leavingtime (Tr :tL), destination region (Tr:rD) and arrival time(Tr :tA). Here, Tr:rO and Tr :rD are spatial features while theothers are temporal features.

Definition 2 (Mobility Pattern). A mobility pattern M is atriple extracted from a transition. Given a transitonTr ¼ ðTr:rO;Tr:rD;Tr:tL;Tr:tAÞ, we obtain two mobilitypatterns: the leaving mobility pattern ML ¼ ðTr:rO;Tr :rD;Tr:tLÞ, and the arriving mobility patternMA ¼ ðTr :rO;Tr :rD;Tr :tAÞ.

Definition 3 (Transition Cuboids). A transition cuboid C isan R�R� T cuboid, where R is the number of regions and Tis the number of time bins. Since there exist two types of mobil-ity patterns, we define two types of transition cuboids: leavingcuboid CL and arriving cuboid CA. The cell with indexði; j; kÞ of the leaving cuboid records the number of mobilitypatterns that leave ri for rj at time tk, i.e.,

CLði; j; kÞ ¼ kfML ¼ ðx; y; zÞ jx ¼ ri; y ¼ rj; z ¼ tkgk:Similarly,

CAði; j; kÞ ¼ kfMA ¼ ðx; y; zÞ jx ¼ ri; y ¼ rj; z ¼ tkgk:

In order to derive mobility semantics (represented by thetransition cuboids defined above) from latent activity trajec-tories, we project each trajectory on the segmented regionunits, turning a trajectory into a transition (note that forboth taxi trajectories and public transit records, the transi-tions obtained pertain to a certain individual). Then, we dis-cretize time of day into time bins in each of which wedeposit the transitions and formulate mobility patterns.Here, we do not differentiate different weekdays but differthe time bins in weekdays from those in weekends. Forexample, setting 2 hours as a bin, we will have 24 bins (12for weekdays and 12 for weekends) in total. Later, two tran-sition cuboids are built using LAT.

3.2 Collaborative POI Feature Vector

To learn the location semantics, we calculate the distribu-tion of POIs for each region. A POI is recorded with a tuple(in a POI database) consisting of a POI category (as listed inTable 2), name and a geo-position (latitude, longitude). Foreach region ri; i ¼ 1; 2; . . . ; R, the number of POIs in eachPOI category can be counted. Later, we calculate the term

Fig. 5. Thinning operator. Fig. 6. Segmented regions after connected component labeling.


frequency-inverse document frequency (TF-IDF) to measure theimportance of a POI in a region. Specifically, for a givenregion ri, we formulate a POI vector, fi ¼ ðvi1; vi2; . . . ; viCÞwhere vij is the TF-IDF value of the jth POI category and C

is the number of POI categories. The TF-IDF value vij isgiven by:

vij ¼ nj

Ni� log

R

kfrijthe j�th POI category 2rigk ; (2)

where nj is the number of POIs belonging to the jth cate-gory and Ni is the number of POIs located in region ri. Theidf term is calculated by computing the quotient of the num-ber of regions R divided by the number of regions whichhave the jth POI category, and taking the logarithm of thatquotient.

However, the TF-IDF vector is still not a good representa-tion of a region’s location semantics, which mainly suffersfrom the following limitations: 1)Missing values. Theremightexist some POIs in a region, which are not in the current POIdatabase, while featuring the location semantics of thatregion. 2) Latent structure. The frequency of POI is not anintrinsic representation of the latent structure of the POI con-figuration for each region [9], thus it is not suitable for mea-suring the similarity of location semantics between regions.

Motivated by the collaborative filtering techniques in rec-ommender systems [10], we employ the singular valuedecomposition (SVD) method to obtain the latent semanticsof each region in terms of POI configuration, which inher-ently tackles the above limitations. Specifically, let

F ¼ ðf1; f2; . . . ; fRÞ

?

; be the matrix containing the TF-IDFvectors for all regions, with dimension R� C. As is shownin Fig. 7, F is a sparse matrix, since for many regions, theremay be no certain categories of POIs. We employ SVD todecompose F by

F ¼ USV

?

; (3)

where U and V are orthogonal matrices with dimensionR�R and C � C respectively, and S is a diagonal matrixwith singular values of F. Then we can approximate F with

F̂ ¼ UlSlV

?

l , where Sl is a l� l low rank matrix containingonly the largest l singular values of S, and Ul, Vl are thereduced matrices with corresponding l columns and Crows, respectively. Now, the location semantics space is

represented by the R� l matrix Ul ¼ ðx1; x2; . . . ; xRÞ

?

,where row xi ¼ ðxi1; xi2; . . . ; xilÞ is termed as the collaborativePOI feature vector for region ri. This representation can beregarded as the coordinates of each region in the locationsemantics space.

As a result, the collaborative POI feature vectorsx1; x2; . . . ; xR are incorporated as metadata in our modelintroduced below.

3.3 Topic Modeling

In text mining, probabilistic topic models have been suc-cessfully used for extracting the hidden semantic structurein large archives of documents [11]. In this model, each doc-ument of a corpus exhibits multiple topics and each word of adocument supports a certain topic. Given all the words ofeach document in a corpus as observations, a topic model istrained to infer the hidden semantic structure behind theobservations.

The problem of identifying the latent activities in a regioncan be analogized to the problem of discovering the latenttopics of a document. As shown in Table 1, we regard aregion as a document and an activity as a topic. In otherwords, a region having multiple activities is just like a docu-ment containing a variety of topics. Meanwhile, we deemthe mobility patterns (representing mobility semantics)associated with a region as words and collaborative POI fea-ture vectors (representing location semantics) as metadataof a document. Since a functional zone is characterized byits agglomeration of activities, its intraregional transportinfrastructure, mobility of people, and inputs are within itsinteraction borders [12].

Fig. 8 further details the analogy using an example. In ourmethod, given the mobility dataset, we build the arrivingand leaving cuboids respectively according to Definition 3.For a specific region ri, the mobility patterns associated withri are counted by CAð1:R; i; 1:T Þ and CLði; 1:R; 1:T Þ, whichare two “slices” extracted from the arriving cuboid and theleaving cuboid (termed as arriving matrix and leavingmatrix respectively). The right part of Fig. 8 shows a“document” we compose for region r1, where a cell (in thematrices) represents a specific mobility pattern and the num-bers in the cell denote the occurrences of the pattern. Forexample, in the right most column of the arriving matrix, thecell containing “5” means on average the mobility that wentto r1 from rj in time bin tk occurred 5 times per day.

Latent dirichlet allocation (LDA) is a generative modelthat includes hidden variables. The intuition behind thismodel is that documents are represented as random mix-tures over latent topics, where each topic is characterized bya distribution over words [13]. Let a and h be the priorparameters for the Dirichlet document-topic distributionand topic-word distribution respectively. Assume there areK topics and b is aK �M matrix whereM is the number ofwords in the vocabulary (all the words in the corpus D).Each bk is a distribution over the vocabulary. The topic pro-portions for the dth document are ud, where ud;k is the topic

Fig. 7. Computing collaborative POI feature vector.

TABLE 1Analogy from Region-Activities to Document-Topics

transition cuboids �! vocabularyregions �! documents

activities of a region �! topics of a documentmobility patterns �! words

collaborative POI feature vector �! metadata of a document


proportion for topic k in the dth document. The topic assign-ments for the dth document are zd, where zd;n is the topicassignment for the nth word in the dth document. Finally,the observed words for document d are wd, where wd;n is thenth word in document d, which is an element from the fixedvocabulary.

Using the above notations, the generative process can bedescribed as follows:

1) For each topic k, draw bk � DirðhÞ.2) Given the dth document d in corpus D, draw

ud � DirðaÞ.3) For the nth word in the dth document wd;n,

a) draw zd;n � MultðudÞ;b) draw wd;n � Multðbzd;n

Þ.Here, Dirð�Þ is the Dirichlet distribution and Multð�Þ is themultinomial distribution. The central problem of topic

modeling is to estimate the posterior distributionP ðu; z;b jw;a; hÞ, which can be accomplished by differentapproaches, such as Gibbs sampling and variational infer-ence [13].

Using the basic LDA model, region topics can be discov-ered using mobility patterns. However, as stated in Section 1,the region topics (i.e., activities) are products of both mobilitysemantics and location semantics. In order to combine theinformation from both of them, we utilize a more advancedtopic model based on LDA andDirichletMultinomial Regres-sion (DMR) [14].

Specifically, we incorporated the learned collaborativePOI feature vectors (introduced in Section 3.2) into ourmodel. The collaborative POI feature vector of ri is denotedby xi ¼ ðxi1; xi2; . . . ; xil; 1Þ where the last “1” is a default fea-ture (as shown in Fig. 8 for region r1) to account for themean value of each topic, as explained in [14]. This vector isregarded as the metadata of each region, which is an ana-logue of the observed features such as author/email/insti-tution of a document. Such information is used as a priorknowledge to generate the “topics” of a document.

The DMR-based topic model (for simplicity, DMR in therest of the paper) takes into account the influence of theobservablemetadata in a document by using a flexible frame-work, which supports arbitrary features [14]. Compared toothermodels designed for specific data such asAuthor-Topicmodel and Topic-Over-Time model (a member in the super-vised-LDA family of topic models), DMR achieves similar orimproved performance while is more computationally effi-cient and succinct in implementation [14].

As presented in Fig. 9, the generative process of the DMRmodel is:

1) For each activity k,

TABLE 2POI Category Taxonomy

code POI category code POI category

1 car service 16 banking and insurance service2 car sales 17 corporate business3 car repair 18 street furniture4 motorcycle service 19 entrance/bridge5 caf�e/tea bar 20 public utilities6 sports/stationery shop 21 Chinese restaurant7 living service 22 foreign restaurant8 sports 23 fastfood restaurant9 hospital 24 shopping mall10 hotel 25 convenience store11 scenic spot 26 electronic products store12 residence 27 supermarket13 governmental agencies and public organizations 28 furniture building materials market14 science and education 29 pub/bar15 transportation facilities 30 theaters

Fig. 8. Analogy between mobility patterns and words based on transitioncuboids. Fig. 9. DMR-based topic model.


a) draw �k � Nð0; s2IÞ;b) draw bk � DirðhÞ.

2) Given the ith region ri,a) for each activity k, let ai;k ¼ expðxTi �kÞ;b) draw ui � DirðaiÞ;c) for the nth mobility pattern in the ith regionmi;n,

i) draw zi;n � MultðuiÞ;ii) drawmi;n � Multðbzi;n

Þ.Here, N is the Gaussian distribution with s as a hyperparameter, and �k is a vector with the same length as thecollaborative POI feature vector. The nth observed mobilitypattern of region ri is denoted as mi;n. Other notations aresimilar to the previous LDA model. In our implementation,the parameters of this model are trained using Gibbs sam-pling following the method provided in [14].

Unlike the basic LDA model, here, the Dirichlet prior a isnow specified to individual regions (ai) based on theobserved collaborative POI feature vector of each region,

i.e., ai;k ¼ expðxTi �kÞ. Therefore, for different combinationsof POI category distributions, the resulting a values are dis-tinct. Thus the activity distributions extracted from the dataare induced by both the collaborative POI features andmobility patterns. As a result, by applying DMR, given themobility patterns and collaborative POI feature vectors, weobtain the activity assignment for each region and themobility pattern distribution of each activity.

4 TERRITORY IDENTIFICATION

4.1 Region Aggregation

This step aggregates similar regions in terms of activity(topic) distributions by performing a clustering algorithm.Regions from the same cluster have similar functions, anddifferent clusters represent different functions. For regionri, after parameter estimations based on the DMR model,the topic distribution is a K dimensional vector ui ¼ðui;1; ui;2; . . . ; ui;KÞ, where ui;k is the proportion of topic k forregion ri. We perform the k-means clustering method onthe K-dimensional points ui; i 2 1; 2; . . . ; R. The number ofclusters can be predefined according to the needs of anapplication or determined using the average silhouette valueas the criterion [15]. The silhouette value of a point i in thedataset, denoted by sðiÞ is in the range of ½�1; 1, where sðiÞclose to 1 means that the point is appropriately clusteredand very distant from its neighboring clusters; sðiÞ close to 0indicates that the point is not distinctly in one cluster oranother; sðiÞ close to �1 means the point is probablyassigned to the wrong cluster. The average silhouette valueof a cluster measures how tightly the data in this cluster isgrouped, and the average silhouette of the entire datasetreflects how appropriately all the data has been clustered.In practice, we perform cross validation on the dataset fordifferent k multiple times and choose an appropriate k withthe maximum overall silhouette value. Consequently, weaggregate the regions into k clusters, each of which istermed as a functional zone.

4.2 Functionality Intensity Estimation

On one hand, the functionality of a functional zone is gener-ally not uniformly distributed within the entire region. Onthe other hand, sometimes, the core functional area may

span multiple regions and may have an irregular shape, e.g.,a hot shopping street crossing several regions. In order toreveal the degree of functionality and glean the essential ter-ritory of a functional zone, we estimate the functionality inten-sity for each aggregated functional zone (a cluster of regions).

Intuitively, the number of visits implicitly reflects thepopularity of a certain functional zone. In other words, peo-ple’s mobility patterns imply the functionality intensity. Asa result, we feed the origin and destination of each mobility(represented by latitude and longitude) into a Kernel Den-sity Estimation (KDE) model to infer the functionality inten-sity in a functional zone. Note that the real place that anindividual visited may not be the destination that we canobtain from a mobility dataset. For example, the drop-offpoints of taxi trajectories may not be people’s final destina-tions like a shopping mall. However, the pick-up/drop-offpoints should not be too far from the really-visited locationsaccording to commonsense knowledge. The farther distancea location to the drop-off point, the lower probability thatpeople would visit the location.

Given n points x1; x2; . . . ; xn located in a 2D spatial space,we estimate the intensity at location s using a kernel densityestimator, defined as:

�ðsÞ ¼Xni¼1

1

nr2K

di;sb

� �; (4)

where di;s is the distance from xi to s, b is the bandwidth andKð�Þ is the kernel function whose value decays with theincreasing of di;b, such as the Gaussian function, Quarticfunction, Conic and negative exponential. The choice of thebandwidth usually determines the smoothness of the esti-mated density—a large b achieves smoother estimationwhile a small b reveals more detailed peaks and valleys. Inour case, we choose the Gaussian function as the kernelfunction, i.e.,

Kdi;sb

� �¼ 1ffiffiffiffiffiffi

2pp exp � d2i;s

2b2

!; (5)

and the bandwidth b is determined according to MISE crite-rion [16].

4.3 Region Annotation

In this step, given the results we have obtained, we try toannotate each cluster of regions with some semantic terms,which can contribute to the understanding of its real func-tions. Note that region annotation is a very challengingproblem in both traditional urban planning and documentprocessing. Essentially, the issue is the visualization of thetopic model, which is listed as a future direction of topicmodeling in the recent survey paper by Blei [11]. A compro-mised method used thus far is to utilize the most frequentwords in a discovered topic to annotate a document. But inour case, listing the frequent mobility patterns (analogue towords) is far from enough to name a functional zone.

In our method, we annotate a functional zone by consid-ering the following four aspects: 1) The POI configuration ina functional zone. We compute an average POI density vec-tor across the regions in functional zone, where the densityrj of the jth POI category in region ri is calculated by:


rj ¼Number of POIs of the jth POI category

Area of region ri ðmeasured by grid� cellsÞ: (6)

According to density value of each POI category in the cal-culated POI density vector, we rank POI categories in afunctional zone (termed as internal ranking) and rank allfunctional zones for each POI category (referred to as theexternal ranking). We will give an example in the experi-ment as shown in Table 5. 2) The most frequent mobilitypatterns of each functional zone. 3) The functionality inten-sity. We study the representative POIs located in each func-tionality kernel, e.g., a function region could be aneducational area if its kernel is full of universities andschools. 4) The human-labeled regions. People may knowthe functions of a few well-known regions, e.g., the regioncontains the Forbidden City is an area of historic interests.After clustering, the human labeled regions will help usunderstand other regions in a cluster. Refer to the experi-ments for the detailed results and analysis.

5 EXPERIMENTS

5.1 Settings

5.1.1 Datasets

We use the following datasets for the evaluation:

1) Data representing location semantics.� Points of interest. The Beijing POI dataset covers

328,669 POIs from the year 2011, where each POIis associated with the information of its latitude,longitude and the category (see Table 2 for acomplete list of categories).

� Road networks. The road network of Beijing isused to segment the urban area into regions,with statistics shown in Table 3.

2) Data representing mobility semantics.3

� Taxi trajectories. We used a GPS trajectory datasetgenerated by Beijing taxis in the year 2011, withthe statistics shown in Table 3. We only choseoccupied trips (identified by the information of ataxi meter) from the data, and accordingly

segmented the trajectories to individual transi-tions (refer to Definition 1). It is worth noticingthat there are over 30 cities in the world withover 10,000 taxicabs, and Beijing has over 67,000taxis. The taxi trips represent a significant por-tion of people’s urban mobility. According reportby the Beijing Transportation Bureau, taxi tripsoccupy over 12 percent of traffic flows on roadsurfaces[17].

� Public transit data. This dataset logs the transac-tions of public transit including buses and sub-ways in 2011. By pre-processing the transactions,we obtained a total of 1.5 M trips (after removingthe trips that have no information of origins anddestinations), which is complementary to thetaxi trips for representing urban mobility.

5.1.2 Platforms and Baselines

We implement our method on a 64-bit server with a Quad-Core 2.67 G CPU and 16 GB RAM. We train our model with10 topics for 1,000 iterations, and optimize the parametersevery 50 iterations. For k-means clustering, we incorporatethe average silhouette value to determine the k and use theaverage results based on a five-fold cross-validation. Theefficiency (on average) is presented in Table 4.

We compare our method with several baselines:

� TF-IDF-based Methods, which include twoapproaches: 1) using POI distribution as feature vec-tors and 2) using collaborative POI feature vectors(introduced in Section 3.2). A k-means clustering isemployed to cluster the regions into k functionalzones based on their POI feature vectors.

� LDA-based Topic Model, which uses only the mobilitysemantics. Similar to our analogy from regions todocuments, this method feeds the mobility patterns(the analogue to words) into an LDA model. Later,we perform a k-means clustering, similar to themethod we used when grouping all regions basedon their topic distributions learned from LDA. Theparameters such as number of iterations, number oftopics are set in accordance with the DMR-basedmethod. As the number of POI categories usuallyhas the same scale as the topics, applying the LDAmodel solely to POIs (as words) will not reduce thedimension of words.

We carried out the following studies to evaluate theeffectiveness of our framework (though it is very difficult).1) We invited 12 local people (who have been in Beijing forover six years) and asked them to label two representative

TABLE 3Statistics of Taxi and Public Transit Trips and Road Networks

Taxi #taxis 13,597#occupied trips 8,202,012#effective days 92average trip distance (km) 7.47average trip duration (min) 16.1average sampling interval (sec) 70.45

PTC #trips 1,503,101#card IDs 295,720

Road #road segments 162,246percentage of major roads 17.1%#segmented unit regions 554size of “vocabulary” (non-0 items) 3,244,901

TABLE 4Overall Efficiency

operation time(min)

map segmentation 0.325building transition cuboids 41.3learning location semantics using SVD 2.127estimating topic model (1,000 iterations) 1372region aggregation 0.124

total 1394

3. Although we used taxi trajectories and public transit data (PTC)to evaluate our framework, we note that other mobility data such asmobile phone traces can also be directly incorporated into thisframework.


regions for each kind of function (in total 24 labeledregions). We checked whether the regions having the samelabels are assigned into the same functional zone andwhether the regions with different labels are improperlyclustered into one functional zone. 2) We matched ourresults against the land use planning of Beijing.

Compared to our experiments in [1], we further con-ducted the following experiments to examine the proposedDMR-based method:

� We compared the performance when using differentTF-IDF-based feature vectors (i.e., non-collaborativeTF-IDF feature vector/collaborative POI featurevector).

� We investigated whether the public transit data canfurther improve the performance of the proposedframework by comparing the results using solely thetaxi data against using the combined data.

5.2 Results

5.2.1 Discovered Functional Zones

Fig. 10 shows the aggregated functional zones discoveredby different methods, with different colors indicating differ-ent functions. Note that in different figures, the same colormay stand for different functions. As a result, TF-IDF-basedmethods forms seven clusters (c0–c6) while LDA-basedmethods and DMR-based methods form nine functionalzones.

The TF-IDF-based methods considering only the locationsemantics perform the worst compared with otherapproaches. For example, as shown in Fig. 10a, region B is auniversity, which should be clustered with region A(another university) and region D (a high school). Mean-while, region F (the Forbidden City) is not distinguishedfrom other commercial areas like region E (Xidan). Anotherexample is the Wangjing area (C), which is an emerging res-idential area with some companies and many living

services, like apartments, shopping malls and restaurants.Unfortunately, the TF-IDF method improperly divides thisarea into many small functional zones as this method onlyconsiders POI distributions. The TF-IDF method using col-laborative feature vectors tends to smoothen the distribu-tion of POIs, thus merging a large portion of regions intoone cluster. However, it still suffers from only consideringlocation semantics, e.g., as shown in Fig. 10b, region A is theuniversity campus, while region B is a developed commer-cial area, which should not be clustered together with A,although some universities lie in region B.

Basically, the LDA-based method and DMR-basedmethod have a similar output of functional zones. However,there still exist several exemplary regions where using bothmobility and location semantics (DMR-based) outperformsusing only mobility semantics (LDA-based) obviously. Forexample, region F in Fig. 10c is a developing commercial/entertainment area in the Wangjing area. But LDA aggre-gates it with the Forbidden City (Region E in Fig. 10a)),which is a region of historical interests; Area B (China Agri-cultural University) and Area D (Tsinghua University) aretypical science and education areas where LDA fails to cor-rectly cluster them together; Area A around Sanlitun is awell-known diplomatic district of Beijing, which is mixedwith a developing commercial area C. The LDA-basedmethod only using mobility semantics overlooks the loca-tion semantics implied by the POIs, thereby drops behindthe DMR-based method (shown in Fig. 10d).

Fig. 10e presents the identified functional zones usingDMR combing both the taxi trips and public transit data.Compared with solely using the taxi data (shown inFig. 10d), some previous sparse regions, such as region D inFig. 10d (where the taxi trajectories are not sufficient to trainthe model), can now be identified.4 In addition, some mis-identified functional regions are further corrected by incor-porating diverse types of human mobility data. For exam-ple, region A in Fig. 10d is a residential area (with manyancient “hutongs” and old neighborhoods), which iswrongly clustered into the diplomatic area if we only usethe taxi data. Region B in Fig. 10d is the famous SOLANAMall (an emerging commercial district in Beijing), whichshould be in the same cluster with the new CBD area(region C in Fig. 10d. Fig. 10f shows the results of DMRwith collaborative POI feature vectors as the metadata,which further improves the performance. As shown inFig. 10e, region A, B and C are actually all residential areas,which failed to be identified until we fed the DMR with col-laborative feature vectors. The reason behind this is proba-bly that both the mobility data and POI configurations ofthese regions are biased due to the small sizes, but could besmoothed by collaborative filtering considering the seman-tic similarity with other regions.

Overall, the method combing location semantics andmobility semantics (including both taxi trajectories and pub-lic transit data) outperforms other approaches in terms ofthe accordance with the labeled functional regions.

TABLE 5Overall POI Density Vector and Ranking of Functional Zones

FD: frequency density, IR: internal ranking.

4. Note that if we solely use the public transit data as mobilitysemantics, many regions will be identified as “sparse” regions due tothe insufficiency (only 1/5 of the taxi trips on average for each region)for learning the model.


5.2.2 Annotation of Functional Zones

Table 5 shows the average POI density vector of each regioncluster (c0-c8, remember that DMR-based method generatednine clusters) and the corresponding internal and externalrankings, where the external rank is represented by thedepth of the color (1 darkest, 4 lightest). Clearly, clusters(functional zones) c0, c2, c5, c6, c7, and c8 are more matureand more developed areas as compared to other clusters,since they have more high ranked POI categories, which areannotated as follows:

Diplomatic/Embassy Areas [c0]. The most characteristic POIcategories in this functional zone are the international res-taurants, pubs/bars, theaters, caf�es and tea bars, with a sig-nificantly higher frequency density than other functionalzones. Most embassies are located in these areas, which arewell configured for the diplomatic function, e.g., they havethe second highest external rank of residential buildings,hospitals, bank and insurance services.

Science/Education/Technology Areas [c2]. This functionalzone contains the maximum number of science and educa-tion POIs (e.g., Tsinghua university and Beijing university),banks and corporate business POIs. In addition, the biggestelectronic market in China, called “ZhongguanCun”,known as the Silicon Valley in China, is located in this func-tional zone.

Developed residential areas [c6]. This functional zone isclearly a mature residential area with the most residentialbuildings, hospitals, hotels, and convenience stores. Withinthis functional zone, an adequate number of services

supports the people’s living, such as the restaurants, shop-ping malls, banking services, schools, and sports centers.

Emerging residential areas [c8]. This area is annotated asthe emerging residential area since it has a balanced POIconfiguration (similar to, but less than c6), such as livingservices, residential buildings, sports centers, hospitals andsome companies.

Old neighborhoods [c7]. The areas within this functionalzone are mostly residential building built before the year1995 or even more anciently, where the old streets areknown as “hutongs”. The POI configuration shows that thistype of zone is less developed than both c6 and c8.

Developed commercial/entertainment areas[c5]. This is a typi-cal entertainment and commercial zone containing severalmature business circles in Beijing, such as the Xidan busi-ness circle,5 Financial Street,6 and Gongzhufen businesscircle.7

Emerging commercia/entertainment areas [c1]. The POI con-figuration (the internal rank) of this cluster is similar to clus-ter c5, but in terms of the absolute quantity, c1 is less than c5.A certain number of shopping malls, restaurants and bank-ing services feature this cluster as a developing commer-cial/business/entertainment functional zone (either ofthem is possible). In the meantime, the functionality inten-sity provides another corroboration for this annotation. As

Fig. 10. Functional zones discovered by different methods ((a),(b) share the left-top label legend, and (c)-(f) share the right-bottom label legend).

5. http://en.wikipedia.org/wiki/Xidan6. http://en.wikipedia.org/wiki/Beijing_Financial_Street7. http://en.wikipedia.org/wiki/Gongzhufen


depicted in Fig. 17a, the core of this functional zone is thenew CBD of Beijing.

Figs. 11 and 12 show the arriving/leaving transitionsmatrix of c0 and c2 during weekdays and weekends respec-tively, where the x-axes are time of day (by hour) andy-axes are the functional zones that people come from andleave for. Both c0 and c2 can generally be considered theworking areas, since trends reveal for both of them that peo-ple come at the morning peak time (8-9 am) and leave in theearly evening (5-6 pm). The results also indicate thatc6; c7; c8 are residential areas since most people come to c0and c2 in the morning are originated from these zones(Figs. 11a and 11c).

Figs. 13 and 14 show transition matrices of c5 and c1 onweekdays and weekends. It’s clear that on weekdays, mostpeople reach and leave these areas after work (5 pm-6 pm),while during weekends, people (mostly from the residentialareas such as c6 and c8) come to and leave for these zonesthroughout the day, which is a typical pattern of the com-mercial area. Another signal showing the commercial func-tion of c5 is that people go to c5 more often on weekends (asshown in Figs. 12b and 14).

With regard to the other identified functional zones,since the frequency densities of POIs are much lower thanthe above functional zones, we identify their semantic func-tions with more consideration of functionality intensity andfrequent mobility patterns derived for each functional zonein addition to the POI configurations.

Historical interests/parks [c4]. If we only consider the POIconfiguration, the characteristic of this cluster does notreveal obviously. However, by considering the functionalityintensity estimated by mobility patterns, we find that theyare places of historic interests in Beijing. As shown in

Fig. 17b, famous historical sites like the Forbidden City andthe Temple of Heaven are located in these areas. In addition,some parks like the Purple Bamboo Park,8 Happy Valley,9

and Wangxinghu Park10 are also successfully clustered intothis functional zone.

Nature areas [c3]. These areas have the fewest POIs inmost POI categories. Actually, a lot of forests and moun-tains cover this cluster, e.g., the Xishan Forest Park, CenturyForest Park, and Baiwang Moutain.

Figs. 15 and 16 show that people come to c3 followingsimilar temporal patterns as c4, but the diversity and quan-tity are reasonably weaker than c4, since many POIs in c4are very famous scenic spots. In addition, people travel to c3and c4 more often on weekends, which coincides withexpected tendencies at that time.

5.2.3 Calibration for Urban Planning

The discovered functional zones provide calibration andreference for urban planning. For example, Fig. 18 presentsthe comparison between governmental land use planning(2002-2010) and the results of our method in 2011. This areaforms an emerging residential area as planned by the gov-ernment, while some small regions become developingcommercial areas, such as A, B and C after two years’development.

6 RELATED WORK

6.1 Urban Computing with Taxicabs

In recent decades, urban computing has emerged as a con-cept where every sensor, device, person, vehicle, building,and street in urban areas can be used as a component toprobe city dynamics and further enable a city-wide comput-ing for serving people and their cities. The increasing avail-ability of GPS-embedded taxicabs provides us with anunprecedented wealth to understand human mobility in acity, thereby enabling a variety of novel urban computingresearch recently. For example, [18] and [17] studied thestrategies for improving taxi drivers’ income by analyzingthe pick-up and drop-off behavior of taxicabs in differentlocations. Yuan et al. [19] aimed to find the fastest practicaldriving route to a destination according to a large numberof taxi trajectories.

Fig. 11. Weekday transitions of c0; c2.

Fig. 12. Weekend transitions of c0; c2.


8. http://bit.ly/RwzcWs9. http://bit.ly/14YDRUG10. http://bit.ly/14YDgCi


The work presented in this paper is also a step towardsurban computing, but unlike the above-mentioned research,we focus on the discovery of functional zones in a city,which we have never seen before in this research theme.

6.2 Map Segmentation

Grid-based map segmentation is extensively used in geo-spatial-related analysis. Krumm and Horvitz [21] provideda method for predicting drivers’ destination by mappingpast trips into grid-cells and learning the destination proba-bilities with respect to each cell. Powell et al. [22] proposedan approach to suggest maximum profit grid for taxi driversby constructing a spatio-temporal profitability mapwith a grid-based segmentation, where the probabilities are calculatedusing historical data. Compared with a grid-based segmen-tation, our solution, which considers high-level roads as theboundary, is more natural for studying the human mobilityon a map, as described in Section 1.

Gonzalez et al. [23] proposed a novel road network parti-tion approach based on road hierarchy. Specifically, roadnetworks are first divided into areas by high level roads,then the partition process is recursively performed for eacharea. The partition process is implemented by finding thestrongly connected components after the removal of theintersection nodes connected to high level roads as well asthe terminals of high level road segments themselves.Fig. 19b presents the results of this approach for a portion ofthe Beijing road network. However, this method does notwork in our scenario since 1) our desired region is bound byhigh level road segments and may contain several stronglyconnected components, and 2) we aim to segment the wholearea instead of just the road nodes into regions, i.e., we needa mapping from any locations represented by latitudes and

longitudes (within the bounding box of the road network)to the region IDs.

Morphology operators are widely used in geographicalinformation systems as well as image processing. Saradjianand Amini [24] employed mathematical morphology for mapsimplification from remote sensing images by extracting skel-etons from the image and converting the structure into vec-tors. Similar works are presented in [25], [26] and [27].Different from the above methods which are based on remotesensing images, we aim to segment the urban area repre-sented by vector-based model into regions, instead of simpli-fying the map or extracting structures from the map. Fig. 19cplots the result using the proposedmorphological-based algo-rithm, as compared to the grid-based method and hierarchy-basedmethodwith respect to the same area of Beijing.

6.3 Discovery of Functional Zones

Functional zones [28] have been studied in traditional fieldsof GIS and urban planning for years, as their discovery canbenefit policy making, resource allocation, and related

Fig. 17. Functionality intensity of the emerging and developed commer-cial functional zones.




Fig. 18. (a) Governmental land use planning (2002-2010), (b) discoveredfunctional zones in 2011.


research. As early as 1970, [29] provided a case study onfunctional regions within Central London using surveyeddata of taxi flows collected in 1962, which is part of the Lon-don Traffic Survey. Karlsson [12] gave a good survey onrelated works which are mainly based on clustering algo-rithms. Some algorithms classify regions in urban areabased on remote-sensing data, as thoroughly compared in[30]. Other network-based clustering algorithms (e.g., spec-tral clustering), however, employed interaction data, suchas economic transactions, communication records [31] andpeople’s movement between regions. In [31], the authorsexploited telecommunication data to partition the GreatBritain into regions. They first rasterized the map into pix-els, then built a transition matrix using telecommunicationdata with respect to each partitioned pixel, which can beregarded as the adjacent matrix of a graph. The techniqueused for partitioning the map is the spectral method basedon modularity, which is usually utilized in communitydetection. As a result, the detected boundaries coincidedwell with either the official administrative boundaries ofthese regions, or results from existing literatures.

As the capital of China, Beijing has experienced pro-found changes especially during the past two decades. Gau-batz [32] presented a historical review of urban planning inBeijing, with a focus on the period during 1979-1995, wherethey indicated that the “new urban-planning ideas, complexlanduse and transportation patterns” are blended by theevolving form. The work reported in this paper, however,focuses on contemporary Beijing and potentially enablescalibration for urban planners in the near future.

Recently, a series of work has aimed to study the geo-graphic distribution of some topic in terms of user-generatedsocial media [33], [34]. For example, [35] studied the distribu-tions of some geographical topics (like the beach, hiking, andsunsets) in the USA using geo-tagged photos acquired fromFlickr. Pozdnoukhov and Kaiser [36] explored the space-time structure of topical content from numerous geo-tweets.The social media generated in a geo-region is still used asstatic features to feature a region. Meanwhile, a few workshave reported that human mobility can describe the func-tions of regions. For instance, Qi et al. [37] observed that thegetting on/off amount of taxi passengers in a region candepict the social activity dynamics in the region.

Our work is different from the research mentioned abovein the following aspects. First, to the best of our knowledge,

our method is the first one that simultaneously considerslocation semantics (e.g., POIs) of a region and mobilitysemantics (i.e., human mobility intentions) between regionswhen identifying functional zones. Second, rather thandirectly using some clustering algorithm, we propose atopic-model-based solution which represents a region witha distribution of socio-economic activities. Moreover, itreduces data sparsity by clustering regions into functionalzones. We demonstrate the advantage of our method overjust using the clustering approach in our experiments.

7 CONCLUSION

This paper has proposed a framework for discoveringfunctional zones (e.g., educational areas, entertainmentareas, and regions of historic interests) in a city usinghuman trajectories, which imply socio-economic activitiesperformed by citizens at different times and in various pla-ces. We have evaluated this framework with large-scaledatasets including POIs, road networks, taxi trajectoriesand public transit data. According to extensive experimen-tal results, our method using both location and mobilitysemantics outperforms the baselines solely using locationor mobility semantics in terms of effectively finding func-tional zones. Meanwhile, we have found that public transitdata can be used as a complement to the taxi trips in repre-senting urban mobility, so as to achieve a better perfor-mance for discovering functional zones. In addition, bymatching the discovered functional zones against Beijingland use planning (2002-2010), we have shown exemplarycalibrated results. The proposed framework provides apowerful tool for computational urban science, and offersemerging implications for human mobility analytics andlocation-based services.

ACKNOWLEDGMENTS

This paper is an expanded version of [1], which appeared inProceedings of the 18th ACM SIGKDD International Conferenceon Knowledge Discovery and DataMining (KDD 2012).

REFERENCES

[1] J. Yuan, Y. Zheng, and X. Xie, “Discovering regions of differentfunctions in a city using human mobility and pois,” in Proc. 18thACM SIGKDD Int. Conf. Knowledge Discovery Data Mining, 2012,pp. 186–194.

Fig. 19. Different map segmentation methods.


[2] S. M. Willhelm, Urban Zoning and Land-Use Theory. Glencoe, IL,USA: Free Press, 1962.

[3] R. Estkowski, “No Steiner point subdivision simplification is NP-Complete,” in Proc. 10th Can. Conf. Comput. Geometry, 1998,pp. 76–77.

[4] J. Snyder, Map Projections–A Working Manual, USGPO, no. 1395,1987.

[5] Z. Guo and R. Hall, “Parallel thinning with two-subiteration algo-rithms,” Commun. ACM, vol. 32, no. 3, pp. 359–373, 1989.

[6] A. Rosenfeld and J. Pfaltz, “Sequential operations in digital pic-ture processing,” J. ACM, vol. 13, no. 4, pp. 471–494, 1966.

[7] L. Lam, S. Lee, and C. Suen, “Thinning methodologies-a compre-hensive survey,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 14,no. 9, pp. 869–885, Sep. 1992.

[8] L. Shapiro, and G. Stockman, Computer Vision. Englewood Cliffs,NJ, USA: Prentice-Hall, 2001.

[9] D. Billsus and M. J. Pazzani, “Learning collaborative informationfilters,” in Proc. 5th Int. Conf. Mach. Learn., 1998, vol. 98, pp. 46–54.

[10] X. Su, and T. M. Khoshgoftaar, “A survey of collaborative filteringtechniques,” Adv. Artif. Intell., vol. 2009, p. 4, 2009.

[11] D. Blei, “Introduction to probabilistic topic models,” Commun.ACM, vol. 55, pp. 77–84, 2012.

[12] C. Karlsson, “Clusters, functional regions and cluster policies,”JIBS CESIS Electron. Working Paper Ser. (84), 2007.

[13] D. Blei, A. Ng, and M. Jordan, “Latent dirichlet allocation,” J.Mach. Learn. Res., vol. 3, pp. 993–1022, 2003.

[14] D. Mimno, and A. McCallum, “Topic models conditioned on arbi-trary features with dirichlet-multinomial regression,” in Proc.Uncertain. Artif. Intell., 2008, pp. 411–418.

[15] P. Rousseeuw, “Silhouettes: A graphical aid to the interpretationand validation of cluster analysis,” J. Comput. Appl. Math., vol. 20,pp. 53–65, 1987.

[16] M. Wand, and M. Jones, Kernel Smoothing, vol. 60. Boca Raton, FL,USA: CRC Press, 1995.

[17] J. Yuan, Y. Zheng, L. Zhang, X. Xie, and G. Sun, “Where to findmynext passenger,” in Proc. 13th Int. Conf. Ubiquitous Comput., 2011,pp. 109–118.

[18] Y. Ge, H. Xiong, A. Tuzhilin, K. Xiao, M. Gruteser, and M. Paz-zani, “An energy-efficient mobile recommender system,” in Proc.16th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2010,pp. 899–908.

[19] J. Yuan, Y. Zheng, X. Xie, and G. Sun, “Driving with knowledgefrom the physical world,” in Proc. 17th ACM SIGKDD Int. Conf.Knowl. Discovery Data Mining, 2011, pp. 316–324.

[20] W. Liu, Y. Zheng, S. Chawla, J. Yuan, and X. Xing, “Discoveringspatio-temporal causal interactions in traffic data streams,” inProc. 17th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining,2011, pp. 1010–1018.

[21] J. Krumm, and E. Horvitz, “Predestination: Where do you want togo today?” Computer, vol. 40, no. 4, pp. 105–107, Apr. 2007.

[22] J. Powell, Y. Huang, F. Bastani, and M. Ji, “Towards reducing taxi-cab cruising time using spatio-temporal profitability maps,” inProc. 12th Int. Symp. Adv. Spatial Temporal Databases, 2011, pp. 242–260.

[23] H. Gonzalez, J. Han, X. Li, M. Myslinska, and J. P. Sondag,“Adaptive fastest path computation on a road network: A trafficmining approach,” in Proc. 33rd Int. Conf. Very Large Data Bases,2007, pp. 794–805.

[24] M. Saradjian, and J. Amini, “Image map simplification usingmathematical morphology,” Int. Arch. Photogrammetry RemoteSens., vol. 33, pp. 36–43, 2000.

[25] E. L�OPez-Ornelas, “High resolution images: Segmenting, extract-ing information and GIS integration,” World Acad. Sci., Eng. Tech-nol., Vol. 54, pp. 172–177, 2009.

[26] C. Martel, G. Flouzat, A. Souriau, And F. Safa, “A morphologicalmethod of geometric analysis of images: Application to the grav-ity anomalies in the Indian Ocean,” J. Geophysical Res., vol. 94,no. B2, pp. 1715–1726, 1989.

[27] M. Ansoult, P. Soille, And J. Loodts, “Mathematical morphol-ogy—A tool for automated GIS data acquisition from scanned the-matic maps,” Photogrammetric Eng. Remote Sens., vol. 56, no. 9,pp. 1263–1271, 1990.

[28] J. Antikainen, “The concept of functional urban area,” Findings onthe ESPON project 1.1.1, Informationen zur Raumentwicklung, vol. 7,pp. 447–456, 2005.

[29] J. B. Goddard, “Functional regions within the city centre: A studyby factor analysis of taxi flows in central London,” Trans. Inst. BritGeographers, pp. 161–182, 1970.

[30] R. R. Vatsavai, E. Bright, C. Varun, B. Budhendra, A. Cheriyadat,and J. Grasser, “Machine learning approaches for high-resolutionurban land cover classification: A comparative study,” in Proc 2ndInt. Conf. Comput. Geospatial Res. Appl., 2011, pp. 11:1–11:10.

[31] C. Ratti, S. Sobolevsky, F. Calabrese, C. Andris, J. Reades, M. Mar-tino, R. Claxton, and S. H. Strogatz, “Redrawing the map of GreatBritain from a network of human interactions,” Plos One, vol. 5,no. 12, p. E14248, 2010.

[32] P. Gaubatz, “Changing Beijing,” Geographical Rev., vol. 85, pp. 79–96, 1995.

[33] J. Cranshaw, R. Schwartz, J. I. Hong, and N. M. Sadeh, “The live-hoods project: Utilizing social media to understand the dynamicsof a city,” presented at the 6th Int. AAAI Conf. Weblogs SocialMedia, Dublin, Ireland, 2012.

[34] F. Kling, and A. Pozdnoukhov, “When a city tells a story: Urbantopic analysis,” in Proc. 20th Int. Conf. Adv. Geographic Inf. Syst.,2012, pp. 482–485.

[35] Z. Yin, L. Cao, J. Han, C. Zhai, and T. Huang, “Geographical topicdiscovery and comparison,” in Proc. Int. Conf. World Wide Web,2011, pp. 247–256.

[36] A. Pozdnoukhov, and C. Kaiser, “Space-time dynamics of topicsin streaming text,” in Proc. 3rd ACM SIGSPATIAL Int. WorkshopLocation-Based Social Netw., 2011, pp. 8:1–8:8.

[37] G. Qi, X. Li, S. Li, G. Pan, Z. Wang, and D. Zhang, “Measuringsocial functions of city regions from large-scale taxi behaviors,” inProc. IEEE PERCOMWorkshops, 2011, pp. 384–388.

Nicholas Jing Yuan is an associate researcher in Microsoft ResearchAsia. Currently, his research interests include spatial-temporal datamining, behavioral mining, and computational social science. He is amember of the ACM and IEEE.

Yu Zheng is a lead researcher at Microsoft Research Asia. Hisresearch interests include location-based services, spatiotemporaldata mining, ubiquitous computing, and mobile social applications. Heis a senior member of the IEEE and ACM.

Xing Xie is a senior researcher in Microsoft Research Asia, and aguest PhD advisor with the University of Science and Technology ofChina. His research interest include spatial data mining, location-based services, social networks, and ubiquitous computing. He is asenior member of the ACM and IEEE.

Yingzi Wang is currently an graduate student in the University of Sci-ence and Technology of China. Her recent research interests includespatial-temporal data mining and human mobility analytics.

Kai Zheng is currently a research fellow at the University of Queens-land. His research interests include efficient spatiotemporal queryprocessing, uncertain data management, and spatial trajectorycomputing.

Hui Xiong is currently an associate professor and the vice departmentchair at the Management Science and Information Systems Depart-ment, Rutgers University. His general area of research is data andknowledge engineering, with a focus on developing effective and effi-cient data analysis techniques for emerging data intensive applica-tions. He is a senior member of the ACM and IEEE.

" For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


712 ieee transactions on knowledge and data …home.ustc.edu.cn/~yingzi/yingzi/func_zone.pdf · ple...

Documents