JP2010186256A - Sequential clustering device, method thereof, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To sequentially perform a DBSCAN algorithm by clustering. <P>SOLUTION: Two kinds of processing are performed in cluster updating, i.e., processing for checking whether a point to be newly a cluster nucleus exists or not among a newly added point p and the adjacent points, and processing for generating the new cluster when the point to be the cluster nucleus exists and integrating existing clusters. Specifically, the aggregation of the point p to be sequentially given and the points included in the ε-neighborhood is given as an input (S22), and it is checked whether the number of the points included in the ε-neighborhood is equal to or more than N or not (S23). When the number is N or more, the point p becomes the cluster nucleus, and the process proceeds to cluster updating portion processing 1 (S24). Then, a cluster nucleus mark is given to the point p, a new cluster ID is issued, and the cluster ID is given to the point p. When the number is N or less, the point p is not the cluster nucleus, so that the process proceeds to cluster updating portion processing 2 (S25). Then, cluster nucleus mark exchange processing is performed. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a sequential clustering apparatus, a method, and a program for use in processing for determining, in real time, a place where the user is staying from user position information obtained by using, for example, GPS (Global Positioning System) About.

  A process for extracting a place where a user stayed by clustering from coordinate data (hereinafter, collectively referred to as point data in this specification) indicating the position of the user acquired at regular time intervals using GPS or the like has been proposed. . As one of such algorithms, a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used, which is effective. The characteristics of the DBSCAN algorithm include that it is strong against data including noise and does not depend on the shape of the cluster.

  However, the existing DBSCAN algorithm does not show the sequential execution method. Therefore, in the situation where the point data is obtained sequentially, when trying to always obtain the latest clustering result for the obtained point data, every time point data is newly obtained, all the points obtained so far Clustering processing must be re-executed for point data. Since this processing is computationally expensive, it is not suitable for processing where the user's place of stay is known in a situation where input data can be obtained in real time.

Martin Ester, Hans-Peter Kriegel, Jorg Sander, Xiaowei Xu, "A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise", ACM KDD'96 Brett Adams, Dinh Phung, and Svetha Venkatesh, "Extraction of Social Context and Application to Personal Multimedia Exploration", ACM MM'06

  As described above, since the existing DBSCAN algorithm does not show the sequential execution method, the latest clustering result is always obtained for the obtained data in a situation where the input data is obtained sequentially. In some cases, every time new data is obtained, the clustering process must be re-executed on all the data obtained up to now.

  The present invention has been made in view of the above circumstances, and a sequential clustering apparatus and method capable of reducing the calculation cost when sequentially executing the DBSCAN algorithm by clustering in a situation where data is sequentially obtained. And to provide a program.

  In order to achieve the above object, the sequential clustering apparatus, sequential clustering method, and sequential clustering program according to the present invention are configured as follows.

  (1) Under a situation in which new point data is sequentially obtained, each time the new point data is input, clustering processing is performed in which neighboring point data within a certain distance range is set as a cluster unit with a certain point data as a nucleus. A sequential clustering apparatus that executes a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, which accumulates a cluster structure that has been subjected to past clustering processing, and is stored when the new point data is input. The existing cluster structure is updated to the new point data, and when the new point data is input, the new point data becomes the nucleus of the cluster or the new point data A determination means for determining whether the neighboring point data becomes a cluster nucleus, and based on the result of the determination, the existing cluster structure The for new point data and its neighboring point data, the emergence of new cluster, extension of existing cluster, coupling between clusters, configured to and a determination to update updating means the occurrence of any change in.

  (2) Under a situation in which new point data is sequentially obtained, each time the new point data is input, clustering processing is performed in which neighboring point data in a certain distance range is set as a cluster unit with a certain point data as a nucleus. A sequential clustering method that executes a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, which accumulates a cluster structure that has been clustered in the past and stores the new point data when it is input. The existing cluster structure is updated to the new point data, and when the new point data is input, the new point data becomes the nucleus of the cluster or the new point data It is determined whether neighboring point data becomes a cluster nucleus, and based on the result of the determination, the new cluster structure is added to the new cluster structure. For point data and its neighboring point data, the emergence of new cluster, extension of existing cluster, coupling between clusters, a configuration in which to determine the occurrence of any change in the update.

  (3) Under a situation in which new point data is sequentially obtained, each time the new point data is input, clustering processing is performed in which neighboring point data within a certain distance range with a certain point data as a nucleus is a cluster unit. A sequential clustering program for causing a computer to execute a DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, which stores a cluster structure obtained by past clustering processing and the new point data A process of updating an existing cluster structure that has been accumulated in accordance with new point data when input, and the update process is performed when the new point data is input when the new point data is input. It is determined whether it becomes a cluster nucleus or the neighboring point data of the new point data becomes a cluster nucleus. Based on the result of the above, with respect to the existing cluster structure, with respect to the new point data and its neighboring point data, any of the appearance of a new cluster, the expansion of the existing cluster, and the connection between clusters occurs. Is determined and updated.

  That is, in the sequential clustering apparatus, method, and program according to the present invention, as a method for sequentially executing the DBSCAN algorithm by clustering, all the points obtained so far are obtained in a situation where new point data is sequentially obtained. Instead of re-executing the DBSCAN algorithm with point data as an input, a new cluster is obtained by updating the cluster result obtained in the previous calculation in accordance with the new point data. By using this method, the calculation cost for executing clustering in a situation where point data is obtained sequentially can be reduced.

  The points of the method are as follows: (1) The newly added point is included in any cluster as a possible change when new point data is added. (2) The newly added point becomes the cluster nucleus. (3) The point near the newly added point becomes the cluster nucleus, and (4) The newly added point becomes NOISE (noise). . Of these, only (2) and (3) can cause changes in the structure of existing clusters. When these changes occur, either (a) the appearance of a new cluster, (b) the expansion of an existing cluster, or (C) the joining between clusters may occur against the structure of the existing cluster. So examine the occurrence of these changes. Since both processes can be performed by examining only the vicinity of the added point, the update range can be made local, and the calculation cost can be reduced.

  ADVANTAGE OF THE INVENTION According to this invention, the sequential clustering apparatus which can reduce the calculation cost at the time of performing the DBSCAN algorithm by clustering sequentially in the situation where point data are obtained sequentially, its method, and a program can be provided. .

1 is a block diagram showing an overall configuration of a user position determination system as an embodiment of a sequential clustering apparatus according to the present invention. The figure which shows the example of the position data of the said embodiment. The figure which shows the example of the cluster data of the said embodiment. The figure which shows the example of the ID conversion table of the said embodiment. In the said embodiment, the flowchart which shows the flow of a process of the whole apparatus. 6 is a flowchart showing a flow of sequential clustering processing of FIG. The flowchart which shows the flow of the proximity position data acquisition process of FIG. The flowchart which shows the flow of the cluster update process of FIG. The flowchart which shows the flow of the cluster partial update process 1 of FIG. The flowchart which shows the flow of the cluster partial update process 1-1 of FIG. The flowchart which shows the flow of the cluster partial update process 2 of FIG. The flowchart which shows the flow of the cluster partial update process 2-1 of FIG.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing the overall configuration of a user position determination system using a sequential clustering apparatus according to the present invention. This system includes an input unit 11, a recording unit 12, a processing unit 13, and an output unit 14.

  The input unit 11 includes position information sequential input means 111 for sequentially inputting position information obtained by a GPS receiver or the like. The input means 11 takes in position data given sequentially and represented as a point having two-dimensional coordinates, and sequentially transmits it to the position data recording means 121 of the recording unit 12.

  The recording unit 12 includes position data recording means 121, clustering result recording means 122, and cluster ID conversion table recording means 123.

  The position data recording unit 121 receives and records the position data transmitted from the input unit 11 as needed. As in the example shown in FIG. 2, the position data is held in a format in which input data IDs given in the order of input to the recording unit 12 are associated with respective coordinates. Once recorded, the data is never modified.

  The clustering result recording unit 122 stores an output obtained by applying the position data recorded in the position data recording unit 121 to the sequential clustering processing unit 131 of the processing unit 13. An example of cluster data recorded by the clustering result recording means 122 is shown in FIG.

  In the cluster data shown in FIG. 3, the cluster ID indicating which cluster each point belongs to is associated with each input data ID in the first and second columns of the table. For example, the points indicated by the input data IDs 1 and 3 indicate that they belong to the same cluster (cluster ID 1). In general, a clustering algorithm assigns a cluster to which input data belongs to input data. As a concept unique to the DBSCAN algorithm, some input points do not belong to any cluster, and a special ID of NOIZE is given to such points.

  In the third column of the table of FIG. 3, a flag indicating whether or not the point is a cluster nucleus is held. Here, a certain point is a cluster nucleus is defined as “there are N or more points within a certain threshold (hereinafter referred to as ε) centered on the coordinates of the point”. . Here, a set of points within a distance ε from a certain point p is defined as “ε-near point p”. The cluster nucleus is a central concept in the DBSCAN algorithm. The DBSCAN algorithm is a method of performing clustering by using a cluster nucleus as a cluster if a point that is a cluster nucleus and its vicinity, and if another cluster nucleus is included in the vicinity, and the vicinity of the cluster nucleus as one cluster. For details, see Non-Patent Document 1.)

  The cluster ID conversion table recording unit 123 stores information that the cluster IDs represent the same cluster when integration of different clusters occurs during the sequential clustering process. Specifically, a table as shown in FIG. 4 is stored.

  The table in FIG. 4 holds the correspondence between the original cluster ID and the cluster ID representing the same cluster as the cluster ID. In this example, for example, the cluster having the cluster ID 2 is equal to the cluster having the cluster ID 3 and further equal to the cluster having the cluster ID 5. Further, among a plurality of cluster IDs issued for a single cluster, the last issued ID is used as the representative ID of the cluster.

  Assuming that the cluster IDs are issued in order of 1, 2, 3,..., The representative ID of a certain cluster is the largest of the cluster IDs issued for that cluster. In the cluster ID conversion table, conversion from an ID to an ID larger than that is always registered. As a result, it is possible to know the representative ID of the corresponding cluster by referring to the table a plurality of times with an arbitrary cluster ID as an input. In the table of FIG. 4, the representative ID of the cluster corresponding to the cluster ID 4 is ID7. However, the representative ID7 can be obtained by searching the table with ID4 to obtain ID6 and then searching the table again with ID6. .

  The processing unit 13 includes sequential clustering processing means 131. The sequential clustering processing unit 131 receives the input newly given from the position data storage unit 121 and the clustering result at a certain point given from the clustering result storage unit 122, and outputs the clustering result after the input is given. To do.

  The output unit 14 includes cluster output means 141. The cluster output unit 141 receives the clustering result from the recording unit 13 and outputs it.

  The overall process flow of the system will be described with reference to the flowchart shown in FIG.

  When the process start is instructed (step S1), the data input from the input unit 11 is waited (step S2), and until there is a new input, the process transits to step S2 according to the determination of step S3 and enters a standby state. . When a new input is obtained, the process proceeds to step S4. In step S4, a sequential clustering process is executed. The flow of this process will be described later.

  The clustering result obtained by the sequential clustering process is recorded in step S5, and the clustering result is recorded in the output unit 14 in step S6. In step S7, it is determined whether or not data is continuously input. If it is determined that data is input, the process returns to a data input standby state (step S2). If it is determined in step S7 that there is no input, the series of processing ends (step S8).

  Next, the sequential clustering process used in step S4 of the overall process flow of FIG. 5 will be described with reference to the flowchart shown in FIG.

  When the process start is instructed (step S11), newly given input data and the result of already clustering are loaded in steps S12 and S13. Here, let p be the newly input data. In step S14, the neighboring position data acquisition process is executed using the newly input data p as an argument. This proximity position data acquisition process is a process for acquiring all the points included in the ε- vicinity of the point p among the points recorded in the position data recording unit 121. Details will be described later. In step S15, NOIZE is given as the initial value of the cluster ID of p. Thereafter, the cluster update process is performed in step S16, and the series of processes is terminated (step S17).

  The proximity position data acquisition process in step S14 will be described with reference to the flowchart shown in FIG.

  First, when an instruction to start processing is given (step S101), the point given as an input in step S102 is received, and the position data stored in the recording unit 12 is read in step S103. In step S104, all the points included in the ε- vicinity of the point given as input in S102 are acquired from the position data read in S103. For the acquisition of nearby points, as in [Non-Patent Document 1], points are recorded in advance by the position data recording means 121 in the form of R-Tree, which is a data structure suitable for high-speed search of spatial data. To reduce the processing time. The point acquired in step S105 is output, and the series of processing ends (step S106).

  The cluster update process S16 will be described with reference to the flowchart shown in FIG. As described above, the cluster update is performed by checking whether there is a new cluster nucleus among the newly added point p and its neighboring points, and if there is a cluster nucleus point, the new cluster is updated. This is done by performing two processes: generation and integration between existing clusters. In the process of FIG. 8, it is checked whether or not the newly added point p becomes a cluster nucleus, and a different process is executed for each case.

  First, when a process start is instructed (step S21), a point p and a set of points included in the vicinity of ε− are given as inputs in step S22. In step S23, it is checked whether the number of points included in the vicinity of ε− is N or more. If it is N or more, since p becomes a cluster nucleus, the process proceeds to the cluster update partial process 1 in step S24. If N or less, since p is not a cluster nucleus, the process proceeds to the cluster update partial process 2 in step S25. When one of the processes at the branch destinations S24 and S25 ends, the series of processes ends (step S26).

  The cluster update partial process 1 in step S24 will be described with reference to the flowchart shown in FIG. This process is an update process when the newly input point p is a cluster nucleus. First, when an instruction to start the process is given (step S32), points p and p in the vicinity of ε − are input in step S32. In step S33, a cluster nucleus mark is given to p, a new cluster ID is issued in step S34, and the cluster ID is given to the point p. Issuing a new cluster ID means that a new cluster around the point p is formed.

  In steps S35 to S40, each point included in the ε- vicinity of p is updated in order. An update procedure of q will be described below, where q is a point included in the ε- vicinity. The contents of the update are first branched depending on whether the point q is already a cluster nucleus. That is, it is determined in step S36 whether q is a cluster nucleus. If it is a cluster nucleus, the process proceeds to step S37, and if not, the process proceeds to S39. In steps S37 and S38, the ID conversion table is used to register that the cluster having q as the cluster nucleus and the cluster having p as the cluster nucleus are equal. Specifically, the representative ID of the cluster represented by q is acquired in step S37, and the conversion from the representative ID of q to the cluster ID of p is registered in the cluster ID conversion table in step S38. Since the cluster ID of p is newly issued in this processing, it is always larger than the representative ID of q. If q is not a cluster nucleus, the process proceeds to the cluster update partial process 1-1 in step S39. When the process of step S38 or S39 is completed, it is determined in step S40 whether there is still a point in the vicinity of p. If there is a point, the process proceeds to step S35. S41).

  The cluster update partial process 1-1 in step S39 will be described with reference to the flowchart shown in FIG. In this cluster update partial process 1-1, it is examined whether or not a point that is not a cluster nucleus in advance among the points near the newly added point p is a new cluster nucleus. After that, when a new cluster nucleus is formed, it is checked whether or not there is a cluster to be bound around it. If it does not become a cluster nucleus, the same cluster ID as p is given.

  First, when a process start is instructed (step S51), in step S52, a point q that is an update target point and is not a cluster nucleus is received as an input. In step S53, all the points near ε− of the point q are obtained by the neighborhood position data obtaining process. In step S54, it is checked whether the size in the vicinity of ε− exceeds the threshold value. If the threshold value is exceeded, q becomes a new cluster nucleus, and the process proceeds to step S55. In step S55, q is registered as a cluster nucleus, and the same cluster ID as p is assigned as the cluster ID. This is because the distance between p and q is equal to or less than ε, so that the same cluster is always obtained.

  In steps S56 to S61, each point included in the ε- vicinity of q is updated. First, one point that is not in the ε- neighborhood of p is selected from the points in the ε- neighborhood of q and set as r (step S56), and it is determined whether the point r included in the ε- neighborhood is a cluster nucleus (step S57). ) If the point r is a cluster nucleus, it is registered that the clusters represented by q and r are equal (steps S58 and S59). If the point r is not a cluster nucleus, a cluster ID of p is assigned to r (step S60). Here, in step S61, it is determined whether or not a nearby point still exists (step S61). If it exists, the process proceeds to step S56, and if it does not exist, a series of processing ends (step S63). On the other hand, if q does not become a cluster nucleus in step S54, q is given to q as the same cluster as a cluster having p as a cluster nucleus (step S62), and a series of processing is performed. The process ends (step S63).

  Next, the cluster update partial process 2 in step S25 will be described with reference to the flowchart shown in FIG. This process corresponds to the process when it is determined in step S23 of the cluster update process in FIG. 8 that the added point is not a cluster nucleus. First, when an instruction to start processing is given (step S71), a set of points p and points in the vicinity of ε- is received as input in step S72. In steps S73 to S79, each point included in the ε- vicinity of p is updated. Hereinafter, q is described as one point in the vicinity of ε−. After deleting the update target mark of q in step S73, it is checked in step S74 whether q is a cluster nucleus. If it is a cluster nucleus, the process proceeds to step S75, and if not, the process proceeds to step S78.

  In steps S75 to S77, processing for changing the cluster ID of the cluster p is performed. First, in step S75, a cluster representative ID corresponding to the cluster IDs of p and q is searched. In step S76, the representative IDs are compared. If the representative ID of q is larger, the process proceeds to step S77, and the ID is given to p. If q is not a cluster nucleus, cluster update partial processing 2-1 is executed in step S78. Q is given as a processing argument. When the process of step S77 or S78 is completed, it is determined in step S79 whether there is still an update target. If there is an update target, the process proceeds to step S73. If there is no update target, the series of processes is terminated. (Step S80).

  The cluster update partial process 2-1 in step S78 will be described with reference to the flowchart shown in FIG. This process has many similarities to the cluster update partial process 1-1, but the process is increased because p is not a cluster nucleus.

  First, when an instruction to start processing is given (step S81), a point q to be updated is given as an input in step S82. In step S83, a point included in the ε- vicinity of q is acquired. In step S84, the size of the point q near ε- is checked to determine whether q is a new cluster nucleus. If it is a cluster nucleus, the process proceeds to step S85, and if not, the process proceeds to step S92.

  In step S85, a newly issued cluster ID is given to q determined to be a cluster nucleus. In steps S86 to S93, all points included in the ε- vicinity of q are updated in order. The following description will be made with r being a point in the vicinity of ε of q.

  First, steps S81 to S84 are the same as steps S51 to S54 of the process 1-1 in FIG. In step S85, a cluster nucleus mark and a newly issued cluster ID are given to q. Subsequently, in step S86, one point near ε− of q is selected, and the selected point is set as r. It is checked whether the point r is a cluster nucleus (step S87). If it is a cluster nucleus, the process proceeds to step S88, and if not, the process proceeds to step S90.

  The processing performed in steps S88 and S89 is almost the same as the processing performed in steps S58 and S59 in the cluster update partial processing 1-1. The only difference is that the ID conversion destination is changed from the ID of p to the ID of q. It is. In step S90, the same cluster ID as q is given to r which is not a cluster nucleus. When the process in step S89 or S90 is completed, it is determined in step S91 whether there is still an update target. If there is an update target, the process proceeds to step S86. If there is no update target, the series of processes is terminated. (Step S93).

  If the size of q near ε− is smaller than N in step S84, the process proceeds to step S92. In step S92, if q is not a cluster nucleus, q is given the same cluster ID as p. When the process is finished, the series of processes is finished (step S93).

  In the user position determination system that performs the above processing, by applying the present invention, the DBSCAN algorithm having characteristics suitable for extraction of a user's place of stay such as being resistant to noise is applied in a situation where data is sequentially input. It becomes possible. As an effect, it is possible to obtain information that the user has stayed at a certain place without time delay, which can lead to the realization of a context storewareness service in real time.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some configurations may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiment examples may be appropriately combined.

11 ... input unit, 111 ... position information sequential input means,
12 ... recording unit, 121 ... position data recording means, 122 ... clustering result recording means, 123 ... cluster ID conversion table recording means,
13 ... processing unit, 131 ... sequential clustering processing means,
14: output unit, 141: cluster output means.

Claims (3)

In a situation where new point data is sequentially obtained, every time the new point data is input, a DBSCAN (Density) by a clustering process that uses a certain point data as a nucleus and neighboring point data within a certain distance range as a cluster unit. -Based Spatial Clustering of Applications with Noise) algorithm,
The cluster structure that has been subjected to past clustering processing is accumulated, and when the new point data is input, the existing cluster structure that has been accumulated is updated according to the new point data,
Determining means for determining whether the new point data becomes a cluster nucleus or the neighboring point data of the new point data becomes a cluster nucleus when the new point data is input;
Based on the result of the determination, with respect to the existing cluster structure, for the new point data and its neighboring point data, any change of appearance of a new cluster, expansion of an existing cluster, or connection between clusters A sequential clustering apparatus comprising: update means for determining and updating the occurrence of In a situation where new point data is sequentially obtained, every time the new point data is input, a DBSCAN (Density) by a clustering process in which neighboring point data within a certain distance range with a certain point data as a nucleus is used as a cluster unit. -Based Spatial Clustering of Applications with Noise) algorithm,
The cluster structure that has been subjected to past clustering processing is accumulated, and when the new point data is input, the existing cluster structure that has been accumulated is updated according to the new point data,
When the new point data is input, determine whether the new point data is a cluster nucleus or whether the neighboring point data of the new point data is a cluster nucleus,
Based on the result of the determination, with respect to the existing cluster structure, for the new point data and its neighboring point data, any change of appearance of a new cluster, expansion of an existing cluster, or connection between clusters A sequential clustering method characterized by determining and updating the occurrence of an error. In a situation where new point data is sequentially obtained, every time the new point data is input, a DBSCAN (Density) by a clustering process that uses a certain point data as a nucleus and neighboring point data within a certain distance range as a cluster unit. -Based Spatial Clustering of Applications with Noise)
An accumulation process for accumulating the cluster structure subjected to past clustering process;
When the new point data is input, the existing cluster structure accumulated is updated to the new point data,
The update process includes
When the new point data is input, determine whether the new point data is a cluster nucleus or whether the neighboring point data of the new point data is a cluster nucleus,
Based on the result of the determination process, for the existing cluster structure, any of the new point data, its neighboring point data, the appearance of a new cluster, the expansion of an existing cluster, and the connection between clusters A sequential clustering program characterized by determining and updating the occurrence of a change.

Images