JP7511278B2 - Information processing device, information processing method, and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program, for example, an information processing device, an information processing method, and a program that learn an input distribution structure for an input vector and cluster the input distribution structure.

Many techniques have been researched and developed, known as clustering, which extracts the input distribution structure of input data and classifies it into classes based on that input distribution structure. For example, the well-known k-means method is a non-hierarchical clustering technique that uses the cluster means to create any specified number of clusters. k-means is a popular clustering method that is a simple algorithm and computationally efficient.

A method called Self-Organizing Incremental Neural Network (SOINN) has also been proposed (Patent Documents 1 and 2). SOINN has many advantages, such as being able to learn non-stationary inputs by autonomously managing the number of nodes, and being able to extract an appropriate number of classes and a topological structure even for classes with complex distribution shapes. As an example of an application of SOINN, for example, in pattern recognition, after learning classes of hiragana characters, it is possible to additionally learn classes of katakana characters, etc.

JP 2008-217246 A JP 2014-164396 A

However, it has been found that when learning input vectors using the above-mentioned clustering method, the learning results obtained differ depending on the input order of the input vectors and differences in parameters. Examples of such different learning results are described below. Figure 24 shows the input data to be learned. In this input data, there are two parts where the two-dimensional input vectors are distributed in an arc shape. Figure 25 shows examples of multiple learning results generated by differences in the input order of the input vectors and parameters when the input data in Figure 24 is learned using SOINN.

In Figure 25, eight different learning results were obtained due to differences in the learning conditions, even though the same input data was learned. When multiple different learning results are obtained like this, it is necessary to manually compare the different learning results and select the appropriate learning result, which is time-consuming and leads to inefficiency.

The present invention has been made in consideration of the above circumstances, and aims to provide an information processing device, information processing method, and program that can perform clustering that more accurately reflects the structure of input data while reducing human work.

An information processing device according to one embodiment of the present invention is an information processing device that learns the input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, and includes: a distance matrix calculation unit that calculates a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results for all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions; an evaluation index calculation unit that obtains the minimum value of each column and each row for each distance matrix and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition; a score calculation unit that calculates a score indicating the degree of similarity of each cluster of each learning result to clusters of other learning results other than the learning result of interest based on the evaluation index calculated for all distance matrices; and a learning result selection unit that selects a final learning result from a plurality of previous learning results based on the score of each cluster of each learning result calculated by the score calculation unit.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, wherein the score calculation unit generates an evaluation index matrix composed of evaluation indexes for all combinations between two clusters selected one from each of two learning results based on the evaluation indexes calculated for all distance matrices, calculates a score indicating whether a cluster included in one learning result is similar to a cluster of another learning result based on the evaluation indexes included in the evaluation index matrix, and generates a score matrix including scores for clusters belonging to all learning results, and it is desirable that the learning result selection unit selects a final learning result from the multiple learning results based on the score matrix.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that the evaluation index calculation unit calculates the average distance between nodes belonging to each of the two clusters, calculates the average value of the two average distances calculated, and calculates an evaluation index for each distance matrix based on the number of the minimum values in each column and each row of each distance matrix that are equal to or greater than the average value.

An information processing device according to one embodiment of the present invention is the information processing device described above, and it is desirable that the evaluation index is a value obtained by dividing the number of minimum values in each column and each row of each distance matrix that are greater than or equal to the average value by the total number of nodes belonging to the two clusters.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that the score calculation unit compares each evaluation index included in the evaluation index matrix with a threshold value, assigns a score based on the comparison result, and calculates a total score for each cluster.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that the learning result selection unit obtains the minimum value among the total scores of the clusters belonging to each learning result, and selects the learning result corresponding to the maximum value among the obtained minimum values.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that the learning result selection unit selects multiple non-overlapping clusters from the clusters of the multiple learning results based on the score matrix, and selects the final learning result composed of the selected clusters.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that the learning result selection unit selects one learning result from the multiple learning results as the final learning result based on the score matrix.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and further includes a deletion processing unit that performs processing to delete clusters from each learning result, and the deletion processing unit preferably includes a score sorting unit that sorts the clusters belonging to each learning result in order of their scores, a reference cluster selection unit that selects the cluster with the highest score as a reference cluster, a target cluster selection unit that selects from the sorted clusters the cluster with the highest score among clusters with scores lower than the reference cluster as a target cluster, a judgment value calculation unit that calculates a judgment value used to judge a cluster based on the reference cluster and the target cluster, a cluster judgment unit that selects a cluster to be deleted based on the judgment value, and a cluster deletion unit that deletes the cluster to be deleted.

An information processing device according to one embodiment of the present invention is the information processing device described above, wherein the distance matrix calculation unit calculates the distance matrix for the reference cluster and the target cluster, the distance matrix having a number of rows equal to the number of nodes in the reference cluster and a number of columns equal to the number of nodes in the target cluster, and the evaluation index calculation unit obtains the minimum value of each column and each row of the distance matrix calculated for the reference cluster and the target cluster, calculates the average distance between nodes belonging to the reference cluster and the target cluster, calculates the average value of the two calculated average distances, and calculates the minimum value of the distance matrix calculated for the reference cluster and the target cluster. It is desirable that the judgment value calculation unit counts a first value which is the number of the minimum values in each row of the distance matrix that are equal to or greater than the average value, and a second value which is the number of the minimum values in each column of the distance matrix that are equal to or greater than the average value, and the judgment value calculation unit calculates a first judgment value by dividing the first value by the number of nodes in the reference cluster and a second judgment value by dividing the second value by the number of nodes in the reference cluster, and if either the first or second judgment value is smaller than a second threshold, the cluster judgment unit selects the target cluster as a cluster to be deleted, and the cluster deletion unit deletes the cluster to be deleted.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and when the deletion processing unit deletes the cluster to be deleted, or when the cluster determination unit determines that the first and second determination values are greater than a second threshold value, it is desirable that the target cluster selection unit selects the cluster having the next lowest score after the target cluster as a new target cluster, and the deletion processing unit performs a process of deleting a cluster based on the reference cluster and the new target cluster.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and when there is no cluster with the next lowest score after the target cluster, it is desirable that the reference cluster selection unit selects, in addition to the reference cluster, the cluster with the next lowest score after the reference cluster as a new reference cluster, and the deletion processing unit performs processing to delete clusters based on the new reference cluster.

An information processing device according to one embodiment of the present invention is the above-mentioned information processing device, and it is desirable that, when there is no cluster with the next lowest score after the reference cluster, the deletion processing unit terminates the process of deleting clusters.

An information processing method according to one embodiment of the present invention is an information processing method for learning an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, in which a distance matrix calculation unit calculates a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results, for all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions, an evaluation index calculation unit obtains the minimum value of each column and each row for each distance matrix, and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition, a score calculation unit calculates a score indicating the degree of similarity of each cluster of each learning result to clusters of other learning results other than the learning result of interest, based on the evaluation index calculated for all distance matrices, and a learning result selection unit selects a final learning result from the plurality of learning results in the previous period, based on the score of each cluster of each learning result calculated by the score calculation unit.

A program according to one embodiment of the present invention is a program that causes a computer to execute a process of learning the input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, and causes the computer to execute the following processes: for all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions, a process of calculating a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results; a process of obtaining the minimum value of each column and each row of each distance matrix, and calculating an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition; a process of calculating a score for each cluster of each learning result that indicates the degree of similarity between clusters of other learning results other than the learning result of interest based on the evaluation index calculated for all distance matrices; and a process of selecting a final learning result from the plurality of learning results in the previous period based on the calculated score for each cluster of each learning result.

According to the present invention, it is possible to provide an information processing device, information processing method and program that can perform clustering that more accurately reflects the structure of input data while reducing human work.

FIG. 1 is a diagram illustrating an example of a system configuration for realizing an information processing device according to a first embodiment. FIG. 13 is a diagram showing the distribution of input vectors in input data. FIG. 3 is a diagram showing learning results generated by k-means and SOINN using the input data of FIG. 2 . This is a diagram showing five examples of different learning results generated by different input vector input orders and parameters when learning the same input data. 1 is a diagram illustrating a basic configuration of an information processing device according to a first embodiment; FIG. 2 is a diagram showing a more detailed configuration of the information processing device according to the first embodiment. 11 is a flowchart of a learning result selection operation of the information processing device according to the first embodiment; FIG. 5 is a diagram showing an example of a score matrix obtained from the five learning results shown in FIG. 4 . 3A to 3C are diagrams showing examples of four different learning results generated by differences in the input order of input vectors and parameters when learning the input data of FIG. 2 . FIG. 11 is a diagram illustrating a functional configuration of an information processing device according to a second embodiment. 13 is a flowchart of a cluster deletion process in the information processing device according to the second embodiment; FIG. 9 is a diagram showing an example of a score matrix obtained from the four learning results shown in FIG. 8 . FIG. 13 is a diagram showing an example of sorting clusters based on the score matrix of FIG. 12. FIG. 11 is a diagram showing the progress of a cluster deletion process according to the second embodiment; FIG. 13 is a diagram showing an example of a score matrix after cluster deletion. FIG. 15 is a diagram showing a final learning result including clusters selected based on the score matrix of FIG. 14 . FIG. 25 is a diagram showing an example in which only a portion of the clusters could be learned for the input data shown in FIG. 24. FIG. 17 is a diagram showing a score matrix for the case of FIG. 16. FIG. 13 is a diagram illustrating a configuration of an information processing device according to a third embodiment. 13 is a flowchart of a learning result deletion process in the information processing device according to the third embodiment. FIG. 20 is a diagram showing an example of sorting clusters based on the score matrix of FIG. 18. FIG. 13 is a diagram showing the progress of a cluster deletion process according to the third embodiment. FIG. 13 is a diagram showing an example of a score matrix after deleting learning results. FIG. 2 is a diagram showing input data to be learned; 25 is a diagram showing examples of multiple learning results generated by differences in the input order of input vectors and parameters when the input data of FIG. 24 is learned by SOINN.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In each drawing, the same elements are given the same reference numerals, and duplicate explanations will be omitted as necessary.

First embodiment
FIG. 1 is a diagram showing an example of a system configuration for realizing an information processing device according to the first embodiment. The information processing device 100 can be realized by a computer 10 such as a dedicated computer or a personal computer (PC). However, the computer does not need to be physically single, and may be multiple when performing distributed processing. As shown in FIG. 1, the computer 10 has a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, and a RAM (Random Access Memory) 13, which are connected to each other via a bus 14. Note that an OS software for operating the computer will not be described, but it is assumed that the computer that constructs this information processing device also has such software.

An input/output interface 15 is also connected to the bus 14. The input/output interface 15 is connected to, for example, an input unit 16 including a keyboard, mouse, sensor, etc., a display including a CRT, LCD, etc., an output unit 17 including headphones, speakers, etc., a storage unit 18 including a hard disk, etc., and a communication unit 19 including a modem, terminal adapter, etc.

The CPU 11 executes various processes in accordance with various programs stored in the ROM 12 or various programs loaded from the storage unit 18 to the RAM 13; in this embodiment, for example, the processes of each unit of the information processing device 100 described below. A GPU (Graphics Processing Unit) may be provided separately from the CPU 11 and be made to execute the same processes as the CPU 11. The GPU is suitable for applications in which standard processes are executed in parallel, and by applying it to the learning process described below, it is possible to improve the processing speed compared to the CPU 11. The RAM 13 also appropriately stores data necessary for the CPU 11 and the GPU to execute various processes.

The communication unit 19 performs communication processing, for example, via the Internet (not shown), transmits data provided by the CPU 11, and outputs data received from a communication partner to the CPU 11, RAM 13, and storage unit 18. The storage unit 18 communicates with the CPU 11 and stores and erases information. The communication unit 19 also performs communication processing of analog or digital signals with other devices.

The input/output interface 15 is also connected to a drive 20 as required, and, for example, a magnetic disk 20A, an optical disk 20B, a flexible disk 20C, or a semiconductor memory 20D is appropriately attached, and computer programs read from them are installed in the memory unit 18 as required.

Next, each process in the information processing device 100 according to the present embodiment will be described. The information processing device 100 has nodes described by n-dimensional vectors (n is an integer equal to or greater than 1), and information on the nodes is stored in a storage unit such as the RAM 13.

The input data is input as an n-dimensional input vector. For example, the input vector is stored in a temporary storage unit (e.g., RAM 13).

The information processing device 100 according to the first embodiment will be specifically described below. The information processing device 100 is configured to select a learning result that appropriately reflects the structure of the input data from a plurality of different learning results by performing the processing described below. Each learning result is a clustering result composed of a plurality of nodes and information indicating the cluster to which each node belongs (e.g., cluster label), and may be, for example, a neural network generated by the above-mentioned E-SOINN or LB-SOINN, or may be, for example, a method of clustering the input data based on an unsupervised clustering method such as the k-means method, and generating some of the input data divided into clusters as nodes. FIG. 2 shows the distribution of the input vector of the input data. FIG. 3 shows the learning result generated by the k-means method and SOINN using the input data of FIG. 2. In the input data of FIG. 2, the input data is distributed in four clusters, three of which are close to each other, and the boundaries are unclear due to noise. When this input data is trained using the k-means method and SOINN, it can be seen that the same four clusters are detected, although the distribution of nodes in the training results is slightly different due to the difference in the methods.

Figure 4 shows five examples of different learning results generated by differences in the input order of input vectors and parameters when learning from the same input data. Here, the same input data as shown in Figure 24 was used. As shown in Figure 4, learning results 1 to 5 differ in the number and shape of clusters.

As shown in FIG. 4, although the learning results are different, it can be seen that there are clusters that are considered to be the same among the different learning results. For example, cluster 0 of learning result 1 is similar to cluster 0 of learning result 2 and cluster 0 of learning result 5, and is considered to be the same. Even if the input order, learning method, parameters, etc. are different, since the same input data is learned, the frequency with which a cluster that the input data truly possesses appears as a result is considered to be higher than the frequency with which a noise-like cluster appears due to the influence of the input order, learning method, parameters, etc. Therefore, it is considered that an appropriate cluster can be determined based on the frequency with which clusters that are considered to be the same appear in multiple learning results. In this embodiment, a process of selecting an appropriate learning result based on the frequency of appearance of each cluster in multiple learning results will be described.

Below, the components and operations of the information processing device 100 will be described with reference to Figures 5-7. Figure 5 shows a schematic diagram of the basic configuration of the information processing device 100 according to the first embodiment. Figure 6 shows a more detailed configuration of the information processing device 100 according to the first embodiment. Figure 7 shows a flowchart of the learning result selection operation of the information processing device 100 according to the first embodiment. The information processing device 100 has a distance matrix calculation unit 1, an evaluation index calculation unit 2, a score calculation unit 3, and a learning result selection unit 4. The evaluation index calculation unit 2 has a calculation unit 2A and an evaluation index matrix generation unit 2B. The score calculation unit 3 has a calculation unit 3A and a score matrix generation unit 3B.

Step S1: Distance Matrix Calculation The distance matrix calculation unit 1 calculates the distance matrix D based on the following steps S11 to S13.

Step S11
The distance matrix calculation unit 1 selects two arbitrary learning results SA and SB from a plurality of learning results (r, where r is equal to or greater than 2).

Step S12
The distance matrix calculation unit 1 selects any one cluster A from p clusters CA1 to CAp included in the learning result SA, and any one cluster B from q clusters CB1 to CBq included in the learning result SB.

Step S13
The distance matrix calculation unit 1 calculates the distances d aibj between nodes for all combinations of nodes a1 to am (ai, 1≦i≦m) included in cluster A and nodes b1 to bn (bj, 1≦j≦n) included in cluster _B , and calculates a distance matrix D with m rows and n columns.

Step S2: Calculate Evaluation Index The calculation unit 2A of the evaluation index calculation unit 2 calculates the evaluation index based on the following steps S21 to S25.

Step S21
The calculation unit 2A obtains the minimum value min _ai (i.e., min _a1 to min _am ) for each row (d _aib1 to d _aibn ) of the distance matrix D, and obtains the minimum value minb _j (i.e., min _b1 to min _bn ) for each column (d _a1bj to d _ambj ) of the distance matrix D.

式［２］において、ｄ_ｉはノードｉとノードｉの属するクラスタに含まれるノードとの距離の最小値である。 Step S22
The calculation unit 2A calculates the average distance d _mean between nodes belonging to the same cluster. Since d _mean is the average distance between nodes identified as being in the same cluster, if the distance is equal to or greater than this average value, it can be considered as an index for identifying the nodes as being in a different cluster, and if the distance is shorter than this average value, it can be considered as an index for identifying the nodes as being in the same cluster. For example, d _mean can be written as follows:

In formula [2], d _i is the minimum distance between node i and the nodes contained in the cluster to which node i belongs.

Note that d _mean in the formula [2] is an example, and may be defined as, for example, the following formula [3].

Furthermore, d _mean may be set to the larger or smaller value of the values shown in the following expressions [4] and [5].

Furthermore, separate d _mean may be prepared for the minimum values min _a1 to min _am of each row and the minimum values min _b1 to min _bn of each column, and used accordingly. For the minimum values min _a1 to min _am of each row, d _mean calculated only for cluster A may be used, for example, as shown in formula [6]. For the minimum values min _b1 to min _bn of each column, d _mean calculated only for cluster B may be used, for example, as shown in formula [7].

In the above, the learning result including multiple nodes has been described, but when the input data is learned by SOINN or the like, the learning result includes multiple nodes and multiple edges connecting two nodes. In this way, when edges exist between nodes, it is desirable to use the average length of edges belonging to the same cluster as d _mean . Since edges represent the relationship between nodes and edge lengths represent their similarity, by using edges, it is possible to prevent the distance to a node that happens to exist in the vicinity due to the influence of noise from being used in the calculation of d _mean . In addition, in this embodiment, the arithmetic mean (arithmetic mean) is used as the average length, but this is not limited as long as the central tendency of the distance distribution between neighboring nodes can be expressed numerically. For example, another average index such as the square mean, the geometric mean, or the harmonic mean may be used as the average length, or another statistical representative value such as the median or the mode may be used. In the following, the "average distance" is used to include the average index and statistical representative value described above in addition to the arithmetic mean.

Step S23
Calculation unit 2A counts the number NUMa of the minimum values min _a1 to min _am in each row that are equal to or greater than the average distance d _mean , and counts the number NUMb of the minimum values min _b1 to min _bn in each column that are equal to or greater than the average distance d _mean .

Step S24
The calculation unit 2A calculates the sum of the acquired numbers NUMa and NUMb divided by the total number of nodes (m+n) in clusters A and B as an evaluation index ES.

Here, the number of nodes that are not similar between the two clusters is detected by counting the number of the minimum values that are equal to or greater than the average value _dmean . In addition, the ratio of the number of nodes that are not similar to the total number of nodes is calculated by dividing the number of nodes that are not similar in the formula [8] by the total number of nodes in the two clusters. In other words, the evaluation index ES can be said to be a value that represents the dissimilarity, where a large value indicates that the two clusters are not similar, and a small value indicates that the two clusters are similar.

In this embodiment, the distance matrix calculation unit 1 performs the calculation for all combinations of any one of the p clusters CA1 to CAp included in the learning result SA and any one of the q clusters CB1 to CBq included in the learning result SB. Then, the calculation unit 2A calculates the evaluation index for all combinations of the two clusters.

Step S25
That is, the calculation unit 2A judges whether the evaluation index has been calculated for all combinations of two clusters. If the evaluation index has not been calculated for all combinations of two clusters, the process returns to step S12. It goes without saying that when returning to step S12, two new clusters are selected so that they are not the same as the combinations of two clusters already selected. If the evaluation index has been calculated for all combinations of two clusters, the process proceeds to step S3.

Step S3: Generate evaluation index matrix When the calculation of evaluation indexes has been completed for all cluster combinations, evaluation index matrix generator 2B of evaluation index calculator 2 generates an evaluation index matrix EM with p rows and q columns. That is, the row numbers of the evaluation index matrix EM indicate the cluster numbers included in learning result SA, and the column numbers indicate the cluster numbers included in learning result SB.

Step S4: Score Calculation The calculation section 3A of the score calculation unit 3 calculates the score based on the following steps S41 and S42.

Step S41
The calculation unit 3A compares each evaluation index included in the evaluation index matrix EM with a threshold value TH, and gives a high score (for example, a score of +1) to both clusters that constitute an evaluation index ES that is equal to or less than the threshold value TH (similar), and gives a low score (for example, a score of 0) to both clusters that constitute an evaluation index ES that is greater than the threshold value (dissimilar). Then, the calculation unit 3A calculates the total score of each cluster (CA1 to CAp) of the learning result SA. In other words, the calculation unit 3A calculates the total of the scores given to each row (ES _i1 , ..., SE _iq ) of the evaluation index matrix EM for each row. The total of the scores of (ES _i1 , ..., SE _iq ) becomes the score of the cluster CAi. The calculation unit 3A also calculates the total score of each cluster (CB1 to CBq) of the learning result SB. In other words, the calculation unit 3A calculates the total of the scores given to each column (ES _1j , ..., SE _pj ) of the evaluation index matrix EM for each column. The sum of the scores of (ES _1j , ..., SE _pj ) is the score of cluster CBj. Note that the method of calculating the score is not limited to this, and it is also possible to add an application such as adding a high score to the one that is equal to or less than the threshold TH and has the lowest evaluation index ES in each row (the one that is most similar), or the one that has the lowest evaluation index ES in each column (the one that is most similar).

In this embodiment, the above total score is calculated for all combinations of any two learning results selected from multiple learning results. That is, the distance matrix calculation unit 1 and the evaluation index calculation unit 2 perform the above-mentioned distance matrix calculation, evaluation index calculation, and evaluation index matrix generation for all combinations of any two learning results selected from multiple learning results. Then, based on these calculation results, the calculation unit 3A obtains the total score for each cluster of each learning result.

Step S42
That is, the calculation unit 3A judges whether the scores have been calculated for all combinations of two learning results. If the scores have not been calculated for all combinations of two learning results, the process returns to step S11. It goes without saying that when returning to step S11, two new learning results are selected so that they are not the same as the combinations of two learning results already selected. If the scores have been calculated for all combinations of two learning results, the process proceeds to step S5.

Step S5
Based on the acquired total score, the score matrix generation unit 3B of the score calculation unit 3 generates a score matrix SM with r columns in which the row numbers are the cluster numbers, as shown in the following formula [10]. Here, an element S _ij of the score matrix SM is the total score in cluster i of learning result j. Specifically, it is the sum of all the total scores of cluster i of learning result j calculated by the calculation unit 3A for all combinations of learning result j and all other learning results. In addition, there are elements that do not exist in the score matrix SM. For example, if cluster i of learning result j does not exist, S _ij is empty and does not participate in calculations, etc.

Here, we consider a specific example of the score matrix SM when five different learning results are generated. Figure 8 shows an example of the score matrix obtained from the five learning results shown in Figure 4. In Figure 8, the maximum number of clusters is 3, and for simplicity, the matrix elements (scores) are shown in a table format.

Step S6: Selection of Learning Result The learning result selection unit 4 selects a final learning result from the multiple learning results (here, learning results 1 to 5) based on the following steps S61 and S62.

Step S61
The learning result selection unit 4 refers to the score matrix SM and obtains the minimum scores Smin1 to Sminr for each learning result.

Step S62
The learning result selection unit 4 selects the learning result with the maximum value among the acquired minimum scores Smin1 to Sminr as the final learning result to be adopted. In other words, the learning result selection unit 4 extracts, for each learning result, the score of the cluster that is most dissimilar to the clusters included in the other learning results, and then selects the learning result having the most similar cluster from among them as the final learning result.

In the score matrix shown in Figure 8, the minimum scores for learning results 1 to 5 are 0, 0, 0, 0, and 2, respectively, and the maximum score among these is 2. Therefore, the learning result selection unit 4 selects learning result 5 as the final learning result. As shown in Figure 4, learning result 5 has the same number of clusters as the input data, which is two, and it can be seen that this is suitable as a representative final learning result for the input data.

Note that there may be cases where two or more learning results have the same minimum score. In this case, the final learning result may be selected based on any rule, such as the learning result with the smallest number or the learning result with the largest number. Alternatively, the scores of all clusters in each learning result with the same minimum score may be added up and compared to select the learning result with the largest score. This makes it possible to select a learning result with a relatively high score.

As described above, with this configuration, when learning the same input data, if multiple different learning results (clustering results) are obtained due to differences in the input order or parameters, the appropriate learning result can be automatically selected.

This eliminates the need for humans to select learning results, and reduces the time required for the work. In addition, by eliminating human judgment, it is possible to prevent variation in the learning results selected depending on the judge.

In this embodiment, the learning result with the maximum score among the minimum scores included in each learning result is set as the final learning result, but the selection of the final learning result is not limited to this. For example, the scores of all clusters in each learning result may be added up, a learning result whose total score is equal to or greater than a predetermined value may be selected, and a learning result with the maximum score among the minimum scores included in each selected learning result may be selected as the final learning result. In this case, a learning result having a cluster with a relatively high score can be selected by setting a threshold value for the total score. In addition, if it is acceptable for the final learning result to include a cluster with a relatively low score, the scores of all clusters in each learning result may be added up, and the learning result with the maximum total score may be selected as the final learning result. In this case, the clusters included in the final learning result may include clusters with low scores, but overall they are clusters with high scores. If the user selects from these clusters as appropriate, an appropriate learning result can be obtained.

また、評価指数ＥＳは、以下の式［１２］のように、類似度として定義してもよい。

評価指数ＥＳとして類似度を用いる場合には、ステップＳ４１において、得点算出部が、評価指数行列ＥＭに含まれる各評価指数を閾値と比較し、閾値以上（似ている）の評価指数ＥＳを構成する両クラスタに高い得点を与え（例えば得点＋１）、閾値よりも小さい（似ていない）の評価指数ＥＳを構成する両クラスタに低い得点（例えば得点０）を与えることで、評価指数ＥＳとして非類似度を用いる場合と同様の処理が可能である。以上説明したように、評価指数ＥＳは、クラスタ間の近似度合いを定量的に示すことができれば、いかように定義することも可能である。 In this embodiment, the formula [8] is shown as a specific example of the evaluation index ES of each cluster used by the evaluation index calculation unit 2, but the evaluation index ES is not limited to this example. For example, the evaluation index ES may be defined as the following formula [11].

The evaluation index ES may also be defined as a similarity as shown in the following equation [12].

When similarity is used as the evaluation index ES, in step S41, the score calculation unit compares each evaluation index included in the evaluation index matrix EM with a threshold value, and gives a high score (e.g., score +1) to both clusters that make up an evaluation index ES equal to or greater than the threshold value (similar), and gives a low score (e.g., score 0) to both clusters that make up an evaluation index ES that is smaller than the threshold value (dissimilar), thereby enabling processing similar to that when dissimilarity is used as the evaluation index ES. As explained above, the evaluation index ES can be defined in any way as long as it can quantitatively indicate the degree of similarity between clusters.

Embodiment 2
In the second embodiment, a modified example of the learning result selection according to the first embodiment will be described. In the first embodiment, a process of selecting an appropriate learning result based on the frequency of occurrence of each cluster in a plurality of learning results has been described. However, in some cases, there may be no appropriate learning results. FIG. 9 shows an example of four different learning results generated by differences in the input order of the input vectors and parameters when learning the input data of FIG. 24. The number of clusters in the input data is two, whereas the number of clusters is three, although all learning results reflect the characteristics of the input data, and it can be seen that the results do not adequately represent all clusters. In this embodiment, a process of selecting an appropriate cluster and creating a learning result based on the frequency of occurrence of each cluster in a plurality of learning results will be described.

Figure 10 shows a schematic configuration of an information processing device 200 according to the second embodiment. The information processing device 200 has a configuration in which a deletion processing unit 7 is added to the information processing device 100 according to the first embodiment. The deletion processing unit 7 has a score sorting unit 71, a reference cluster selection unit 72, a target cluster selection unit 73, a judgment value calculation unit 74, a cluster judgment unit 75 and a cluster deletion unit 76, and together with the distance matrix calculation unit 1 and the evaluation index calculation unit 2, it performs a process of deleting overlapping clusters.

Figure 11 shows a flowchart of the cluster deletion process in the information processing device 200 according to embodiment 2. Cluster deletion process step S7 is inserted between steps S5 and S6 in Figure 7, and includes the following steps S701 to S713.

Step S701
The score sorting unit 71 sorts all the clusters in order of scores, for example, in descending order, based on the score matrix SM. Fig. 12 shows an example of a score matrix obtained from the four learning results shown in Fig. 9. Fig. 13 shows an example of the clusters sorted based on the score matrix of Fig. 12.

Step S702
The reference cluster selection unit 72 selects the cluster with the highest score as the reference cluster.

Step S703
The target cluster selection unit 73 selects the cluster in the row immediately below the reference cluster in the sorted order (FIG. 13) as the target cluster.

Needless to say, the reference cluster and the target cluster are selected so that they are not the same as any combination of two clusters that have already been selected.

Step S704
Similar to step S1, distance matrix calculation unit 1 calculates a distance matrix D for a reference cluster (cluster A) and a target cluster (cluster B).

Step S705
As in step S21, the evaluation index calculation unit 2 obtains the minimum value min _ai (i.e., min _a1 to min _am ) for each row (d _aib1 to d _aibn ) of the distance matrix D, and obtains the minimum value minb _j (i.e., min _b1 to min _bn ) for each column (d _a1bj to d _ambj ) of the distance matrix D.

Step S706
Similar to step S22, the evaluation index calculation unit 2 calculates the average distance d _mean between nodes belonging to the same cluster.

Step S707
As in step S23, the evaluation index calculation unit 2 counts the number NUMa of the minimum values min _a1 to min _am in each row that are greater than or equal to the average distance d _mean , and counts the number NUMb of the minimum values min _b1 to min _bn in each column that are greater than or equal to the average distance d _mean .

Step S708
The judgment value calculation unit 74 divides the count number NUMa for each row by the number of nodes m in cluster A to calculate a judgment value RA indicating the proportion of nodes in cluster A that are dissimilar to cluster B. The judgment value calculation unit 74 also divides the count number NUMb for each column by the number of nodes n in cluster B to calculate a judgment value RB indicating the proportion of nodes in cluster B that are dissimilar to cluster A.

Step S709
The cluster determination unit 75 compares the judgment values RA and RB with a predetermined threshold value THr, and if either of the judgment values RA and RB is smaller than the predetermined threshold value THr, it determines that one of clusters A and B is included in or matches the other, and determines that cluster B (i.e., the target cluster) with a lower score or the same score should be deleted.

Step S710
When the cluster determination unit 75 determines that the target cluster should be deleted, the cluster deletion unit 76 deletes the score of the target cluster with the low score from the score matrix SM. As described later, the deleted cluster cannot be selected as the final learning result. Also, the row corresponding to the deleted cluster is deleted from the cluster sorting (FIG. 13). That is, the deleted cluster will not be selected as a reference cluster or a target cluster in the next process and thereafter. However, it can be used as a current target cluster for calculating the next target cluster in steps S711 and S712. That is, in this embodiment, when it is described that "a cluster one step below a certain cluster is selected in the sorted order (FIG. 13)", it means that a cluster one step below is selected in the order in which the row corresponding to the deleted cluster is deleted from the sorted order (FIG. 13).

Step S711
The target cluster selection unit 73 determines whether there is another cluster that can be selected as a new target cluster in the row immediately below the current target cluster in the sorted order (FIG. 13).

Step S712
If it is determined in step S711 that there is another cluster that can be selected as a new target cluster, the target cluster selection unit 73 selects that other cluster (i.e., the cluster one row below the current target cluster in the sorted order (Figure 13)) as the new target cluster, and then returns the process to step S704.

Step S713
If it is determined in step S711 that there is no other cluster that can be selected as a new target cluster, the reference cluster selection unit 72 determines whether there is another cluster that can be selected as a new reference cluster that remains undeleted in the row immediately below the current reference cluster in the sorted order (FIG. 13). If it is determined that there is no cluster that can be selected as a new reference cluster, the cluster deletion process of step S7 is terminated.

Step S714
If it is determined in step S713 that a cluster that can be selected as a new reference cluster exists, the reference cluster selection unit 72 selects the cluster one cluster below the current reference cluster in the sorted order (FIG. 13) as the new reference cluster, and then returns the process to step S703.

The cluster deletion process in step S7 above will be explained using a specific example based on the four learning results in Figure 9. Figure 14 shows the progress of the cluster deletion process in the second embodiment.

In the cluster deletion process in step S7, processing is performed between the reference cluster with the highest score and the target cluster in descending order of score. After that, the cluster with the highest score among the remaining clusters is selected as the reference cluster, and the cycle of processing between the new reference cluster and the target cluster in descending order of score is repeated until a new reference cluster cannot be selected.

Delete Transition 1
First, cluster 0 of learning result 1, which has the highest score, becomes the reference cluster, and processing is performed between it and the other clusters. Rows marked with * in the table of deletion transition 1 in Fig. 14 are determined to match cluster 0 of learning result 1, to include cluster 0 of learning result 1, or to be included in cluster 0 of learning result 1. Therefore, rows marked with * are deleted.

Delete Transition 2
After the process shown in deletion transition 1 is completed, the process shown in deletion transition 2 is executed. In deletion transition 2 in FIG. 14, rows corresponding to the clusters deleted in deletion transition 1 in sorted order are displayed excluding them. Here, processing is executed between cluster 1 of learning result 3, which has the highest score among the remaining clusters excluding cluster 0 of learning result 1, and the remaining clusters other than cluster 0 of learning result 1. Rows marked with * in the table of deletion transition 2 in FIG. 14 are determined to match cluster 1 of learning result 3, include cluster 1 of learning result 3, or be included in cluster 1 of learning result 3. Therefore, the rows marked with * are deleted.

Delete Transition 3
After the process shown in deletion transition 2 is completed, the process shown in deletion transition 3 is executed. In deletion transition 3 in FIG. 14, rows corresponding to the clusters deleted in deletion transitions 1 and 2 in the sorted order are excluded and displayed. Since there are no unprocessed cluster combinations in the state of deletion transition 3, the cluster deletion process ends here. FIG. 15 shows an example of the score matrix after cluster deletion. The scores of each cluster before cluster deletion are the same as in FIG. 12, but here, the clusters marked with an x have been deleted in step S7.

Step S6
After the process of step S7, only clusters without overlaps remain. The learning result selection unit 4 extracts only clusters with a score equal to or higher than a predetermined score threshold from the above, and sets them as the final learning results. In FIG. 15, for example, if the predetermined score threshold is set to 1, cluster 0 of learning result 1 and cluster 1 of learning result 3 become the final learning results. Here, the predetermined score threshold is set to 1, but it is not limited to this and an appropriate value can be set. If the predetermined score threshold is set low, any cluster can be extracted, and if it is set high, only clusters that appear frequently can be extracted. FIG. 16 shows the final learning result including the clusters selected based on the score matrix of FIG. 15. As shown in FIG. 16, it can be seen that the clusters appropriately express the characteristics of the input data.

As described above, with this configuration, by deleting clusters that are common to multiple learning results, it is possible to select high-scoring clusters from all learning results without selecting similar clusters. This makes it possible to create appropriate clusters even if there are no appropriate clusters among the learning results.

In this embodiment, if either of the judgment values RA and RB is smaller than a predetermined threshold value THr in step S708, it is determined that one of clusters A and B is contained within the other, and the target cluster, cluster B, is determined to be deleted and is deleted, but for example, it may be attached to the cluster as the final learning result and shown to the user as a subordinate class.

Embodiment 3
In the third embodiment, an example will be described in which the cluster deletion process of the second embodiment is simplified and functions as a learning result deletion process. Depending on the learning method, it may be assumed that only partial features of the input data can be learned. FIG. 17 shows an example in which only a part of the clusters can be learned for the input data shown in FIG. 24. Although the probability of the result shown in FIG. 17 being obtained is low, it is possible that it can be obtained as a learning result. If such an inappropriate result in which learning has only progressed partially is included in the options, the first embodiment may not function properly. For example, a case is assumed in which multiple learning results in which the result of FIG. 17 is added to the result of FIG. 4 are targeted, and the score matrix at this time is shown in FIG. 18. In this case, if the final learning result is selected by the process of step S62 of the first embodiment, the inappropriate learning result of FIG. 17 is selected. Therefore, in this embodiment, a process will be described in which an inappropriate learning result is deleted and an appropriate learning result is selected even if an inappropriate result that only partially captures the features of the input data is included.

Figure 19 shows a schematic configuration of an information processing device 300 according to embodiment 3. The information processing device 300 has a configuration in which the deletion processing unit 7 of the information processing device 200 according to embodiment 2 is replaced with a deletion processing unit 8. The deletion processing unit 8 has a configuration in which the cluster deletion unit 76 of the deletion processing unit 7 is replaced with a cluster exclusion unit 86, and a learning result deletion unit 87 is added. The other configurations of the deletion processing unit 8 are similar to those of the deletion processing unit 7, and therefore description thereof is omitted.

Figure 20 shows a flowchart of the learning result deletion process in the information processing device according to embodiment 3. Learning result deletion process step S8 is inserted between steps S5 and S6 in Figure 7, and includes the following steps S801 to S817.

Steps S801 to S808
Steps S801 to S808 are similar to steps S701 to S708 in FIG. 11, and therefore a description thereof will be omitted.

Step S809
The cluster determination unit 75 compares the determination values RA and RB with a predetermined threshold value THr and determines whether or not either of the determination values RA and RB is smaller than the predetermined threshold value THr. If both of the determination values RA and RB are equal to or greater than the predetermined threshold value THr, the cluster determination unit 75 determines that cluster A and cluster B are different clusters, and proceeds to step S811.

Step S810
If it is determined in step S809 that either of the judgment values RA and RB is smaller than a predetermined threshold value THr, the cluster exclusion unit 86 excludes the cluster B with a lower score or the same score (i.e., the target cluster) from the calculation, assuming that one of the clusters A and B is included in or matches the other. Rows corresponding to the clusters excluded from the calculation from the cluster sorting are deleted, and are not selected as the reference cluster or the target cluster for the next process. However, it is possible to use them as the current target cluster for calculating the next target cluster in steps S811 and S812. That is, in this embodiment, when it is described that "a cluster one step below a certain cluster is selected in the order in which the rows corresponding to the clusters excluded from the calculation are deleted from the sorting order," it means that the cluster one step below is selected.

Steps S811 to S812
Steps S811 to S812 are similar to steps S711 to S712 in FIG. 11, and therefore a description thereof will be omitted.

Step S813
The cluster determination unit 75 determines whether or not there is a learning result that does not hold a cluster common to the reference cluster. Here, the same method as the determination method in step S809 can be used to determine whether each cluster is common to the reference cluster. That is, in this step, the cluster excluded from the calculation is determined to be a cluster common to the reference cluster, and the learning result that does not hold the reference cluster or a cluster common to the reference cluster is determined to be deleted. Note that the determination method is not limited to this, and for example, the determination value RA of the reference cluster and the determination value RB of each cluster (target cluster) may be compared using a threshold value THr2 different from the threshold value THr1, and if both the determination values RA and RB are smaller than the threshold value THr2, the target cluster may be determined to be a common cluster. If there is no learning result that does not hold the reference cluster or a cluster common to the reference cluster, the cluster determination unit 75 proceeds to the process of step S815, assuming that there is no learning result to be deleted.

Step S814
If it is determined in step S813 that there is a learning result that does not hold a reference cluster or a cluster common to the reference cluster, the learning result deletion unit 87 deletes the learning result as it should be deleted. Then, the process proceeds to step S815. In addition, rows that correspond to all clusters held by the deleted learning result in the sorted order are excluded from the calculation.

Step S815
The reference cluster selection unit 72 determines whether there is another cluster that remains undeleted and can be selected as a new reference cluster in the row immediately below the current reference cluster in the sorted order. If it is determined that there is no cluster that can be selected as a new reference cluster, the reference cluster selection unit 72 ends the process of step S8.

Step S816
If it is determined in step S815 that there is a cluster that can be selected as a new reference cluster, the reference cluster selection unit 72 determines whether the score of the cluster in the row immediately below the current reference cluster in the sorted order is equal to or greater than a predetermined score threshold THp. If the score of the cluster in the row immediately below the reference cluster is smaller than the predetermined score threshold THp, the reference cluster selection unit 72 ends the process of step S8.

Step S817
If it is determined in step S816 that a cluster that can be selected as a new reference cluster exists, the reference cluster selection unit 72 selects the cluster one cluster below the current reference cluster in the sorted order as the new reference cluster, and then returns the process to step S803.

The cluster deletion process in step S8 will be explained using a concrete example based on six learning results, the five learning results in FIG. 4 and the learning result in FIG. 17. The score matrix is shown in FIG. 18. FIG. 21 shows an example of sorting clusters based on the score matrix in FIG. 18.

Figure 22 shows the progress of the cluster deletion process according to the third embodiment. In the cluster deletion process in step S8, two clusters other than the clusters that have already been deleted are selected in descending order of score, and the deletion process is repeated for the two selected clusters.

Delete Transition A
First, processing is performed between cluster 0 of learning result 1, which has the highest score, and the other clusters. Rows marked with ** in the table of deletion transition 1 in FIG. 22 indicate the reference cluster, and in this case, cluster 0 of learning result 1 is the reference cluster. Rows marked with * are clusters that match the reference cluster, contain the reference cluster, or are determined to be included in the reference cluster. Since there are no learning results that are not marked with ** or *, no learning results are deleted here. Furthermore, rows marked with * are excluded from the sorted table, and therefore are excluded from the options for the deletion process (i.e., options as the reference cluster or target cluster).

Delete Transition B
After the process shown in the deletion transition A is completed, the process shown in the deletion transition B is executed. In the deletion transition B in FIG. 22, the rows corresponding to the clusters excluded from the calculation in the deletion transition A in the sorted order (i.e., the rows marked with * in the deletion transition A) are displayed excluding them. Here, the cluster 1 of the learning result 3, which has the highest score excluding the cluster 0 of the learning result 1, becomes the reference cluster, and the process is executed between the clusters other than the cluster 0 of the learning result 1. That is, the cluster 1 of the learning result 3 in the row marked with ** in the table of the deletion transition B in FIG. 22 is the reference cluster, and the row marked with * is the cluster that is determined to match the reference cluster, include the reference cluster, or be included in the reference cluster. Since none of the clusters in the learning result 6 are marked with ** or *, the learning result 6 is deleted. In addition, the row marked with * is excluded from the sorted table, and therefore is excluded from the options for the deletion process (i.e., the options as the reference cluster and the target cluster).

Delete Transition C
After the process shown in deletion transition B is completed, the process shown in deletion transition C is executed. In deletion transition C in FIG. 22, rows corresponding to the clusters excluded from the calculation in deletion transitions A and B in the sorted order (i.e., rows marked with * in deletion transitions A and B) are excluded from the display. In the state of deletion transition C, there are no unprocessed cluster combinations, so the learning result deletion process ends here. An example of the score matrix after learning result deletion is shown in FIG. 23. It can be seen that learning result 6 marked with an x has been deleted by step S8.

Step S6
When the process of step S8 has been completed for all the sorted clusters, the learning result selection unit 4 selects the final learning result based on the score matrix SM, as in the case of Fig. 4. According to the score matrix in Fig. 23, it can be seen that the learning result selection unit 4 selects learning result 5 as the final learning result, as in the first embodiment.

As described above, with this configuration, by deleting learning results that do not have clusters common to multiple learning results, it is possible to eliminate results that only partially capture the characteristics of the input data and select appropriate learning results.

Other embodiments The present invention is not limited to the above-mentioned embodiments, and can be modified as appropriate without departing from the spirit of the present invention. For example, as for the distance measure indicating the distance between nodes, any distance measure can be applied as the distance measure between nodes, such as Euclidean distance, cosine distance, Manhattan distance, and fractional distance. In general, Euclidean distance is used when the input data is low-dimensional, and cosine distance, Manhattan distance, and fractional distance are used when the input data is high-dimensional.

Although the present invention can be applied to multiple results obtained from various clustering methods, it is preferable to apply it to multiple results from a single method. Clustering methods each have different characteristics, and when the input data has a complex structure, the differences in clustering results due to the characteristics can become significant. By narrowing down the clustering method to one, the differences in the characteristics of the methods of each result are eliminated, and the effects of the present invention can be more effectively achieved.

Although the present invention is effective for the results of general clustering methods such as the k-means method, it is preferable to apply it to the results of a method such as SOINN, which inputs input vectors sequentially and learns the input distribution structure of the input vectors as a network structure in which multiple nodes and multiple edges connecting two nodes are arranged. This is because noise nodes can be suppressed by selecting nodes based on frequency, and the relevance of nodes can also be understood by edges, making it possible to more accurately calculate the distance between nodes and the similarity between clusters, which are important in the present invention.

In the above embodiment, the judgment of the magnitude of two values has been described, but this is merely an example, and the case where the two values are equal in the judgment of the magnitude of two values may be handled as necessary. That is, for the judgment of whether the first value is equal to or greater than the second value or smaller than the second value, either may be adopted as necessary. For the judgment of whether the first value is equal to or less than the second value or larger than the second value, either may be adopted. In other words, when the magnitude judgment of two values is performed to obtain two judgment results, the case where the two values are equal may be included in either of the two judgment results as necessary.

In the above embodiment, the present invention has been described mainly as a hardware configuration, but the present invention is not limited to this, and any process can be realized by having a CPU (Central Processing Unit) execute a computer program. In this case, the computer program can be stored using various types of non-transitory computer readable medium and supplied to the computer. Non-transitory computer readable medium includes various types of tangible storage medium. Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R/Ws, and semiconductor memories (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, and RAMs (random access memories)). The program may also be provided to the computer by various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. The transient computer-readable medium can provide the program to the computer via a wired communication path such as an electric wire or optical fiber, or via a wireless communication path.

Some or all of the above embodiments may be described as follows, but are not limited to the following:

(Supplementary Note 1) In an information processing device that learns the input distribution structure of multiple input vectors as a network structure including multiple nodes described by multidimensional vectors, the information processing device has: a distance matrix calculation unit that calculates a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results, for all combinations of two learning results included in multiple learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions; an evaluation index calculation unit that obtains the minimum value of each column and each row for each distance matrix and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition; a score calculation unit that calculates a score indicating the degree of similarity of each cluster of each learning result to clusters of other learning results other than the learning result of interest, based on the evaluation index calculated for all distance matrices; and a learning result selection unit that selects a final learning result from multiple previous learning results based on the score of each cluster of each learning result calculated by the score calculation unit.

(Appendix 2) The information processing device described in Appendix 1, wherein the score calculation unit generates an evaluation index matrix composed of evaluation indexes for all combinations between two clusters selected one from each of two learning results based on the evaluation indexes calculated for all distance matrices, calculates a score indicating whether a cluster included in one learning result is similar to a cluster of another learning result based on the evaluation indexes included in the evaluation index matrix, and generates a score matrix including scores for clusters belonging to all learning results, and the learning result selection unit selects a final learning result from the multiple learning results based on the score matrix.

(Appendix 3) The information processing device described in Appendix 2, wherein the evaluation index calculation unit calculates the average distance between nodes belonging to each of the two clusters, calculates the average of the two average distances calculated, and calculates an evaluation index for each distance matrix based on the number of the minimum values in each column and each row of each distance matrix that are equal to or greater than the average.

(Appendix 4) The information processing device described in Appendix 3, wherein the evaluation index is the number of the minimum values in each column and each row of each distance matrix that are greater than or equal to the average value divided by the total number of nodes belonging to the two clusters.

(Appendix 5) An information processing device as described in Appendix 2 or 3, wherein the score calculation unit compares each evaluation index included in the evaluation index matrix with a threshold value, assigns a score based on the comparison result, and calculates a total score for each cluster.

(Appendix 6) The information processing device described in Appendix 5, wherein the learning result selection unit obtains the minimum value among the total scores of the clusters belonging to each learning result, and selects the learning result corresponding to the maximum value among the obtained minimum values.

(Appendix 7) An information processing device described in any one of Appendices 2 to 5, wherein the learning result selection unit selects multiple non-overlapping clusters from the clusters of the multiple learning results based on the score matrix, and selects the final learning result composed of the selected clusters.

(Appendix 8) An information processing device described in any one of Appendices 2 to 5, wherein the learning result selection unit selects one learning result from the multiple learning results as the final learning result based on the score matrix.

(Appendix 9) An information processing device as described in any one of Appendices 3 to 7, further comprising a deletion processing unit that performs processing to delete clusters from each learning result, the deletion processing unit comprising a score sorting unit that sorts clusters belonging to each learning result in order of their scores, a reference cluster selection unit that selects the cluster with the highest score as a reference cluster, a target cluster selection unit that selects from the sorted clusters the cluster with the highest score among clusters with scores lower than the reference cluster as a target cluster, a judgment value calculation unit that calculates a judgment value used to determine a cluster based on the reference cluster and the target cluster, a cluster judgment unit that selects a cluster to be deleted based on the judgment value, and a cluster deletion unit that deletes the cluster to be deleted.

(Appendix 10) The distance matrix calculation unit calculates the distance matrix for the reference cluster and the target cluster, the distance matrix having a number of rows equal to the number of nodes in the reference cluster and a number of columns equal to the number of nodes in the target cluster, and the evaluation index calculation unit obtains the minimum value of each column and each row of the distance matrix calculated for the reference cluster and the target cluster, calculates the average distance between the nodes belonging to the reference cluster and the target cluster, calculates the average value of the two calculated average distances, and calculates the minimum value of each row of the distance matrix calculated for the reference cluster and the target cluster. The information processing device of claim 9, wherein the first value is the number of values equal to or greater than the average value, and the second value is the number of values among the minimum values in each column of the distance matrix that are equal to or greater than the average value, the judgment value calculation unit calculates a first judgment value by dividing the first value by the number of nodes in the reference cluster, and a second judgment value by dividing the second value by the number of nodes in the reference cluster, and if either the first or second judgment value is smaller than a second threshold, the cluster judgment unit selects the target cluster as a cluster to be deleted, and the cluster deletion unit deletes the cluster to be deleted.

(Appendix 11) An information processing device as described in Appendix 10, in which, when the deletion processing unit deletes the cluster to be deleted, or when the cluster determination unit determines that the first and second determination values are greater than a second threshold, the target cluster selection unit selects the cluster having the next lowest score after the target cluster as a new target cluster, and the deletion processing unit performs a process of deleting a cluster based on the reference cluster and the new target cluster.

(Appendix 12) An information processing device as described in Appendix 11, in which, if there is no cluster with the next lowest score after the target cluster, the reference cluster selection unit selects, in addition to the reference cluster, the cluster with the next lowest score after the reference cluster as a new reference cluster, and the deletion processing unit performs processing to delete a cluster based on the new reference cluster.

(Appendix 13) If there is no cluster with the next lowest score after the reference cluster, the deletion processing unit terminates the process of deleting clusters, in the information processing device described in Appendix 12.

(Appendix 14) An information processing device as described in any one of Appendices 3 to 7, further comprising a deletion processing unit that performs processing to delete a learning result selected from the multiple learnings, the deletion processing unit comprising a score sorting unit that sorts clusters belonging to each learning result in order of their scores, a reference cluster selection unit that selects the cluster with the highest score as a reference cluster, a target cluster selection unit that selects from the sorted clusters the cluster with the highest score among clusters with scores lower than the reference cluster as a target cluster, a judgment value calculation unit that calculates a judgment value used for cluster judgment based on the reference cluster and the target cluster, a cluster judgment unit that selects clusters based on the judgment value, a cluster exclusion unit that performs processing to exclude clusters based on the results of the judgment, and a learning result deletion unit that deletes learning results that do not have a cluster in common with the reference cluster.

(Appendix 15) The distance matrix calculation unit calculates the distance matrix for the reference cluster and the target cluster, the distance matrix having a number of rows equal to the number of nodes in the reference cluster and a number of columns equal to the number of nodes in the target cluster, and the evaluation index calculation unit obtains the minimum value of each column and each row of the distance matrix calculated for the reference cluster and the target cluster, calculates the average distance between the nodes belonging to the reference cluster and the target cluster, calculates the average value of the two average distances calculated, and calculates a first value which is the number of the minimum values of each row of the distance matrix calculated for the reference cluster and the target cluster that are equal to or greater than the average value, and a second value which is the number of the minimum values of each column of the distance matrix that are equal to or greater than the average value. the judgment value calculation unit calculates a first judgment value by dividing the first value by the number of nodes of the reference cluster and a second judgment value by dividing the second value by the number of nodes of the reference cluster, the cluster judgment unit determines that the target cluster is a cluster different from the reference cluster if either the first or second judgment value is smaller than a second threshold, and determines that the target cluster is a cluster in common with the reference cluster if the first and second judgment values are larger than the second threshold, and when it is determined that the target cluster is a cluster different from the reference cluster, the cluster exclusion unit excludes the target cluster from a learning result to which the target cluster belongs.

(Appendix 16) An information processing device as described in Appendix 15, in which, when the cluster exclusion unit excludes the target cluster, or when the cluster determination unit determines that the target cluster is a cluster common to the reference cluster, the target cluster selection unit selects the cluster having the next lowest score after the target cluster as a new target cluster, and the deletion processing unit performs a process of deleting the learning result based on the reference cluster and the new target cluster.

(Appendix 17) In an information processing device as described in Appendix 16, if there is no cluster with the next lowest score after the target cluster, the learning result deletion unit deletes from the multiple learning results any learning result that does not have a cluster in common with the reference cluster.

(Appendix 18) An information processing device as described in Appendix 16, in which when the learning result deletion unit deletes the learning result, or when there is no cluster with the next lowest score after the target cluster and there is no cluster that does not have a cluster in common with the reference cluster, the reference cluster selection unit selects, other than the reference cluster, a cluster with the next lowest score after the reference cluster if its score is greater than a predetermined value, and the deletion processing unit performs processing to delete the learning result based on the new reference cluster.

(Appendix 19) An information processing device as described in Appendix 18, wherein if the score of the cluster with the next lowest score after the reference cluster is smaller than the predetermined value, or if there is no cluster with the next lowest score after the reference cluster, the deletion processing unit terminates the process of deleting the learning results.

(Appendix 20) An information processing method for learning an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, in which a distance matrix calculation unit calculates a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results, for all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions; an evaluation index calculation unit obtains the minimum value of each column and each row for each distance matrix, and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition; a score calculation unit calculates a score for each cluster of each learning result indicating the degree of similarity between the cluster of each learning result and clusters of other learning results other than the learning result of interest, based on the evaluation index calculated for all distance matrices; and a learning result selection unit selects a final learning result from the plurality of learning results in the previous period, based on the score of each cluster of each learning result calculated by the score calculation unit.

(Supplementary Note 21) A program that causes a computer to execute a process of learning an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, the program causing the computer to execute the following processes: for all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained as a result of learning the same input vector under different conditions, a process of calculating a distance matrix indicating a distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results; a process of obtaining a minimum value for each column and each row of each distance matrix, and calculating an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition; a process of calculating a score for each cluster of each learning result that indicates a degree of similarity to clusters of other learning results other than the learning result of interest, based on the evaluation indexes calculated for all distance matrices; and a process of selecting a final learning result from the plurality of previous learning results based on the calculated score for each cluster of each learning result.

The present invention has been described above with reference to the embodiment, but the present invention is not limited to the above. Various modifications that can be understood by a person skilled in the art can be made to the configuration and details of the present invention within the scope of the invention.

This application claims priority based on Japanese Patent Application No. 2020-218716, filed on December 28, 2020, the disclosure of which is incorporated herein in its entirety.

REFERENCE SIGNS LIST 1 Distance matrix calculation unit 2 Evaluation index calculation unit 2A Calculation unit 2B Evaluation index matrix generation unit 3 Score calculation unit 3A Calculation unit 3B Score matrix generation unit 4 Learning result selection unit 7, 8 Deletion processing unit 10 Computer 11 CPU
12 ROM
13 RAM
REFERENCE SIGNS LIST 14 bus 15 input/output interface 16 input unit 17 output unit 18 storage unit 19 communication unit 20 drive 20A magnetic disk 20B optical disk 20C flexible disk 20D semiconductor memory 71 score sorting unit 72 reference cluster selection unit 73 target cluster selection unit 74 judgment value calculation unit 75 cluster judgment unit 76 cluster deletion unit 86 cluster deletion unit 87 learning result deletion unit 100, 200, 300 information processing device

Claims (15)

An information processing device that learns an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector,
a distance matrix calculation unit that calculates a distance matrix indicating a distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results in all combinations of two learning results included in a plurality of learning results having different input distribution structures obtained by learning the same input vector under different conditions;
an evaluation index calculation unit that obtains the minimum value of each column and each row for each distance matrix, and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition;
a score calculation unit that calculates a score indicating a degree of similarity between each cluster of each learning result and clusters of other learning results other than the learning result of interest, based on the evaluation indexes calculated for all distance matrices;
a learning result selection unit that selects a final learning result from a plurality of learning results in the previous period based on the score of each cluster of each learning result calculated by the score calculation unit.
Information processing device. The score calculation unit,
generating an evaluation index matrix composed of evaluation indexes for all combinations between two clusters selected one by one from each of the two learning results based on the evaluation indexes calculated for all distance matrices;
calculating a score indicating whether a cluster included in one learning result is similar to a cluster of another learning result based on the evaluation indexes included in the evaluation index matrix;
Generate a score matrix including scores for the clusters belonging to all the learning results;
the learning result selection unit selects a final learning result from the plurality of learning results based on the score matrix.
The information processing device according to claim 1 . The evaluation index calculation unit is
calculating an average distance between nodes belonging to each of the two clusters, and calculating an average value of the two calculated average distances;
Calculating an evaluation index for each distance matrix based on the number of minimum values equal to or greater than the average value among the minimum values in each column and each row of each distance matrix;
The information processing device according to claim 2 . The evaluation index is a value obtained by dividing the number of the minimum values in each column and each row of each distance matrix that are equal to or greater than the average value by the total number of nodes belonging to the two clusters.
The information processing device according to claim 3 . the score calculation unit compares each evaluation index included in the evaluation index matrix with a threshold, assigns a score based on a comparison result, and calculates a total score for each cluster.
4. The information processing device according to claim 2 or 3. the learning result selection unit obtains a minimum value among the total scores of the clusters belonging to each learning result, and selects the learning result corresponding to the maximum value among the obtained minimum values.
The information processing device according to claim 5 . the learning result selection unit selects a plurality of non-overlapping clusters from the clusters of the plurality of learning results based on the score matrix, and selects the final learning result constituted by the selected clusters.
The information processing device according to claim 2 . the learning result selection unit selects one learning result from the plurality of learning results as the final learning result based on the score matrix;
The information processing device according to claim 2 . A deletion processing unit that performs a process of deleting clusters from each learning result is further provided,
The deletion processing unit:
a score sorting unit that sorts clusters belonging to each learning result in order of the scores;
a reference cluster selection unit that selects the cluster with the highest score as a reference cluster;
a target cluster selection unit that selects, from the sorted clusters, a cluster having a highest score among clusters having scores lower than the reference cluster, as a target cluster;
a judgment value calculation unit that calculates a judgment value used for cluster judgment based on the reference cluster and the target cluster;
a cluster determination unit for selecting a cluster to be deleted based on the determination value;
A cluster deletion unit that deletes the deletion target cluster.
The information processing device according to claim 3 . the distance matrix calculation unit calculates, for the reference cluster and the target cluster, the distance matrix having a number of rows equal to the number of nodes of the reference cluster and a number of columns equal to the number of nodes of the target cluster;
The evaluation index calculation unit is
Obtaining the minimum values of each column and each row of the distance matrix calculated for the reference cluster and the target cluster;
calculating an average distance between nodes belonging to the reference cluster and the target cluster, and calculating an average value of the two calculated average distances;
counting a first value which is the number of the minimum values in each row of the distance matrix calculated for the reference cluster and the target cluster that are equal to or greater than the average value, and a second value which is the number of the minimum values in each column of the distance matrix that are equal to or greater than the average value;
the determination value calculation unit calculates a first determination value by dividing the first value by the number of nodes in the reference cluster and a second determination value by dividing the second value by the number of nodes in the reference cluster;
the cluster determination unit selects the target cluster as a deletion target cluster when either the first or second determination value is smaller than a second threshold value;
The cluster deletion unit deletes the cluster to be deleted.
The information processing device according to claim 9. When the deletion processing unit deletes the deletion target cluster, or when the cluster determination unit determines that the first and second determination values are greater than a second threshold value,
the target cluster selection unit selects a cluster having a second lowest score after the target cluster as a new target cluster;
The deletion processing unit performs a process of deleting a cluster based on the reference cluster and the new target cluster.
The information processing device according to claim 10. When there is no cluster having the next lowest score after the target cluster, the reference cluster selection unit selects, in addition to the reference cluster, a cluster having the next lowest score after the reference cluster as a new reference cluster;
The deletion processing unit performs a process of deleting clusters based on the new reference cluster.
The information processing device according to claim 11. If there is no cluster having the next lowest score after the reference cluster, the deletion processing unit ends the process of deleting clusters.
The information processing device according to claim 12. An information processing method for learning an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by multidimensional vectors, comprising:
a distance matrix calculation unit calculates, for all combinations of two clusters selected one from each of the two learning results, the two learning results having different input distribution structures obtained by learning the same input vector under different conditions, a distance matrix indicating distances between a node belonging to one cluster and a node belonging to the other cluster;
an evaluation index calculation unit obtains a minimum value for each column and each row of each distance matrix, and calculates an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition;
a score calculation unit calculates, for each cluster of each learning result, a score indicating a degree of similarity to clusters of learning results other than the learning result of interest, based on the evaluation indexes calculated for all distance matrices;
a learning result selection unit selects a final learning result from the plurality of learning results in the previous period based on the score of each cluster of each learning result calculated by the score calculation unit;
Information processing methods. A program for causing a computer to execute a process of learning an input distribution structure of a plurality of input vectors as a network structure including a plurality of nodes described by a multidimensional vector, the program comprising:
A process of calculating a distance matrix indicating the distance between a node belonging to one cluster and a node belonging to the other cluster for all combinations between two clusters selected one from each of the two learning results in all combinations of two learning results having different input distribution structures obtained by learning the same input vector under different conditions;
A process of obtaining the minimum value of each column and each row for each distance matrix, and calculating an evaluation index for each distance matrix based on the number of the obtained minimum values that satisfy a predetermined condition;
A process of calculating a score indicating the degree of similarity of each cluster of each learning result with clusters of other learning results other than the learning result of interest based on the evaluation index calculated for all distance matrices;
selecting a final learning result from the plurality of learning results in the previous period based on the calculated score of each cluster of each learning result;
program.

Images