JP2013182341A - Clustering device and clustering method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a clustering device which performs clustering sequentially with a small operation quantity.SOLUTION: A clustering device includes: a data storage part 40 for storing data items to which clustering is applied, and a cluster to which the data items belong, in such a manner that they are associated with each other; a centroid storage part 50 for storing the cluster and a representative value representing feature quantities of the data items belonging to the cluster, in such a manner that they are associated with each other; a data acquisition part 10 for acquiring a new data item which is subject to clustering; a data distance calculating part 31 for calculating a data distance between a feature quantity of the new data item and the representative value stored in the centroid storage part 50; a cluster determining part 32 for determining, on the basis of the data distance, a cluster to which the new data item belongs; a centroid calculating part 33 for calculating, on the basis of the feature quantity of the new data item, a representative value of the cluster to which the new data item belongs; and a data updating part 60 for updating, on the basis of the new data item, the data storage part 40 and the centroid storage part 50.

Description

  The present invention relates to a clustering apparatus and a clustering method.

  Conventionally, a method for automatically clustering data to be clustered without setting an external reference is known. Clustering methods are roughly classified into a hierarchical method represented by a tree diagram and a non-hierarchical method based on the validity of the cluster (for example, see Non-Patent Document 1). In either method, data to be clustered is prepared before clustering, and similar data is grouped.

  Patent Document 1 discloses a technique for performing clustering every time new data is added based on a criterion of minimum description length using past clustering results and probability parameters associated therewith.

Japanese Patent No. 3243669

Tadashi Okuno, Hitoshi Kume, Toshiro Haga, Tadashi Yoshizawa “Multivariate Analysis”, Nikka Giren Publishing Co., Ltd. p. 124-157

  In the technique of Patent Document 1, since clustering can be performed every time new data is added, it is not necessary to prepare all data to be clustered before clustering. However, in the technique of Patent Document 1, although it is not necessary to review all data that has been clustered in the past, it is necessary to repeat cluster integration as much as possible every time data is added, resulting in a large amount of computation. there were.

  The present invention has been made in view of the above, and an object of the present invention is to provide a clustering apparatus and a clustering method capable of performing sequential clustering each time data is added with a small amount of calculation.

  In order to solve the above-described problems and achieve the object, the present invention is a clustering device for sequentially clustering acquired data, and stores the already clustered data and the cluster to which the data belongs in association with each other. A data storage unit, a representative value storage unit that associates and stores a representative value that represents a feature value of the data belonging to the cluster, and a data acquisition unit that acquires new data to be clustered A data distance calculation unit that calculates a data distance between the feature value of the new data acquired by the data acquisition unit and the representative value stored in the representative value storage unit, and based on the data distance A cluster determination unit for determining a cluster to which the new data belongs, and a cluster to which the new data belongs based on the feature quantity of the new data. The representative value calculation unit for calculating the representative value of the star, the new data, and the cluster to which the new data belongs are associated and written to the data storage unit, and the new data is associated with the cluster to which the new data belongs. And a data updating unit for writing the representative value of the cluster to which the cluster belongs to the representative value storage unit.

  Further, the present invention is a clustering method executed by a clustering device that sequentially clusters acquired data, and the clustering device stores data that is already clustered and the cluster to which the data belongs in association with each other. A data acquisition unit that includes a storage unit, the cluster, and a representative value storage unit that stores a representative value that represents a feature value of the data belonging to the cluster in association with each other, and acquires new data to be clustered A data distance calculating step of calculating a data distance between the feature value of the new data acquired in the data acquisition step and the representative value stored in the representative value storage unit, and based on the data distance A cluster determining step for determining a cluster to which the new data belongs, and the special feature of the new data. A representative value calculating step of calculating the representative value of the cluster to which the new data belongs, and the new data and the cluster to which the new data belongs are associated and written to the data storage unit based on the amount; And a data updating step of writing the representative value of the cluster to which the new data belongs to the representative value storage unit in association with the cluster to which the new data belongs.

  According to the present invention, it is possible to perform sequential clustering each time data is added with a small amount of calculation.

FIG. 1 is a block diagram illustrating a configuration of the clustering apparatus according to the embodiment. FIG. 2 is a diagram for explaining the centroid. FIG. 3 is a diagram schematically illustrating the data configuration of the data storage unit. FIG. 4 is a diagram schematically illustrating a data configuration of the centroid storage unit. FIG. 5 is a flowchart showing clustering processing by the clustering apparatus. FIG. 6 is a diagram illustrating a cluster generation process. FIG. 7 is a diagram showing a process of deleting old data as sound data is acquired. FIG. 8A is a diagram of a clustering result for random data in the data range by the clustering apparatus according to the first embodiment. FIG. 8-2 is a diagram illustrating a clustering result of the group averaging method according to Comparative Example 1. FIG. 9A is a diagram of a clustering result by the clustering apparatus according to the second embodiment. FIG. 9-2 is a diagram illustrating a clustering result of the group averaging method according to Comparative Example 2. FIG. 10 is a diagram illustrating a data range of data input in the third embodiment. FIG. 11 is a diagram illustrating an input order of data input in the third embodiment and cluster transition. FIG. 12A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 12B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set. FIG. 13A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 13B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set. FIG. 14A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 14B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set. FIG. 15A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 15B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set. FIG. 16A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 16B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set. FIG. 17A is a diagram illustrating a clustering result when the distance threshold (D _p ) 70 is set. FIG. 17B is a diagram illustrating a clustering result when the distance threshold (D _p ) 80 is set.

  Exemplary embodiments of a clustering apparatus and a clustering method will be described below in detail with reference to the accompanying drawings.

  FIG. 1 is a block diagram illustrating a configuration of a clustering apparatus 1 according to the embodiment. The clustering apparatus 1 sequentially performs data clustering every time data is acquired. The clustering apparatus 1 includes a data acquisition unit 10, a feature parameter calculation unit 20, a sequential clustering unit 30, a data storage unit 40, a centroid storage unit 50, and a data update unit 60.

  The data acquisition unit 10 acquires data to be clustered. The feature parameter calculation unit 20 calculates a feature parameter as a feature amount of the data acquired by the data acquisition unit 10. Note that the clustering apparatus 1 of the present embodiment acquires sound data as clustering target data and calculates an LPC cepstrum coefficient as a feature parameter. However, the types of data to be clustered and the types of feature parameters are not limited to the embodiment.

  The data acquisition unit 10 acquires, for example, an acoustic signal quantized with 16 kHz and 16 bits (hereinafter referred to as sound data) as clustering target data. Further, the feature parameter calculation unit 20 calculates a 16th-order LPC cepstrum coefficient with an analysis frame length of 64 msec (1024 p) and an analysis frame interval of 16 msec (256 p), for example, thereby representing a 16-dimensional feature parameter representing the frequency structure of sound. Is calculated.

  As another example, feature parameters may be acquired from the outside together with target data to be clustered. In this case, the clustering device 1 does not have to include the feature parameter calculation unit 20.

  The sequential clustering unit 30 sequentially performs data clustering each time the data acquisition unit 10 acquires data based on the feature parameters calculated by the feature parameter calculation unit 20.

  The data storage unit 40 stores sound data clustered by the sequential clustering unit 30. The centroid storage unit 50 stores the centroid value of the cluster generated by the sequential clustering unit 30. Here, the centroid is the center of gravity of the feature parameter of the sound data belonging to each cluster. In this embodiment, a centroid is used as a representative value indicating the characteristics of the cluster. However, the representative value may be any value that indicates the characteristics of the cluster. For example, the centroid such as an average value of sound data belonging to the cluster may be used. Values other than Lloyd may be used.

  FIG. 2 is a diagram for explaining the centroid. In the present embodiment, for the sake of convenience of explanation, a case where the feature parameter of sound data is two-dimensional data having values on the x and y coordinates will be described as an example. As shown in FIG. 2, the centroid position of the position on the xy plane of the characteristic parameters of the data a1 to a5 belonging to the cluster A is the centroid of the cluster A.

  Returning to FIG. 1, when the data acquisition unit 10 acquires new sound data, the data update unit 60 updates data in the data storage unit 40 and the centroid storage unit 50 based on the new sound data. In the clustering device 1 of the present embodiment, the data acquisition unit 10 sequentially acquires sound data to be clustered, and the data update unit 60 stores the number of sound data stored in the data storage unit 40. Exceeds the preset threshold value, the oldest sound data is deleted, and new sound data is written in the data storage unit 40 instead. That is, the data update unit 60 also functions as a deletion unit that deletes sound data.

  FIG. 3 is a diagram schematically illustrating a data configuration of the data storage unit 40. As shown in FIG. 3, the data storage unit 40 uses the feature parameter calculation unit 20 for the acquisition order of the sound data acquired by the data acquisition unit 10, the data ID for identifying the sound data, the sound data, and the sound data. The calculated feature parameter and the cluster ID for identifying the cluster to which the sound data belongs are stored in association with each other. Note that the acquisition order may be a serial number given every time data is acquired, or may be an acquisition time as another example.

  FIG. 4 is a diagram schematically illustrating a data configuration of the centroid storage unit 50. As illustrated in FIG. 4, the centroid storage unit 50 stores the cluster ID of the cluster generated by the sequential clustering unit 30 and the centroid of the cluster in association with each other.

  Returning to FIG. 1, the sequential clustering unit 30 includes a data distance calculation unit 31, a cluster determination unit 32, and a centroid calculation unit 33. The data distance calculation unit 31 calculates the data distance. Here, the data distance is a distance between the feature parameter of the sound data (new data) newly acquired by the data acquisition unit 10 and the centroid of the cluster already generated by the sequential clustering unit 30. The data distance calculation unit 31 calculates the data distance based on the feature parameter calculated by the feature parameter calculation unit 20 and the cluster centroid stored in the centroid storage unit 50. When the centroids of a plurality of clusters are stored in the centroid storage unit 50, the data distance calculation unit 31 calculates the data distances with all centroids.

  The cluster determination unit 32 compares the data distance calculated by the data distance calculation unit 31 with a preset distance threshold. Then, when the data distance from the predetermined cluster is equal to or smaller than the distance threshold, the cluster determining unit 32 determines the cluster for which the data distance equal to or smaller than the distance threshold is calculated as the cluster to which the new data belongs. If the data distance is greater than the distance threshold, the cluster determination unit 32 also generates a new cluster as a cluster to which the new data belongs, and determines this as a cluster to which the new data belongs.

  Note that the distance threshold is set in the clustering apparatus 1 in advance. Although the value of the distance threshold is arbitrary, by setting the distance threshold to a large value, the number of clusters to be generated can be reduced, and data with different feature values can belong to the same cluster. On the other hand, by setting the distance threshold to be small, a relatively large number of clusters can be generated, and only data having relatively similar feature amounts can belong to the same cluster.

  The centroid calculation unit 33 calculates the centroid of the new data cluster determined by the cluster determination unit 32. The centroid calculation unit 33 also calculates the centroid of the cluster to which the deleted sound data belonged when the predetermined sound data is deleted from the data storage unit 40 by the data update unit 60. That is, the centroid calculating unit 33 determines the centroid of the cluster to which the deleted sound data belonged based on the characteristic parameters of the remaining sound data belonging to the deleted sound data cluster after the sound data is deleted. calculate.

  When the cluster to which the new data belongs is determined by the cluster determination unit 32, the data update unit 60 is determined for the new data, the data ID of the new data, the feature parameter of the new data, and the new data. The cluster ID of the cluster is written into the data storage unit 40. When the centroid is calculated by the centroid calculation unit 33, the data update unit 60 writes the centroid calculated by the centroid calculation unit 33 in the centroid storage unit 50 in association with the cluster ID.

  FIG. 5 is a flowchart showing the clustering process performed by the clustering apparatus 1. The clustering process is executed every time the data acquisition unit 10 of the clustering apparatus 1 acquires sound data. In the clustering process, first, when the data acquisition unit 10 acquires sound data (step S100), the feature parameter calculation unit 20 calculates feature parameters of the sound data (step S101). Next, the data update unit 60 refers to the data storage unit 40 and confirms whether data is stored in the data storage unit 40.

  If no sound data is stored in the data storage unit 40 and a cluster has not yet been generated (No in step S102), the data update unit 60 uses the address [j of the data array X [j] that the data update unit 60 has. = 0], the feature parameter of the new data calculated in step S101 is stored (step S103). The data update unit 60 further writes new data, a new data acquisition order, a data ID, and a feature parameter in the data storage unit 40.

  Next, the cluster determination unit 32 of the sequential clustering unit 30 newly generates a cluster to which the new data belongs (new cluster) (step S104). Correspondingly, in the data storage unit 40, the data update unit 60 writes the cluster ID of the new cluster in association with the new data.

  Next, the centroid calculation unit 33 calculates the centroid of the new cluster (step S105). Note that since the data belonging to the new cluster generated in step S104 is only new data, the feature parameter of the new data is calculated as the centroid of the cluster in step S105.

  The calculated centroid is written to the centroid storage unit 50 in association with the cluster ID of the new cluster by the data updating unit 60, and the clustering process when the first sound data is acquired is completed.

  On the other hand, in step S102, when the sound data is already stored in the data storage unit 40 and the cluster has been generated (step S102, Yes), the data update unit 60 uses the address [ j] is advanced by 1 (step S110). If the address [j] is larger than the final address of the data array X [j] (step S111, Yes), the data update unit 60 returns the address of the data array X [j] to [0]. (Step S112). In step S111, when the address [j] is equal to or lower than the final address of the data array X [j] (step S111, No), the process proceeds to step S113.

  Next, the centroid calculation unit 33 selects the address [based on the feature parameter other than the feature parameter stored in the address [j] among the feature parameters stored in each address of the data array X [j]. j] and the data updating unit 60 updates the centroid of the centroid storage unit 50 based on the calculation result (step S113).

  Here, since the feature parameter stored at the address [j] is excluded, the centroid value of the cluster to which the feature parameter at the address [j] belongs may be different from the previously calculated centroid value. There is no change in the centroid values of other clusters. Therefore, in step S113, the centroid of the cluster to which the feature parameter of the address [j] belongs may be calculated and updated.

Next, the data distance calculation unit 31 calculates the data distances between the new data and the centroids of all the already generated clusters (step S114). The data distance d _i between the new data feature parameter x _k (k = 0, 1, 2... K) and the centroid C _{i, k} of cluster [i] (i = 0, 1, 2... I) is Calculated by (Equation 1). Here, k is the dimension of the feature parameter, and I is the number of generated clusters. In the present embodiment, the feature parameter is two-dimensional.

Next, when there are a plurality of clusters, the cluster determination unit 32 determines the minimum value of the data distance of the new data calculated for each of the plurality of clusters, a preset distance threshold (D _p ), Compare When only one cluster exists, the calculated data distance is compared with the distance threshold value (D _p ). When the minimum value of the data distance is equal to or less than the distance threshold (D _p ) (step S115, Yes), the cluster determination unit 32 determines the cluster from which the minimum value of the data distance is obtained as the cluster to which the new data belongs. (Step S116).

  Next, the centroid calculation unit 33 calculates the centroid of the cluster of new data, and the data update unit 60 updates the centroid of the centroid storage unit 50 based on the calculation result (step S117). Specifically, the centroid calculating unit 33 calculates the centroid of the cluster to which the new data belongs based on the new data and the feature parameters of the sound data belonging to the cluster to which the new data belongs. The data update unit 60 then adds the centroid value associated with the cluster ID of the cluster to which the new data belongs to the centroid value calculated by the centroid calculation unit 33, that is, the cluster after adding the new data. Update to the value of the centroid.

  Next, the data update unit 60 stores the feature parameter of the new data at the address [j], and writes the new data, the acquisition order of the new data, the data ID, the feature parameter, and the cluster ID in the data storage unit 40 (step S118). ). This is the end of the process.

In step S115, if the minimum value of the data distance is larger than the distance threshold (D _p ) (step S115, No), it is determined that the new data does not belong to any already generated cluster, and step Proceeding to S104, a new cluster having only new data as belonging data is generated.

  As described above, when new data is added, the clustering apparatus 1 according to the present embodiment updates the cluster based only on the feature parameter of the new data and the centroid of the already generated cluster. . That is, the clustering apparatus 1 according to the present embodiment can perform sequential clustering with a small amount of calculation.

  In addition, when the number of data more than the number of data array is acquired, it is deleted in order from the old sound data, and in this case, it is only necessary to update the cluster for the cluster to which the deleted data belonged, For data that changes from moment to moment, it is possible to appropriately cluster the data obtained in the latest fixed period with a small amount of computation.

  Furthermore, since the maximum number of data to be clustered can be set by setting the number of data arrays, the user sets the desired number of data or the number of data corresponding to the desired period. Thus, it is possible to automatically obtain a clustering result for a desired number of data.

  FIG. 6 is a diagram illustrating a cluster generation process. It is assumed that sound data 1 to 5 are input to the clustering device 1 in numerical order. In this case, first, a cluster A having 1 data is generated for input of sound data 1. The centroid of cluster A is the value of the characteristic parameter of sound data 1.

Next, when the sound data 2 is input, the data distance between the sound data 2 and the centroid of the cluster A is calculated. It is assumed that the data distance of the sound data 2 is larger than the distance threshold (D _p ). In this case, a new cluster B to which the sound data 2 belongs is generated, and the centroid of the cluster B becomes the value of the characteristic parameter of the sound data 2.

Next, when the sound data 3 is input, the data distance between the sound data 3 and the centroid of the cluster A and the data distance between the sound data 3 and the centroid of the cluster B are calculated. Here, it is assumed that any data distance is larger than the distance threshold (D _p ). In this case, a new cluster C (C1 in the figure) to which the sound data 3 belongs is generated, and the centroid of the cluster C becomes the value of the characteristic parameter of the sound data 3.

Next, when the sound data 4 is input, a data distance between the sound data 4 and each centroid of the clusters A to C, that is, three data distances are calculated. Here, it is assumed that the data distance to the cluster C is the minimum among the calculated three data distances, and the minimum value of the data distance is a value equal to or smaller than the distance threshold (D _p ). In this case, the cluster to which the sound data 4 belongs is determined as the cluster C. Furthermore, the centroid of cluster C (C2) is updated with the belonging data of cluster C (C2 in the figure) as sound data 3 and sound data 4.

Next, when the sound data 5 is input, the data distance between the sound data 5 and the centroids of the clusters A to C is calculated. Here, it is assumed that, among the calculated three data distances, the data distance to the cluster C (C2) is the minimum and the minimum value of the data distance is a value equal to or less than the distance threshold (D _p ). In this case, the cluster to which the sound data 5 belongs is determined as the cluster C. Furthermore, the centroid of the cluster C (C3) is updated with the belonging data of the cluster C (C3) as the sound data 3, the sound data 4, and the sound data 5.

  FIG. 7 is a diagram showing a process of deleting old data as sound data is acquired. In the example of FIG. 7, 10 sound data are stored in the data array X [j], and when more data is input to the clustering apparatus 1, the old sound data is deleted in order. And It is assumed that the sound data 1 to 10 are clustered into clusters A to D by the timing of t1 in FIG.

  Then, new data 11 is input to the clustering apparatus 1 at the timing of t1. In this case, at the timing t1, the sound data 1 that is the oldest data stored in the data array X [j] is deleted. Then, the cluster to which the new data 11 belongs is determined as the cluster C at the timing t2, and the centroid of the cluster C is updated.

  Further, along with the deletion of the sound data 1, the centroid of the cluster A to which the sound data 1 belongs is updated at the timing t2. Furthermore, it is assumed that new data 12 is input at the timing t2. In this case, the sound data 2 which is the oldest data stored in the data array X [j] is deleted at the timing t2. Then, a new cluster E to which the new data 12 belongs is generated at the timing of t3. Further, with the deletion of the sound data 2, the cluster B to which the sound data 2 belongs disappears. When there is no sound data belonging to the cluster, the data updating unit 60, in the centroid storage unit 50, the centroid value associated with the cluster ID of the cluster to which no sound data belongs, or Delete both the cluster ID and the centroid value.

Example 1
Clustering of two-dimensional random number data was performed using the clustering apparatus 1 according to the embodiment. As data to be input to the clustering apparatus 1, (Expression 2)
(X, y) = (− 50 ≦ x, y ≦ 50) (Formula 2)
100 random data in the data range were used. The data distance threshold (D _p ) was set to 50. FIG. 8A shows a clustering result for random data within the data range by the clustering apparatus 1.

(Comparative Example 1)
Using the same data as in Example 1, clustering was performed using a group averaging method as a conventional method. Fig. 8-2 shows the clustering result of the group averaging method.

(Example 2)
Using the clustering device 1, in addition to 100 random data in the data range of (Expression 2), 103 data of 3 data (referred to as singular data) outside the data range of (Expression 2) are added. Clustering was performed. The data distance threshold (D _p ) was set to 50. FIG. 9A shows a clustering result by the clustering apparatus 1.

(Comparative Example 2)
Using the same data as in Example 2, clustering was performed using the group averaging method. FIG. 9-2 shows the clustering result of the group averaging method.

  As shown in FIGS. 8A and 8B, the clustering of the clustering apparatus 1 according to the present embodiment can obtain the same result as the clustering by the group averaging method. Furthermore, as illustrated in FIG. 9A, the clustering apparatus 1 can cluster the singular data into clusters different from other data in the clustering process for the data group including the singular data. Furthermore, as shown in FIGS. 9-1 and 9-2, the clustering apparatus 1 performed clustering on a data group including singular data, and a result similar to clustering by the group averaging method could be obtained.

(Example 3)
Clustering of two-dimensional data in the data range shown in FIG. 10 was performed using the clustering apparatus 1 according to the embodiment. Note that the data input to the clustering apparatus 1 is random data in different data ranges of groups A to D as shown in FIG. 10, and the number of data in each group is as shown in FIG. FIG. 11 shows the data input order, the data number and the number of data, the data range, and the transition of the cluster. In the cluster transition column, the upper row corresponding to each input order indicates the distance threshold (D _p ) 70, and the lower row indicates the cluster transition in the case of the distance threshold (D _p ) 80. The data in each data range was sequentially input to the clustering apparatus 1 in the input order shown in FIG.

12-1 to 17-2 show the clustering results after data input in the input order shown in FIG. Note that branch number 1 and branch number 2 in each figure show the clustering results when the distance threshold (D _p ) is set to 70 and 80, respectively.

After inputting 100 pieces of data in the input order 1, in the case of the distance threshold (D _p ) 70, four clusters are generated as shown in FIG. On the other hand, in the case of the distance threshold (D _p ) 80, three clusters were generated as shown in FIG. 12-2.

Subsequently, after 120 pieces of data in the input order 2 are input, in the case of the distance threshold (D _p ) 70, a cluster 5 is newly generated as shown in FIG. On the other hand, in the case of the distance threshold (D _p ) 80, a cluster 4 is newly generated as shown in FIG. 13-2.

Further, after the 120 data items in the order 3 are input, in the case of the distance threshold (D _p ) 70, as shown in FIG. 14-1, the cluster 5 disappears and the cluster 6 is newly generated. It was done. On the other hand, in the case of the distance threshold value (D _p ) 80, as shown in FIG. 14-2, the cluster 4 disappears and the cluster 5 is generated. Further, in both cases of the distance thresholds (D _p ) 70 and 80, data different from other data (single data) and predicted data are clustered into clusters 6 and 5 which are clusters different from the other data, respectively. I was able to confirm that.

Subsequently, after the 140 pieces of data in the input order 4 are input, in the case of the distance threshold (D _p ) 70, a cluster 7 is newly generated as shown in FIG. On the other hand, in the case of the distance threshold (D _p ) 80, the cluster 6 is generated as shown in FIG. Again, it was confirmed that the unique data and the predicted data were clustered into clusters 7 and 6 which are different clusters from the other data.

Subsequently, after inputting 120 data items in the order of entry 5, in the case of the distance threshold value (D _p ) 70, as shown in FIG. 8 was newly generated. On the other hand, in the case of the distance threshold (D _p ) 80, as shown in FIG. 16-2, the cluster 6 disappears and the cluster 7 is newly generated.

Subsequently, after inputting 100 pieces of data in the input order 6, in the case of the distance threshold (D _p ) 70, the cluster 8 disappears as shown in FIG. On the other hand, in the case of the distance threshold (D _p ) 80, as shown in FIG.

  As described above, the clustering apparatus 1 according to the present embodiment can perform clustering that follows changes in data feature amounts for a data group in which the feature amounts of input data change over time. It was confirmed.

Further, although different results can be obtained in the number of clusters and the cluster structure by making the distance threshold (D _p ) different, if a value within an appropriate range is set as the distance threshold (D _p ), the peculiarity It was confirmed that the data can be separated with high accuracy.

Note that the number of clusters increases when a relatively small distance threshold (D _p ) is set for data distribution, and the number of clusters when a relatively large distance threshold (D _p ) is set. Will be less. The optimum distance threshold (D _p ) depends on the feature amount of the input data. Therefore, it is desirable to predict variations in the input data feature amount and input data group, and to set an optimal distance threshold (D _p ) in advance based on these values.

  As described above, in the clustering apparatus 1 according to the present embodiment, all input data is not continuously stored, but is deleted sequentially from old data, so that the memory can be used effectively. Furthermore, the clustering apparatus 1 according to the present embodiment can perform clustering that reflects only relatively new data trends. Further, in the clustering apparatus 1 according to the present embodiment, only the newly input data and other data belonging to the cluster to which this data belongs are subject to calculation when data is input. There is no need for repetitive calculations for cluster generation. That is, in the clustering apparatus 1 according to the present embodiment, it is possible to reduce the amount of calculation at the time of new data input and improve the processing efficiency.

DESCRIPTION OF SYMBOLS 1 Clustering apparatus 10 Data acquisition part 20 Feature parameter calculation part 30 Sequential clustering part 31 Data distance calculation part 32 Cluster determination part 33 Centroid calculation part 40 Data storage part 50 Centroid storage part 60 Data update part

Claims (6)

A clustering device for sequentially clustering acquired data,
A data storage unit that stores data that has already been clustered and the cluster to which the data belongs;
A representative value storage unit that associates and stores the cluster and a representative value that represents a feature amount of the data belonging to the cluster;
A data acquisition unit for acquiring new data to be clustered;
A data distance calculation unit that calculates a data distance between the feature value of the new data acquired by the data acquisition unit and the representative value stored in the representative value storage unit;
A cluster determination unit for determining a cluster to which the new data belongs based on the data distance;
A representative value calculating unit that calculates the representative value of the cluster to which the new data belongs, based on the feature amount of the new data;
The new data and the cluster to which the new data belongs are associated and written to the data storage unit, and the representative value of the cluster to which the new data belongs is associated with the cluster to which the new data belongs and the representative value storage unit A clustering apparatus comprising: a data update unit for writing. The cluster determination unit compares the data distance calculated by the data distance calculation unit with a distance threshold, and if the data distance is larger than the distance threshold, the cluster stored in the data storage unit A new cluster other than is determined as a cluster to which the new data belongs,
The clustering apparatus according to claim 1, wherein the representative value calculation unit calculates the representative value of the new cluster based on a feature amount of the new data. The cluster determination unit compares the data distance calculated by the data distance calculation unit with a distance threshold, and calculates the data distance equal to or less than the distance threshold when the data distance is equal to or less than the distance threshold. Determining the cluster for the representative value as the cluster of the new data;
The representative value calculation unit refers to the data storage unit, and based on the feature amount of the data belonging to the cluster determined for the new data and the feature amount of the new data, the new data Calculate the representative value of the cluster to which it belongs,
The data update unit updates the representative value of the cluster determined for the new data stored in the representative value storage unit to the representative value calculated by the representative value calculation unit. The clustering apparatus according to claim 1, wherein: The data storage unit further stores an acquisition order in which the data acquisition unit acquires the data in association with the data,
When the data acquisition unit acquires the data, the total number of data stored in the data storage unit is compared with a preset data number threshold, and the total number of data is greater than the data number threshold. Is also larger, further comprising a deletion unit for deleting the data in the acquisition order earliest among the data stored in the data storage unit,
The representative value calculation unit calculates the representative value of the cluster to which the deleted data belonged based on the feature amount of the data stored in the data storage unit after deletion by the deletion unit. ,
The data update unit updates the representative value of the cluster, to which the deleted data belongs, stored in the representative value storage unit, to the representative value calculated by the representative value calculation unit. The clustering device according to any one of claims 1 to 3, wherein:   The clustering device according to any one of claims 1 to 4, wherein the representative value is a barycentric position of a feature amount of the target data belonging to the cluster. A clustering method executed by a clustering apparatus that sequentially clusters acquired data,
The clustering device includes a data storage unit that stores data that has already been clustered and a cluster to which the data belongs,
A representative value storage unit that stores the cluster and a representative value that represents the feature amount of the data belonging to the cluster in association with each other;
A data acquisition process for acquiring new data to be clustered;
A data distance calculation step of calculating a data distance between the feature value of the new data acquired in the data acquisition step and the representative value stored in the representative value storage unit;
A cluster determining step for determining a cluster to which the new data belongs based on the data distance;
A representative value calculating step of calculating the representative value of the cluster to which the new data belongs, based on the feature quantity of the new data;
The new data and the cluster to which the new data belongs are associated and written to the data storage unit, and the representative value of the cluster to which the new data belongs is associated with the cluster to which the new data belongs and the representative value storage unit And a data updating step for writing.

Images