Urban Population Migration Pattern Mining Based on Taxi Trajectories

Transcription

1 Urban Population Migration Pattern Mining Based on Taxi Trajectories ABSTRACT Bing Zhu Tsinghua University Beijing, China Leonidas Guibas Stanford University Stanford, CA, U.S.A. Understanding urban population migration patterns is very helpful for urban operation and management, including the traffic forecasting, epidemic prevention, commercial resource allocation, emergency response and future urban planning. The large taxi fleet equipped with GPS comprises ubiquitous mobile probes in urban areas, and their trajectories reveal interesting phenomena in the city. We investigate the urban population migration pattern mining based on taxi trajectories, including 1) the hotzone identification and 2) the principal Origin- Destination traffic flow (OD pair) extraction. We show how to apply state-of-art point clustering algorithms to address these two tasks. The performance of the algorithms are evaluated on a Beijing taxi trajectory dataset generated by 8,000 taxicabs in May The results show interesting insights of the time-resolved results, which is consistent with semantic interpretations. Keywords ubiquitous computing, GPS trajectory, population migration pattern, density-based clustering 1. INTRODUCTION Taxis, as one of the most important and widely used public transportation methods, can tell lots of stories about the city. In modern cities, large taxi fleets are always equipped with GPS, turning them into ubiqui- Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 200X ACM X-XXXXX-XX-X/XX/XX...$5.00. Qixing Huang Stanford University Stanford, CA, U.S.A. huangqx@stanford.edu Lin Zhang Tsinghua University Beijing, China linzhang@tsinghua.edu.cn tous mobile probes of the urban areas. For instance, there are,00 licensed taxicabs in Beijing in 2012, which generate over 1.2 million ridden trips per day. The taxi fleet also provides a very good coverage of the city in both space and time. Driven by drivers profitseeking objective and passengers demands, impacted by the traffic pattern on road network, the taxi trajectories can provide great insights into the urban hotspot distribution, population migration pattern, traffic pattern and road network, which will serve the Location Based Service(LBS) applications, transportation system management, and urban planning. We investigate the pattern mining of urban dynamics from taxi trajectories from two aspects. One is to identify the hotzones that indicate the high level of human activities, indicated by large number of taxi pickups and drop-offs. The other is to explore the principal OD flows that reveal the resident migration pattern. Our work is motivated from the foundation work of [] by Gonzalez, et al. They found that human trajectories showed a high degree of temporal and spatial regularity. We formulate both the hotzone identification and the principal OD pair extraction as the spatial-temporal point clustering problem and apply density-based clustering techniques to solve them. The contribution of this paper lies in two aspects: Spatial-temporal hotzone identification: We modify the DBSCAN(Density-Based Spatial Clustering of Applications with Noise) algorithm to identify the time-resolved hotzones, and apply it to the picking-up and dropping-off location point clouds to identify hotzones in Beijing. OD pair clustering: We define a parametric distance metrics to measure the difference between two OD pairs. We apply the DBSCAN method to reveal the characteristic travel patterns in the city

2 Tsinghua Univ.-Summer Palace Zhongguancun Xizhimen Fuxingmen Xidan business district Tiananmen Square Sanlitun business district Olympic Park Beijing Railway Residential area Terminal 2 BCIA Figure 1: Hotzones (8:00-9:00, May 1st, 2009) during different time period of the day. In particular, the distance metric can be tuned to favor long or short trips. Our algorithms are evaluated on a large-scale GPS trip dataset comprising over 8,000 taxis in Beijing. By checking the algorithmic results with Beijing s geographicsocial background information, we found that the identified urban hotzones and principal population migration patterns align well with the semantic information specified on the map. 1.1 Related work There is a growing interest in utilizing taxi trip data to analyze and understand traffic patterns and urban environment. We refer to [10] [9] [8] [] [] [13] [11] [1] [12] for some recent advances in this domain. In particular, Yue et al. [11] propose to extract attractive areasfrom the taxi trips, which is similar to our hotzone clustering task. The goal of our work is to analyze the agreement of the resulting clusters and the semantics of the urban environment. In addition, we find that the DBSCAN clustering method leads to better clusters than the single-linkage method used in their paper. The OD pair cluster extraction problem can also be addressed using the machinery of well-separated pair decomposition in the computational geometry community (See Chapter 3 of [5]). In our setting, we propose to formulate it as solving a clustering problem due to the noise presented in the input data. 2. MINING URBAN POPULATION MIGRA- TION PATTERN In this section, we describe the technical aspect of our methods for mining population migration patterns using taxi trips. Suppose we are given N trips T = {t 1,,t N }. Each trip t i is represented as an originaldestination pair (OD pair) O i D i, where O i and D i denote the original (pick-up) and the destination (dropoff) locations, respectively. Each location O i (or D i ) is given by three coordinates, i.e., normalized longitude Oi x (or Dx i ), latitude Oy i (ordy i ) and time sample Ot i (or Di t ). To make the notations uncluttered, we will always use to denote the 2D Euclidean distance between two locations, e.g., O i D i = ((Oi x Dx i )2 +(O y i D y i )2 ) 1 2. As mentioned above, we consider two clustering problems for mining migration patterns. In the first problem, we identify time-dependent hotzones by clustering the collection of pick-up and drop-off locations of all input trips. In the second problem, we identify the flow pattern by clustering the OD pairs. For both clustering problems, we employ the DBSCAN algorithm described by Ester et al [2], which has proven to be successful on large datasets [1, 3]. In the reminder of this section, we first present an overview of the DBSCAN algorithm. Then we show how to apply it to solve the two proposed clustering problems. 2.1 DBSCAN Clustering The DBSCAN clustering algorithm proposed by Ester et al [2] takes as input a collection of points D = {p} and a pre-defined distance measure d(, ) : D D R between pairs of points, and outputs a decomposition of D = C 1 C m N, such that points in each cluster C i are close to each other in terms of the pre-defined distance measure, and isolated points are categorized into noise set N. The DBSCAN algorithm is controlled by two parameters ǫ d and n min. ǫ d defines the neighbor of each p D: N(p) = {p d(p,p ) < ǫ d }, and n min is used to classify a point p into a border point or a core point based on the size of its neighbor set, i.e., p is a core point if and only if N(p) > n min. Given ǫ d and n min, the algorithm builds an oriented adjacency graph G over D. An edge (p,p ) G if and only if p is a core-point and p N(p). The clusters C i are simply given by the the connected components of G. Note that clusters, which contain a single point, are removed from the output, i.e, they are considered noise. We choose the DBSCAN algorithm for our clustering purposes due to its simplicity. In particular, it does not require the number of clusters as input. Moreover, it can handle clusters with arbitrary shape.

3 E DOSKD D DOSKD distances between OD pairs. In this case, long OD pairs are unlikely to be clusters. In contrast, when α becomes larger, we allow bigger absolute distances between two long OD pairs, and thus they are more preferred to be clustered. To prevent the chain phenomenon in clustering, we check the bounding boxes of the origins and destinations of each OD pair cluster Ci. If the area of one of the bounding boxes is bigger than amax, we then apply the DBSCAN algorithm to further subdivide Ci. To ensure that Ci is subdivided into multiple clusters, we reduce the distance threshold by half, i.e., ǫd = ǫd /21. This process is iterated until all clusters satisfy the area constraints. 3. EVALUATION F DOSKD 1.1 In this section, we carry out our methods on a largescale GPS trajectory dataset and evaluate their performances. Figure 2: OD pair clusters using different parameters of α (See Equation 1). 3.1 Dataset 2.2 Hotzone Extraction The trajectory dataset was generated by approximately 8,000 taxicabs in Beijing in May, We process the log file of each taxicab to extract individual trips. There are about 120k trips generated in one day. To extract hotzones, we feed the pick-up and drop-off locations of all input trips into the DBSCAN algorithm. In this case, the input point set D = {Oi, 1 i n} {Di, 1 i n}. As there is no obvious scaling factor between the time sample and the longitude(latitude). we add another threshold ǫt to determine the neighbors of each point p, i.e., N (p) = {p kp p k < ǫd, pt p t < ǫt }. We will analyze the clustering results in Section Origin-Destination traffic flow clustering To obtain clusters of original-destination traffic flows, we consider each OD pair Oi Di a point (Oix, Dix, Oiy, Diy ) in R. When applying the DBSCAN algorithm, we found that a meaningful distance measure between OD pairs is central to the meaningfulness of the resulting clusters. Motivated from the needs to discover both long and short characteristic migration patterns and by the fact that for clusters of OD pairs with similar purposes, clusters consist of longer OD pairs tend to be more diverse then those of shorter ones, we define the distance measure between two OD pairs Oi Di and Oj Dj as: 1 dα (Oi Di, Oj Dj ) = (koi Oj k2 + kdi Dj k2 ) 2 (koi Di k + koj Dj k)α (1) where α( 0) is a normalization parameter. In other words, dα (, ) is a relative distance, which is normalized by length of the OD pairs. The normalization parameter α adjusts the length of the OD pairs in the resulting clusters. When α = 0, the distance metric degenerates to the absolute Euclidean 3.2 Urban hotzone identification We applied out hotzone extraction method on the trips whose staring time is between 8:00 a.m. and 9:00 a.m. on May 1st As shown in Fig. 1, our algorithm identifies 3 clusters. Projecting the clusters onto the map, we can see that the locations of these clusters and their shapes matched perfectly to certain urban functional areas in Beijing including the Capital International Airport terminals, the railway stations, the public transportation center(xizhimen), the Central Business District () and the university town. By comparing this result with those achieved from the trip data between 11:00 and 12:00 pm, as shown in Fig. 3, in which only 19 clusters in the residential areas and transportation centers are grouped, one can show that the algorithm reveals an obviously less active status in the late night. 3.3 OD pair clustering Figure 2 shows the results of the OD pair clustering algorithm on the trips from 8:00 to 9:00 a.m., May 1st, Here we tested three different values of α = 0, 0.5 and 1. As illustrated Figure 2, the OD pair clustering algorithm is able to identify popular flows such as between the airport and the downtown area and between large residential area and the business districts. 1 The distance threshold is further reduced by half if this procedure is applied on the sub-clusters of Ci

4 5 5 West Transportation Center Dongdan East Transportation Center Jianguomen Transportation Center Wangfujing Shopping Street (a) Jianguomen Dongdan (b) Figure 3: (a)hotzones (11:00-12:00pm, May 1st, 2009), (b) OD pair clusters of the same period. Adjusting α yields different types of clusters. When α = 0, the algorithm only considers absolute distance and finds clusters that are more concentrated, e.g., from the airport to the closest subway station. In contrast, whenα = 1, thealgorithmtolerantsmoreerrorsforlong trip and outputs clusters that composed of more scattered long trips, e.g, clusters of trips from the airport to the union of large tourist zones. We then applied the OD pair clustering algorithm on trips of different time period. The following are some interesting results and their semantic interpretations, which justify the effectiveness of the OD pair clustering algorithm. As illustrated in Figure 3, and Table 1, we observe that more OD pair clusters are found during Time period N C r t 5am-am,1st May % 8am-9am,1st May % 2pm-3pm,1st May % 11pm-12pm,1st May % 2pm-3pm,25th May % Table 1: Statistics of the OD pair clusters. N: number of input trips; C : number of clusters. r t : percentage of clustered trips. the daytime than early morning and late night, because more pick-up events occur and more OD pair clusters are grouped. Second, higher ratio of OD pair are included into the clusters during the active hours, so we can infer that people s moving behaviors are more similar and converge to the principal OD pairs, especially the ones involving business districts and transportation centers, which brings great opportunities and challenges for the future city planning. Third, apart from the level of activity corresponding to the traffic volume, in different time periods through a day, specific traffic patterns vary. For example, in the early morning (Fig. (a)), the main flows are those connecting the residential areas and transportation centers, while at night, a large cluster from to the suburban residential area (the red cluster in the right side of Fig. 3(b)) emerges. This cluster doesn t show up before 11:00 p.m, when the subway Line 1 stops service, and this is consistent to the fact that subway is a preferable transportation method to taxi in Beijing and also implies the subway service should probably be extended in time. Different patterns also can be found between workdays and holidays. In contrast to the workday afternoon traffic pattern on May 25th, 2009 (Fig. (c)), people behave much more actively in the holiday afternoons such as Fig. (b) on May 1st, a national holiday. Comparing to the hotzone clustering, the high-order OD flow clustering algorithm reveals more details. Based on the data from 11:00 to 12:00pm, in Fig. 3(b) the destinations of the blue flow cluster from the airport and the red one from the Wangfujing Shopping Street are distinguished, while they both fall into the green hotzone between Dongdan and Jianguomen in Fig. 3(a). The blue destination area turn out to be some hotels while the red one is a residential area.. CONCLUSION In this paper, we investigate the urban population migration pattern based on the taxi trips. Focusing on the pick-up and drop-off events, the urban hotzone i- dentification and principal traffic flow extraction problems are solved by the spatial-temporal density-based clustering method we proposed. Specifically, the principal traffic flow analysis is formulated as a -D point clustering problem and the relative distance between t- wo OD pairs is defined, including a preference factor which can be used to fine-tune the preference to cluster length. Experiments on the Beijing taxi trajectory dataset have been conducted to evaluate the performances of the methods we proposed. The trajectory dataset was generated by 8,000 Beijing taxicabs in May The results show that, first the urban hotzones can be identified, whose locations and shapes matched satisfactorily to the functional areas in Beijing. Second,

5 Terminal 2 BCIA Subway stations [3] [] (a) 5:00-:00am, 1 May [5] Transportation Center Terminal 2 BCIA [] [] Railway Jinsong Subway station (b) 2:00-3:00pm, 1 May [8] Terminal 2 BCIA Transportation Center [9] Railway - Jinsong Subway station [10] (c) 2:00-3:00pm, 25 May Figure : (Left) Clustered trips. (Right) Summarized behavior of representative clusters. the main traffic flow clusters are successfully grouped by the -D point clustering algorithm and the clustering preference factor does have a great control over the clustering results. Third, combining the time-dependent clustering results with Beijing s geographic-social background information, some insights underlying the results are interpreted, which can justify the effectiveness of the method. 5. [11] [12] [13] REFERENCES [1] Z. Aoying and Z. Shuigeng. Approaches for scaling dbscan algorithm to large spatial database. Journal of Computer Science and Technology, 15():509 52, [2] M. Ester, H. Kriegel, and J. Sander. A density-based algorithm for discovering clusters in [1] large spatial databases with noises. In Proceedings of Second International Conference on Knowledge Discovery and Data Mining, pages , 199. M. Ester, H. Kriegel, and J. Sander. Clustering for mining in large spatial database. KI-Journal(Artificial Intelligence) Sepcial Issue on Data Mining, 12(1):18 2, M. Gonzalez, C. Hidalgo, and A. Barabasi. understanding individual human mobility patterns. nature, 53:9 82, Har-peled and Sariel. Geometric Approximation Algorithms. American Mathematical Society, Boston, MA, USA, A. Rosenfeld. Connectivity in digital pictures. Journal of the ACM (JACM), 1(1):1 10, 190. L. Wei, Y. Zheng, and W. Peng. Constructing popular routes from uncertain trajectories. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, pages , L. Wei, Y. Zheng, and L. Zhang. T-finder: A recommender system for finding passengers and vacant taxis. IEEE Transctions on knowledge and data engineering, PP(99):1, J. Yuan, Y. Zheng, and X. Xie. Discovering regions of different functions in a city using human mobility and pois. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 18 19, J. Yuan, Y. Zheng, and L. Zhang. Where to find my next passenger? In Proceedings of the 13th international conference on Ubiquitous computing, pages , Y. Yue, Y. Zhuang, and Q. Li. Mining time-dependent attractive areas and movement patterns from taxi trajectory data. In 1th International Conference on Geoinformatics, pages 1, D. Zhang, B. Guo, and Z. Yu. The emergence of social and community intelligence. Computer, ():21 28, W. Zhang, B. Zhu, and L. Zhang. Exploring urban dynamics and pervasive sensing: Correlation analysis of traffic density and air quality. In Sixth International Conference oninnovative Mobile and Internet Services in Ubiquitous Computing (IMIS), pages 9 1, Y. Zheng, Y. Liu, and J. Yuan. Urban computing with taxicabs. In Proceedings of the 13th international conference on Ubiquitous computing, pages 89 98, 2011.