JP6907772B2 - Information processing equipment and programs - Google Patents

Description

The present invention relates to an information processing device and a program.

Regarding clustering, the inventor has proposed in Patent Document 1 a method for detecting a cluster structure having overlaps and hierarchies from a network based on "modular decomposition of Markov chains". In the calculation of clustering (community extraction) based on the module decomposition of Markov chains, the probability of each node in the network changes repeatedly (random walk) through the link to another link (random walk). It was calculated and based on the information when the steady state was reached, it was determined which cluster each node belonged to.

Japanese Unexamined Patent Publication No. 2013-168127

The vector data handled by the network composed of two types of nodes (hereinafter referred to as the two-part network) is not only used for expressing the characteristics of a document, but also for expressing physical quantities obtained by various measurements and test values in various diagnoses. Conceivable. In this case, for example, the temperature value may be zero or a negative value, but unlike the feature representation of the document, the magnitude of the value does not represent the weight of the link connecting the node and the data point. Therefore, when trying to represent such vector data in a two-part network, the method of obtaining the weight of the link from the magnitude of the value cannot be adopted.

As a clustering method for handling vector data containing negative values, for example, there is a non-parametric method based on the relative positional relationship between data. In this case, it is necessary to obtain the distance between data pairs in advance. As the number of data increases, the amount of calculation increases and clustering takes time.

An object of the present invention is to provide a technique for clustering vector data containing negative values with a smaller amount of calculation than a nonparametric method.

The information processing apparatus according to claim 1 of the present invention includes acquisition means for acquiring vector data in which a plurality of components are represented by vectors, first clustering means for clustering the vector data by a parametric method, and the vector data. A generation means for generating a two-part network having the data points to be represented and the feature points of each cluster obtained by the first clustering means as nodes, and a link weight connecting the node of the data point and the node of the feature point. Clustering of the nodes by determining the calculation means to be calculated and the transition probability between the nodes via the link in the two-part network according to the link weight and executing the iterative calculation of the probability process of the transition between the nodes. A second clustering means for performing the above is provided.

In the information processing apparatus according to claim 2 of the present invention, the first clustering means performs clustering by the k-means method, and the center of the cluster obtained by the first clustering means is set as the feature point.

In the information processing apparatus according to claim 3 of the present invention, the weight of the link is calculated by an activation function that increases the weight of the link as a positive value as the Euclidean distance between the feature point and the data point becomes shorter.

In the information processing apparatus according to claim 4 of the present invention, when the data point is x _n and the feature point is m _k , the Euclidean distance is calculated by the equation (1).

In the information processing apparatus according to claim 5 of the present invention, the first clustering means performs clustering by the K-medoids method, and the representative points of the clusters obtained by the first clustering means are the feature points.

In the information processing apparatus according to claim 6 of the present invention, the first clustering means performs clustering by a mixed Gaussian model, sets each of a plurality of Gaussian distributions as the feature points, and contributes the data points to the cluster. Let the degree be the link weight.

In the information processing apparatus according to claim 7 of the present invention, the mixed Gaussian model approximates the distribution of the vector data with an ellipsoid.

In the information processing apparatus according to claim 8 of the present invention, the second clustering means calculates the importance of the clusters and extracts clusters whose importance satisfies a predetermined condition.

In the information processing apparatus according to claim 9 of the present invention, the calculation means calculates the Euclidean distance between the feature point and the data point, and determines in advance in the order in which the Euclidean distance from the data point is close. Select the selected number of the feature point nodes, set the link weight between the selected feature point node and the data point as a positive value, and set the feature point nodes other than the selected feature point and the data point to each other. The link weight is set to 0.

The program according to claim 10 of the present invention includes an acquisition means for acquiring vector data in which a plurality of components are represented by vectors, a first clustering means for clustering the vector data by a parametric method, and the vector data. A data point representing the above and a generation means for generating a two-part network having the feature points of each cluster obtained by the first clustering means as nodes, and a link weight connecting the node of the data point and the node of the feature point. By determining the calculation means for calculating the above and the transition probability between the nodes via the link in the two-part network according to the link weight and executing the iterative calculation of the probability process of the transition between the nodes. This is a program for functioning as a second clustering means for performing clustering.

According to the information processing apparatus according to claim 1 of the present invention, vector data including negative values can be clustered with a reduced amount of calculation.
According to the information processing apparatus according to claim 2 of the present invention, clustering can be performed faster than in a configuration in which clustering is performed by a nonparametric method.
According to the information processing apparatus according to claim 3 of the present invention, clustering can be performed without setting the weight of the link to a negative value.
According to the information processing apparatus according to claim 4 of the present invention, even if the vector data contains a negative value, the weight of the link can be prevented from becoming a negative value.
According to the information processing apparatus according to claim 5 of the present invention, the data closest to the center of the cluster can be used as a feature point.
According to the information processing apparatus according to claim 6 of the present invention, the number of feature points can be reduced as compared with the configuration not using the mixed Gaussian model.
According to the information processing apparatus according to claim 7 of the present invention, the amount of clustering calculation can be suppressed.
According to the information processing apparatus according to claim 8 of the present invention, important clusters can be extracted.
According to the information processing apparatus according to claim 9 of the present invention, the accuracy of clustering can be improved.
According to the program according to claim 10 of the present invention, vector data including negative values can be clustered with a reduced amount of calculation.

The figure which showed the structure of the information processing apparatus which concerns on one Embodiment of this invention. The flowchart which showed the flow of the process performed by the control unit 10. The figure which showed an example of the two-part network. The flowchart which showed the flow of the process performed by the control unit 10.

[Embodiment]
FIG. 1 is a diagram showing an example of the configuration of the information processing device 1 according to the present invention. The information processing device 1 is a computer device, and includes a control unit 10, a storage unit 11, an operation unit 12, a display unit 13, and a communication unit 14.

The communication unit 14 is connected to a communication line and has a function of a communication interface for communicating with other computer devices. The display unit 13 is a display device, and displays the result of processing performed by the control unit 10. The operation unit 12 is, for example, a keyboard, a mouse, or the like for operating the information processing device 1.

The storage unit 11 includes a storage device that permanently stores data, and stores vector data representing data points. The vector data stored here is data represented by a plurality of components that are real values as in Equation 2. Each of the plurality of components represents, for example, measured values (real values) of various sensors in the image forming apparatus, and can include negative values. Each of the plurality of components includes, for example, a measured value of a temperature sensor, and the measured value may take not only a positive value but also zero or a negative value.

When the entire data consists of N data points x ₁ , ..., X _n , this is represented by the N × D design matrix shown in Equation 3.

Further, the storage unit 11 stores a program executed by the control unit 10. The program stored in the storage unit 11 is a program that performs clustering from vector data. The program stored in the storage unit 11 may be a program acquired by the communication unit 14 via a telecommunication line or a program acquired from a computer-readable recording medium.

The control unit 10 includes a CPU (Central Processing Unit) and a RAM (Random Access Memory), and executes a program stored in the storage unit 11. When the control unit 10 executes the program stored in the storage unit 11, the acquisition unit 101, the first clustering unit 102, the generation unit 103, the calculation unit 104, and the second clustering unit 105 are realized, and clustering is performed on the vector data. The function to perform is realized.

The acquisition unit 101, which is an example of the acquisition means according to the present invention, acquires vector data from the storage unit 11. The first clustering unit 102, which is an example of the first clustering means according to the present invention, clusters vector data by a parametric method. The generation unit 103, which is an example of the generation means according to the present invention, generates a two-part network in which individual data points of vector data and the average of individual clusters obtained by clustering by the first clustering unit 102 are nodes. The calculation unit 104, which is an example of the calculation means according to the present invention, calculates the weight of the link connecting the node of each data point in the two-part network and the average node of the cluster. The second clustering unit 105, which is an example of the second clustering means according to the present invention, determines the transition probability between the nodes via the link in the two-part network generated by the generation unit 103 according to the link weight, and determines the transition probability between the nodes. Clustering the nodes of the two-part network is performed by performing iterative calculations of the transition stochastic process.

FIG. 2 is a flowchart showing a flow of processing performed by the control unit 10 that executes the program. First, the control unit 10 (acquisition unit 101) acquires the vector data stored in the storage unit 11 (step SA1). Next, the control unit 10 (first clustering unit 102) clusters the acquired vector data by a predetermined method (step SA2). Here, the method for clustering vector data is a parametric clustering method, for example, a k-means method (K-means method). In the k-means method, the vector data is divided into K clusters specified by the user of the information processing apparatus 1. dividing the vector data in k-means, the center m _k of K clusters by the numerical formula 4 is obtained as the average of vector points belonging to each cluster. This center m _k is a temporary center of the cluster and is a feature point of the cluster. In the equation of Equation 4, C _k represents the cluster k (k = 1, ..., K), and N _k represents the number of elements (data points) belonging to the cluster k. By performing clustering by the k-means method and using the central _mk as the feature point of the cluster, the data distribution of the feature points is accurately locally densed as compared with the configuration in which parametric clustering is not performed. It will be selected as the center of the part.

In the process of step SA2, it is assumed that the data points in each cluster are distributed according to a specific model, and that the data points in each cluster are distributed in a spherical or elliptical shape. In the case of the parametric clustering method, it is not necessary to obtain the distance between all the data pairs as in the nonparametric method, so that clustering is performed with a smaller amount of calculation than the nonparametric method.

By performing the process of step SA2, the control unit 10 _{identifies the center mk,} which is the average feature point of each cluster, as the feature node. When the feature node is specified, the control unit 10 (generation unit 103) generates a two-part network having individual data points and individual cluster averages obtained in step SA2 as nodes (step SA3). A bipartite network is also called a bipartite graph, and is a network (graph) in which a set of nodes is divided into two subsets and there is no link between nodes in the same subset. An example of the two-part network is illustrated in FIG. In FIG. 3, the triangle represents the data point node n corresponding to the data point, and the circle represents the feature node m corresponding to the average of the clusters. A straight line connecting the data point node n and the feature node m is a link.

Next, the control unit 10 (calculation unit 104) calculates the weight w _nk of the link connecting the data point node n and the feature node m (step SA4). Here, the control unit 10 determines the _{weight w nk} through an activation function centered on _{the average m k of} the clusters shown in Equation 5, for example. Equation (1) of Equation 5 is the Euclidean distance between the _{center m k of} each cluster and each data point x _n. The equation (2) of Equation 5 is an activation function that increases the weight of the link connecting the node n of the data point and the feature node m as a positive value as the Euclidean distance is shorter. According to the equation of Equation 5, even if the component of the vector data or the component of the feature point has a negative value, the weight w _nk of the link can be set to positive or 0 so that it does not become a negative value.

Next, the control unit 10 (second clustering unit 105) performs community decomposition by the method of module decomposition of the network for the two-part network generated in step SA3 (step SA5). The module decomposition of this network is expressed by the following equation 6.

In the equation of Equation 6, p (n) is the probability that the node n has (the probability that a random walker exists at that node). Further, π _k is a prior probability of the cluster k and indicates the importance of the cluster k. The sum of π _k for k is 1. Further, p (n | k) is the probability of the node n in the cluster k. K is the total number of clusters k. The equation of Equation 6 represents that the probability p (n) of the node n can be decomposed into a combination of the probabilities p (n | k) of the node n in each cluster k.

The specific calculation method performed by the control unit 10 (second clustering unit 105) may be the same as the method described in Japanese Patent Application No. 2017-034888. Hereinafter, as a specific calculation process, an example of the process based on the method described in Japanese Patent Application No. 2017-034888 will be described with reference to the flowchart of FIG.

In the procedure of FIG. 4, first, the control unit 10 generates a transition probability matrix T _nm for the generated two-part network (step SB1). The transition probability matrix T _nm expresses the probability (that is, the transition probability) that the agent (in other words, the probability value of the node m) transitions (random walks) by following the link from the node m in the network to the node n. It is a thing. _{In the present embodiment, the control unit 10 determines the transition probability matrix T nm} by assuming that, for example, one or more links exiting from the node are selected by the agent with a probability corresponding _{to the weight w nk set in step SA4.} That is, the control unit 10 increases the value of the transition probability for the link as the value of _{the weight w nk increases.} For the transition probability matrix, see also Japanese Patent Application Laid-Open No. 2013-168127, Japanese Patent Application Laid-Open No. 2016-029526, and Japanese Patent Application Laid-Open No. 2016-218531.

Next, the control unit 10 calculates the steady-state link probability (step SB2). In this calculation, first, the transition probability matrix T _nm of the two-part network obtained in step SB1 is used, and the probability that each node has in the steady state of the stochastic transition (random walk) in the two-part network (node probability in the steady state). ) Is calculated. In this calculation, for example, the calculation of the following equation 7 is repeated until a steady state is reached.

In the equation of equation 7, _pt (n) is the probability that node n has at discrete time t. P _t when it becomes a steady state repeatedly calculate the number 7 formula (n) is a node probability p ^Stead at steady state of the node n (n).

Next, the control unit 10 calculates the link probability in the steady state from ^{the node probability p stead} (n) of each node n in the steady state according to the following equation (8).

The link probability is _{obtained by multiplying the node probability pt} (n) by the transition probability of the link l (L) exiting the node. The steady-state link probability for the link l (the left side of the equation of equation 8) is the starting point of the link l included in the _{transition probability matrix T nm with respect to the steady-state node probability of the node starting from the link l.} It is the product of the transition probability from the node to the end node.

In Japanese Patent Application Laid-Open No. 2016-029526 and Japanese Patent Application Laid-Open No. 2016-218531, passage information τ _n ^(d) (d is an integer from 1 to D. N is an integer from 1 to D, which is observation data obtained by virtual observation of D times. The node identification number) was used as training data. On the other hand, in the example described below, under the reasonable assumption that the number of observations D is sufficiently large (much more than the number of nodes N _{), instead of τ n} ^(d) , the passing information regarding the actual link l is used. The equation of equation 9 is used.

Here, n is a node identification number. Δ is Kronecker's δ. That is, the passage information (learning data) regarding the real link l of the node n defined by the equation of Equation 9 is such that the node n is the terminal end of link l or the starting node (initial end of link l) of the real link l. ), The value is 1, otherwise the value is 0. The control unit 10 generates such passage information as learning data from the information of the two-part network. The generated passage information is used in the calculation of the EM (Expectation Maximization) algorithm described later.

^{Further, in the present embodiment, instead of the ratio γ (d)} (k) of the cluster k to the entire plurality of clusters (components) in each virtual observation in JP-A-2016-209526, the actual link is used. _{For l, the ratio γ lk} (with tilde) defined by the equation (3) of Equation 11 described later is used.

Further, by replacing the number of observations d with the number l of the actual link, the expression of the sum of the functions is replaced as follows.

The second term on the right-hand side of the equation (1) of Equation 11 described later is such a replacement for the same equation described in JP-A-2016-029526 and the like.

Returning to the explanation of the procedure of Figure 4, then the control unit 10, the probability p _t (n | k) and severity [pi _k ^{new new,} and the initial value of the ratio gamma _lk provisionally decided, the number of repetitions of the counter g Initialize the value to 0 (step SB3). The probability _pt (n | k) is the probability of the node n in the cluster k. The importance π _k ^new is the importance of the cluster k. Further, γ _lk is the ratio of the cluster k to the entire plurality of clusters in the link l.

Next, the control unit 10 repeatedly calculates the EM algorithm using the equations (1), (2), and (3) of the following equation 11.

That is, first, the control unit 10 _{calculates the ratio γ lk} using the equation (3) (step SB4) (E step of the EM algorithm). In the first iteration of this calculation, the initial value tentatively determined in step SB3 is used.

Next, the control unit 10 replaces _{the current probability pt} (n | k) and the importance π _k ^new with the values _pt-1 (n | k) _{one time ago and the importance π k} ^{old (} Step SB5). _{Then, the probability pt} (n | k) and the importance π _k ^new are calculated according to the equations (1) and (2) (step SB6) (M step of the EM algorithm). More specifically, in step SB6, ^{the probability p is calculated by first calculating the new} _{importance π k} new according to the equation (2) and then calculating the equation (1) using this new importance. _{Find t} (n | k). Here, α is a positive real number and is a parameter that determines the size of the cluster, and a predetermined value may be used.

Then, the control unit 10 increments the counter g of the number of repetitive calculations (step SB7), determines whether or not the counter g has reached a predetermined value G (step SB8), and if not, step SB4. The process of ~ SB7 is repeated. The value G is the number of repetitions required for the calculations in steps SB4 to SB6 to converge in the calculation method of the present embodiment, and is predetermined by experiments, empirical knowledge, and the like.

When the control unit 10 determines in step SB8 that the counter g has reached the value G, the control unit 10 considers that the iterative calculation has converged and ends the process of FIG. After the determination result in step SB8 is Yes, the control unit 10 calculates the degree of belonging γ (k | n) of the node n to the cluster k according to the equation of Equation 12 (step SA6).

Of [pi _k and p in the formula (n | k) is calculated of the EM algorithm (step SB4~ step SB6) repeated by finally the obtained [pi _k ^{new new} and p _t of | a (n k). The equation of equation 12 is an equation for calculating γ (k | n) indicating the degree (affiliation degree) of the node n belonging to the cluster k from _{π k and p (n | k) according to Bayes' theorem.} The control unit 10 outputs the degree of belonging γ (k | n) thus obtained as a clustering result (step SA7). The degree of belonging γ (k | n) is information representing the result of soft clustering of the node n.

As another example, the control unit 10 may output a obtained binarized degree of belonging γ (k | n) with a predetermined threshold value as a clustering result. This clustering result indicates that the node n belongs to the cluster k in which the value of the degree of belonging γ (k | n) is equal to or greater than the threshold value (the value of the binarization result is 1). Depending on the value of the set threshold value, there may be a plurality of clusters k in which the binarization result is 1 for the node n, but this can be regarded as a kind of soft clustering result.

Further, the control unit 10 extracts only the clustering results for some important clusters from the clustering results for all K clusters with k = 1 to K (total number of clusters) used in the iterative calculation, and finally. It may be output as a clustering result. Important clusters may be determined based on the _{importance π k. For example, a cluster k whose final importance π k} obtained when the iterative calculation converges is equal to or higher than a predetermined threshold value is extracted as an important cluster, or the importance π _k is within a predetermined order from the top. A certain cluster k may be extracted as an important cluster.

In the determination of convergence in step SB8, instead of the method illustrated in FIG. 4, the same methods as described in JP2013-168127A, 2016-029526A, and 2016-218531A are the same. , When the amount of change in the evaluation value Qt for each repetition becomes a minute value (less than the threshold value), it may be determined that the repetition calculation has converged.

[Modification example]
Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and can be implemented in various other embodiments. For example, the present invention may be carried out by modifying the above-described embodiment as follows. The above-described embodiment and the following modifications may be combined with each other.

In the above-described embodiment, _{the parameter of the weight w nk} may be obtained from the data. For example, the control unit 10 _{may obtain the parameter d k} ² of the equation of equation 5 according to the equation of equation 13.

In the equation of equation 13, C _k represents the cluster obtained by the k-means method in step SA2, and N _k represents the number of data points belonging to C _k. According to this configuration, the parameters of the activation function of Equation 5 can be obtained from the vector data.

In the above-described embodiment, the weight of the link is obtained by the equation of Equation 5, but the method of obtaining the weight of the link is not limited to the method of the embodiment. _{For example, the link weight w nk} may be obtained by the equation of equation 14 which is a function to be reversed. In the equation of equation 14, C is a constant of C> 0 and γ> 0.

In addition, M predetermined feature nodes are selected in order of increasing Euclidean distance from the data point x _n, and the weight of the link between the selected node and the data point is set to a positive value, _{for example, w nk = 1.} The weight of the link between the feature node other than the selected feature node and the data point _{may be w nk} = 0. In the present invention, it is necessary to select a plurality of feature points (or feature nodes corresponding to them), and the accuracy of clustering depends on how to select these. Choosing feature points where the distribution of data points is dense is one of the methods to improve the accuracy of clustering. According to this modification, since the feature point is selected as the center of the portion where the data distribution is locally dense, the accuracy of clustering is improved.

In the above-described embodiment, the vector data is clustered by the k-means method in step SA2, but the method of clustering the vector data in step SA2 is not limited to the k-means method. For example, instead of the k-means method, clustering may be performed by the K-medoids method. The K-means method is similar to the k-means method in that it divides the vector data into K clusters specified by the user, but instead of defining the center of the cluster by the average of the data points, the data points belonging to each cluster. Determine the representative point of the cluster from among them. When clustering is performed by the k-means method, the feature points do not always match the data points, but when clustering by the K-means method, the data points closest to the center of the locally dense part of the data distribution. Will be selected as a feature point.

Then, the representative point of the cluster k obtained by clustering K-Medoids method and r _k, is the corresponding characteristic node to each of the representative points of the K clusters. The representative point is a point in the cluster, and the total dissimilarity with points in the cluster other than that point is the minimum. Then, when clustering is performed by the K-medoids method in step SA2, the link weight w _nk obtained in step SA4 is calculated by the formula of equation 15. The equation (1) of Eq. 15 is the Euclidean distance between the _{representative point r k of} each cluster and each data point x _n. Equation (2) of Equation 15 is an activation function that increases the weight of the link connecting the node n of the data point and the feature node r as a positive value as the Euclidean distance is shorter. The link weight may be obtained by the equation (3) of Equation 15, which is a function to be reversed when the vector data is clustered by the K-medoids method. In equation (3) of Eq. 15, C is a constant of C> 0 and γ> 0. When the vector data is clustered by the K-medoids method, _{M predetermined feature nodes are selected in order of proximity to the data points x n,} and the weight of the link between the selected nodes and the data points is weighted. For example, w _nk = 1 may be set to a positive value, and the weight of the link between the feature node other than the selected feature node and the data point may be set _{to w nk = 0.}

Further, in step SA2, vector data may be clustered using a mixed Gaussian model. For mixed Gaussian models, see, for example, Chapter 9 of Bishop, CM Pattern Recognition and Machine Learning (Springer). In the method of clustering vector data using a mixed Gaussian model, K Gaussian distributions can be obtained. The control unit 10 associates a feature node with each of the obtained K Gaussian distributions. Then, when the vector data is clustered using the mixed Gaussian model in step SA2, the link weight w _nk obtained in step SA4 is calculated by the equation of equation 16. The equation of equation 16 defines the weight of the link between the feature node k corresponding to the Gaussian distribution and the node n of the data points as the contribution of _{the data points x n to the cluster k.} Then, for example, the contribution degree γ _nk is set as the link weight w _nk .

The mixed Gaussian model locally approximates the shading of the data distribution with an ellipsoid. The length of each axis of the ellipsoid is set, for example, by being specified by the user of the information processing apparatus 1 in the operation unit 12 according to the data. On the other hand, the k-means clustering method or the K-means method assumes that the local data distribution is a sphere. When the mixed Gaussian model is used, the number of feature points is smaller, and the number of feature points is smaller, so that the amount of clustering calculation can be suppressed.

1 ... Information processing device, 10 ... Control unit, 11 ... Storage unit, 12 ... Operation unit, 13 ... Display unit, 14 ... Communication unit, 101 ... Acquisition unit, 102 ... First clustering unit, 103 ... Generation unit, 104 ... Calculation unit, 105 ... Second clustering unit.

Claims (10)

An acquisition method for acquiring vector data in which multiple components are represented by vectors,
A first clustering means for clustering the vector data by a parametric method,
A generation means for generating a two-part network having a data point representing the vector data and a feature point of each cluster obtained by the first clustering means as a node, and a generation means.
A calculation means for calculating the link weight connecting the node of the data point and the node of the feature point, and
A second clustering means for clustering the nodes by determining the transition probability between the nodes via the link in the two-part network according to the link weight and executing the iterative calculation of the stochastic process of the transition between the nodes. Information processing device equipped with. The first clustering means performs clustering by the k-means method and performs clustering.
The information processing apparatus according to claim 1, wherein the center of the cluster obtained by the first clustering means is the feature point. The information processing apparatus according to claim 2, wherein the weight of the link is calculated by an activation function that increases the Euclidean distance between the feature point and the data point as a positive value as the distance is shorter. When the data point is x _n and the feature point is m _k , the Euclidean distance is calculated by the equation (1).
The information processing device according to claim 3.

The first clustering means performs clustering by the K-medoids method and performs clustering.
The information processing apparatus according to claim 1, wherein the representative point of the cluster obtained by the first clustering means is the feature point. The first clustering means performs clustering by a mixed Gaussian model and performs clustering.
Each of the plurality of Gaussian distributions is used as the feature point.
The information processing apparatus according to claim 1, wherein the contribution of the data points to the cluster is the link weight. The information processing apparatus according to claim 6, wherein the mixed Gaussian model approximates the distribution of the vector data with an ellipsoid. The second clustering means calculates the importance of the cluster and calculates the importance of the cluster.
The information processing apparatus according to any one of claims 1 to 7, which extracts clusters whose importance satisfies a predetermined condition. The calculation means is
Calculate the Euclidean distance between the feature point and the data point,
A predetermined number of nodes of the feature points are selected in order of increasing distance from the data points to the Euclidean distance.
The first aspect of claim 1, wherein the link weight between the node of the selected feature point and the data point is a positive value, and the link weight between the node of the feature point other than the selected feature point and the data point is 0. Information processing device. Computer,
An acquisition method for acquiring vector data in which multiple components are represented by vectors,
A first clustering means for clustering the vector data by a parametric method,
A generation means for generating a two-part network having a data point representing the vector data and a feature point of each cluster obtained by the first clustering means as a node, and a generation means.
A calculation means for calculating the link weight connecting the node of the data point and the node of the feature point, and
A second clustering means for clustering the nodes by determining the transition probability between the nodes via the link in the two-part network according to the link weight and executing the iterative calculation of the stochastic process of the transition between the nodes. A program to function as.

Images