Information processing method and information processing system

Abstract

The present technology relates to an information processing method and an information processing system that can improve the prediction accuracy of a prediction model constructed through federated learning. The information processing method according to the present technology includes: applying soft clustering to a plurality of pieces of data used in training a prediction model and extracting a feature amount of each category related to a prediction item of the prediction model; and determining, based on the feature amount, an optimal number of categories for performing training of the prediction model. For example, the present technology can be applied to a training system that constructs a prediction model capable of predicting a future cell state through federated learning.

Description

Technical Field

[0001] This technology relates to information processing methods and systems, and more specifically to information processing methods and systems that can improve the prediction accuracy of prediction models constructed through joint learning. Background Technology

[0002] In immunotherapy where cells are used as therapeutic agents, it is necessary to homogenize the cell state to a certain extent during the culture process in order to achieve a stable response rate. Because the initial cell state depends on the patient, collection date, treatment history, etc., even when cells are cultured under the same conditions, the final cell configuration, cell state, and cell count do not become constant. Therefore, labor and resources are required to achieve cell state homogenization.

[0003] A predictive model needs to be built that can predict future cell states based on initial cell states and culture conditions. When building this predictive model, learning must be performed while ensuring the confidentiality of patient data.

[0004] Therefore, it is believed that by performing joint learning using patient data at each medical institution and updating the predictive model at the server, it is possible to obtain a predictive model that can predict future cell states with high accuracy while ensuring the confidentiality of patient data.

[0005] In joint learning, optimizing the number of categories into which data is classified and the category labels assigned to the corresponding data fragments leads to improvements in the prediction accuracy of the constructed predictive model. For example, Patent Document 1 describes a technique for generating a model with high learning accuracy by performing machine learning while merging similar categories into one category.

[0006] Reference List

[0007] Patent documents

[0008] Patent Document 1: JP 2018-41300 A Summary of the Invention

[0009] The problem to be solved by the present invention

[0010] In the elbow method, which is typically used to determine the number of categories, it can be difficult to identify points where the graph bends and to determine the optimal number of categories. Furthermore, the elbow method requires assigning a category label to each data point separately.

[0011] When the technology described in Patent Document 1 is applied to joint learning, there is a possibility that the number of categories during the learning process will not match between the medical institutions performing the learning, since the optimal number of categories is obtained in the machine learning system.

[0012] This technique was developed in light of this situation and is able to improve the prediction accuracy of prediction models built through joint learning.

[0013] Solution to the problem

[0014] The information processing method according to this technology includes: applying soft clustering to multiple data sets used for learning a prediction model to extract a feature quantity for each category of the prediction terms relative to the prediction model; and determining the optimal number of categories for learning the prediction model based on the feature quantity.

[0015] The information processing system according to this technology includes: a soft clustering unit that applies soft clustering to multiple data sets used for learning a prediction model to extract a feature quantity for each category of the prediction terms of the prediction model; and an optimal number of categories determination unit that determines the optimal number of categories for learning the prediction model based on the feature quantities.

[0016] In one aspect of this technique, soft clustering is applied to multiple datasets for learning a predictive model, whereby the feature quantity of each category relative to the predictive model's predictions is extracted, and based on the feature quantity, the optimal number of categories for learning the predictive model is determined. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating an embodiment of the configuration of a learning system according to one implementation of the present technology.

[0018] Figure 2 It is used to explain by Figure 1 The first diagram shows the process of joint learning executed by the learning system.

[0019] Figure 3 It is used to explain by Figure 1 The second diagram shows the process of joint learning executed by the learning system.

[0020] Figure 4 This is a diagram illustrating an embodiment of a process including soft clustering performed by a server.

[0021] Figure 5 This is a diagram illustrating an example of a matrix that applies soft clustering.

[0022] Figure 6 This is a diagram illustrating an embodiment of the record entries corresponding to each column of the initial cell state matrix.

[0023] Figure 7 This is a diagram illustrating an embodiment of the record items corresponding to each column of the stimulus condition matrix.

[0024] Figure 8This is a diagram illustrating an embodiment of the record entries corresponding to each column of the cell state matrix after stimulus application.

[0025] Figure 9 This is a diagram used to illustrate soft clustering in jNMF.

[0026] Figure 10 This is a diagram illustrating an embodiment of the initial cell state matrix and the stimulus condition matrix.

[0027] Figure 11 This is a diagram illustrating an example of applying a cell state matrix after stimulation.

[0028] Figure 12 This is a diagram illustrating an embodiment of the basis matrix and coefficient matrix.

[0029] Figure 13 This is a diagram illustrating an embodiment of the feature quantities used for prediction terms.

[0030] Figure 14 This is a diagram illustrating an example of the temporal variation of characteristic quantities for each category of CD8+ cell counts.

[0031] Figure 15 This is a diagram illustrating an example of the correlation coefficients of the time variations of characteristic quantities in seven categories.

[0032] Figure 16 This is a diagram illustrating an example of the temporal variation of characteristic quantities for each category after the aggregation of CD8+ cell counts.

[0033] Figure 17 This is a graph illustrating an example of the correlation coefficients of the time-varying characteristics between categories after merging.

[0034] Figure 18 This is a diagram used to explain the details of the initial category labeling process.

[0035] Figure 19 This is a diagram illustrating examples of the prediction accuracy of the prediction model when the number of categories is four and when the number of categories is seven.

[0036] Figure 20 This is a block diagram illustrating an embodiment of the server's functional configuration.

[0037] Figure 21 This is a block diagram illustrating an embodiment of the functional configuration of an information processing device.

[0038] Figure 22 It is a flowchart used to illustrate the processes performed by the server.

[0039] Figure 23 It is used to explain in Figure 22The flowchart for determining the optimal number of categories in step S2.

[0040] Figure 24 It is used to explain in Figure 22 The flowchart shows the initial category labeling process performed in step S3.

[0041] Figure 25 It is a flowchart used to illustrate the processes performed by the information processing equipment.

[0042] Figure 26 It is used to explain in Figure 25 The flowchart shows the category merging process performed in step S62.

[0043] Figure 27 It is used to explain in Figure 25 The flowchart shows the initial category labeling process performed in step S63.

[0044] Figure 28 This is a diagram illustrating another embodiment of a process including soft clustering performed by a server.

[0045] Figure 29 It is a diagram used to explain the details of initial data labeling and group labeling processes.

[0046] Figure 30 This is a diagram illustrating an example of the prediction accuracy of a prediction model when group labels are assigned based on clustering results and when random group labels are assigned.

[0047] Figure 31 This is a block diagram illustrating an embodiment of a computer's hardware configuration. Detailed Implementation

[0048] The modes used to implement this technology will be described below.

[0049] The descriptions will be given in the following order.

[0050] 1. Overview of Learning Systems

[0051] 2. Configuration and operation of each device

[0052] 3. Variations

[0053] <1. Overview of Learning Systems>

[0054] <Configuration of the learning system>

[0055] Figure 1 This is a diagram illustrating an embodiment of the configuration of a learning system according to one implementation of the present technology.

[0056] Figure 1The learning system is an information processing system that builds a predictive model through joint learning. This predictive model can predict future cell states while ensuring the confidentiality of patient data. In this joint learning, learning using patient data is performed at each medical institution and the predictive model is updated by the server.

[0057] In the following description, it is assumed that cell state includes cell state, configuration, count, etc.

[0058] Figure 1 The learning system is configured such that server 101 and information processing devices 102 located at medical institutions 1 to n are connected via a network such as the Internet.

[0059] Server 101 (first information processing device) consists of, for example, a single computer or multiple computers operating collaboratively. Server 101 performs learning using a dataset including multiple pre-collected data points and generates a predictive model. Additionally, server 101 aggregates the predictive model retrained by the information processing device 102 of the medical institution.

[0060] For example, information processing device 102 (second information processing device) consists of a single computer or multiple computers operating collaboratively. Information processing device 102 uses a dataset including multiple patient data collected at each medical institution to retrain the predictive model generated by server 101.

[0061] Next, we will see Figure 2 and Figure 3 Description by Figure 1 The learning system executes a collaborative learning process.

[0062] First, during joint learning, server 101 applies soft clustering to the data included in the pre-collected dataset to determine the optimal number of classes (optimal number of classes) for performing joint learning and assigns an initial class label to each data point. The class label indicates which class the data point with the assigned class label is classified into.

[0063] Server 101 uses the data with assigned initial class labels and the optimal number of classes as input to perform learning, while updating the class labels to minimize the difference between predicted and true values, ultimately generating a prediction model corresponding to the optimal number of classes. For example, in Figure 2 As shown in the upper part, server 101 generates, for example, four prediction models A0, B0, C0, and D0.

[0064] Next, as Figure 2As shown by arrow #1, server 101 sends four prediction models A0, B0, C0, and D0, along with information indicating the optimal number of categories, to medical institutions 1 through n.

[0065] Next, as Figure 3 As shown in the lower left corner, the information processing device 102 of medical institution 1 performs soft clustering on the local data (patient data) included in the dataset 1 obtained by medical institution 1, and merges the categories until the number of categories becomes the optimal number of categories determined by server 101. Thereafter, the information processing device 102 of medical institution 1 assigns initial category labels to each patient data point based on the results of the soft clustering.

[0066] The information processing device 102 of medical institution 1 uses dataset 1, which includes patient data with initial category labels already assigned, to retrain four prediction models A0, B0, C0, and D0 sent from server 101. Specifically, the information processing device 102 of medical institution 1 retrains while updating the category labels, thereby reducing the difference between the predicted values ​​and the true values.

[0067] Next, as Figure 3 As indicated by arrow #21-1, the information processing device 102 of medical institution 1 sends the retrained prediction models A1, B1, C1, and D1 (model parameters) to server 101.

[0068] In addition, such as Figure 3 As shown in the lower center, the information processing device 102 of medical institution 2 performs soft clustering on the local data (patient data) included in the dataset 2 obtained by medical institution 2, and merges the categories until the number of categories becomes the optimal number of categories determined by server 101. Thereafter, the information processing device 102 of medical institution 2 assigns initial category labels to each patient data point based on the results of the soft clustering.

[0069] The information processing device 102 of medical institution 2 retrains four prediction models A0, B0, C0, and D0 sent from server 101 using dataset 2, which includes patient data with initial category labels already assigned. Specifically, the information processing device 102 of medical institution 1 retrains while updating the category labels, thereby reducing the difference between the predicted values ​​and the true values.

[0070] like Figure 3 As shown by arrow #21-2, the information processing device 102 of medical institution 2 sends the retrained prediction models A2, B2, C2, D2 (model parameters) to server 101.

[0071] In addition, such as Figure 3As shown in the lower right corner, the information processing device 102 of medical institution n performs soft clustering on the local data (patient data) included in the dataset n obtained by medical institution n, and merges the categories until the number of categories becomes the optimal number of categories determined by server 101. Afterwards, the information processing device 102 of medical institution n assigns initial category labels to each patient data point based on the results of the soft clustering.

[0072] The information processing device 102 of medical institution 1 uses a dataset n containing patient data with initially assigned category labels to retrain four prediction models A0, B0, C0, and D0 sent from server 101. Specifically, the information processing device 102 of medical institution 1 retrains while updating the category labels, thereby reducing the difference between the predicted values ​​and the true values.

[0073] like Figure 3 As indicated by arrow #21-n, the information processing device 102 of medical institution n sends the retrained prediction model A to server 101. n B n C n D n (Model parameters).

[0074] Server 101 receives predictive models sent from medical institutions. For example... Figure 3 As shown in the upper part, server 101 uses aggregated prediction models A1 to A n Generate a new prediction model A, and aggregate prediction models B1 to B. n A new prediction model B is generated. Server 101 aggregates prediction models C1 to C2. n To generate a new prediction model C, and to aggregate prediction models D1 to D2 n To generate a new prediction model D.

[0075] Next, as Figure 3 As indicated by arrow #22, server 101 sends four prediction models A, B, C, and D to information processing devices 102 in medical institutions 1 to n, respectively. Each information processing device 102 in each medical institution retrains the prediction models A, B, C, and D sent from server 101 and sends the retrained prediction models back to server 101. Server 101 aggregates the prediction models sent from the information processing devices 102 in the medical institutions.

[0076] By repeating this process, even without aggregating patient data from outside the healthcare institution (e.g., within server 101), the large amount of patient data collected at the healthcare institution is used to train the predictive model, and the predictive model is updated to have high predictive accuracy.

[0077] Figure 4 This is a diagram illustrating an embodiment of a soft clustering process performed by server 101.

[0078] Before generating the predictive model, server 101 first performs soft clustering on the pre-collected data, such as... Figure 4 As shown in #11, soft clustering is used to extract the feature values ​​of the corresponding categories for the predictions of the prediction model (e.g., CD8+ cell count) and the contribution of the corresponding data segments to the corresponding categories.

[0079] Next, server 101 determines the optimal number of classes in the joint learning process, such as... Figure 4 As indicated by #12 in [reference needed]. Specifically, server 101 determines the optimal number of categories by merging multiple categories into one category based on the similarity of the feature quantities extracted by soft clustering.

[0080] Next, server 101 assigns an initial category label to each piece of data, such as... Figure 4 As indicated in #13. Specifically, after considering the categories merged when determining the optimal number of categories as a single category, server 101 determines the category labels to be assigned to each data point based on the contribution of each data point to the categories extracted by soft clustering.

[0081] <Details of each process>

[0082] Here, for reference Figures 5 to 19 The details of the corresponding processing performed by server 101 before generating the predictive model will be described. The following describes the processing flow in the case where the aim is to construct a predictive model capable of predicting the CD8+ cell count on the last day of culture using joint learning with cell culture data.

[0083] Cell culture data is a summary of time-series data, including various measurements, for each sample. Cell culture data is created by, for example, culturing cells for several days while simultaneously varying conditions for each sample (such as combinations of antibody and cytokine stimulation and stimulation doses) and periodically recording stimulation doses, flow cytometry (FCM) measurements, etc., during the culture period (on days 0, 4, 7, 11, and 14).

[0084] Soft clustering

[0085] Here, embodiments will be described in which soft clustering, such as joint nonnegative tensor decomposition (jNMF), is applied to cell culture data.

[0086] When applying jNMF to cell culture data, multiple cell culture datasets are first divided into, for example, three matrices: an initial cell state matrix, a stimulus condition matrix, and a post-stimulus application cell state matrix, such as... Figure 5 As shown.

[0087] In the initial cell state matrix, measurements such as FCM are recorded on day 0 of culture (the time point before stimulation). In the stimulation condition matrix, the applied stimulation values ​​and the cumulative applied stimulation values ​​are recorded. In the post-stimulation cell state matrix, measurements of FCM are recorded on days 4, 7, 11, and 14 of culture (the time points after stimulation).

[0088] In each of the three matrices, each row indicates the data for each sample, and each column indicates the data for each record. In the stimulus condition matrix and the post-stimulation application cell state matrix, which include data at four time points, the data for the same record at four time points are arranged in the row direction, as shown below. Figure 5 As indicated by the marker in the text.

[0089] In each of the three matrices, for the data on CD8+ cell counts as a prediction term, preprocessing is performed, where the values ​​are increased tenfold so that their contribution to clustering in jNMF becomes higher than their contribution to the data on other record terms.

[0090] Figure 6 This is a diagram illustrating an embodiment of the record entries corresponding to each column of the initial cell state matrix.

[0091] For example, in the first column of the initial cell state matrix, it is recorded whether each sample is from a patient who has already received chemotherapy.

[0092] Figure 7 This is a diagram illustrating an example of the record items corresponding to each column of the stimulus condition matrix.

[0093] For example, the concentration of antibody-stimulated OKT3 in the culture medium on days 4, 7, 11, and 14 of culture was recorded from the first to the fourth column of the stimulation condition matrix.

[0094] Figure 8 This is a diagram illustrating an example of the record entries corresponding to each column of the cell state matrix after stimulus application.

[0095] For example, the ratio of lymphocytes to total white blood cells was recorded from the first to the fourth column of the cell state matrix after stimulation administration on days 4, 7, 11, and 14 of culture.

[0096] Figure 9 This is a diagram used to illustrate soft clustering in jNMF.

[0097] In jNMF, such as Figure 9 As shown, the data matrices X1 to XM, each with n rows, are decomposed into the basis matrix W and the coefficient matrices H1 to HM.

[0098] Data matrix X1 is a matrix with n rows and p1 columns, data matrix X2 is a matrix with n rows and p2 columns, and data matrix XM is a matrix with n rows and p1 columns. M The matrix consists of n rows and k columns. The basis matrix W is a matrix indicating the clustering results for each sample, and it has n rows and k columns. The coefficient matrix H1 is a matrix representing the extracted features of data matrix X1, and it has k rows and p1 columns. The coefficient matrix H2 is a matrix representing the extracted features of data matrix X2, and it has k rows and p2 columns. The coefficient matrix HM is a matrix representing the extracted features of data matrix XM, and it has k rows and p1 columns. M A matrix of columns.

[0099] In jNMF, the basis matrix W and coefficient matrices H1 to HM are computed such that the sum of the square norms of the data matrix Xm and the reconstruction matrix WHm, as expressed by the following formula (1), becomes the minimum.

[0100] [Mathematical Expression 1]

[0101] satisfy

[0102] For example, jNMF is considered to be applied to data matrices X1 to X3, where the initial cell state matrix is ​​shown in Figure 10 The left side is used as the data matrix X1, and the stimulus condition matrix is ​​shown in... Figure 10 The right side serves as data matrix X2, where the cell state matrix after stimulus application is displayed. Figure 11 The data matrix X1 to X3 is set as data matrix X3, and the number of categories (number of bases) is set to 7. By applying jNMF, the data matrices X1 to X3 are decomposed into... Figure 12 The basis matrix W and coefficient matrices H1 to H3 are shown in the figure.

[0103] In the basis matrix W, the degree of contribution of each sample to each class is represented, and in the coefficient matrices H1 to H3, the degree of contribution of each record item to each class is represented.

[0104] Determining the optimal number of categories

[0105] In determining the optimal number of categories, firstly, the CD8+ cell count, as a predictor, is extracted from the coefficient matrix H3, which represents the feature quantity of the cell state matrix (data matrix X3) after stimulus application. Figure 13In this embodiment, a matrix from columns 53 to 56 is extracted, as indicated by being enclosed in a gray frame. In the extracted matrix, each row indicates the feature quantity for each category, and each column indicates the feature quantity for each day. The extracted matrix also represents the temporal variation (trend) of the feature quantities for each category relative to CD8+ cell counts.

[0106] Figure 14 This is a diagram illustrating an example of the time-varying characteristics of a corresponding category of CD8+ cell counts. Figure 14 In the diagram, the horizontal axis represents days, and the vertical axis represents characteristic quantities.

[0107] exist Figure 14 In the figure, the temporal changes of characteristic quantities from day 4 to day 14 of culture are shown for categories 1 to 7 respectively.

[0108] Next, as Figure 15 As shown, the correlation coefficients of the time variations of the feature quantities are calculated across the seven categories 1 to 7. Next, combinations of categories 1 to 7 with correlation coefficients equal to or greater than a threshold are merged into a single category.

[0109] For example, when the threshold is 0.7, in Figure 15 In this embodiment, since the correlation coefficient between category 1 and category 5 is 0.9835, categories 1 and 5 are merged into one category. Furthermore, since the correlation coefficient between category 2 and category 4 is 0.9933, the correlation coefficient between category 4 and category 7 is 0.8899, and the correlation coefficient between category 7 and category 2 is 0.9362, categories 2, 4, and 7 are merged into one category. In the following text, the category in which categories 1 and 5 are merged is also referred to as category 1 / 5, and the category in which categories 2, 4, and 7 are merged is also referred to as category 2 / 4 / 7.

[0110] When merging categories, the feature values ​​of the corresponding categories that are merged into one category are averaged for each day.

[0111] Figure 16 This is a diagram illustrating an example of the temporal variation of characteristic quantities of the corresponding category after CD8+ cell count consolidation. Figure 16 In the diagram, the horizontal axis represents the number of days, and the vertical axis represents the characteristic quantity.

[0112] exist Figure 16 The table shows the temporal changes of feature values ​​from day 4 to day 14 for categories 1 / 5, 2 / 4 / 7, 3, and 7, respectively.

[0113] As described above, after performing a series of processes including calculating the correlation coefficient between categories, comparing the correlation coefficient with a threshold, merging categories with correlation coefficients equal to or greater than the threshold, and averaging the feature values ​​of the merged categories, this series of processes is repeated until a combination of categories with no correlation coefficients equal to or greater than the threshold remains.

[0114] For example, such as Figure 17 As shown, the correlation coefficients of the time series changes of feature quantities are calculated among categories 1 / 5, 2 / 4 / 7, 3, and 7. Figure 17 In the embodiment, since there are no combinations with a correlation coefficient of 0.7 or greater, the series of processes ends.

[0115] The number of categories in a combination that does not have a correlation coefficient equal to or greater than a threshold is set as the optimal number of categories in joint learning. Figure 17 In this embodiment, the optimal number of categories is 4. After determining the optimal number of categories, the category numbers (category names) are reassigned to the individual categories. For example, category number 1 is assigned to category 1 / 5, category number 2 is assigned to category 2 / 4 / 7, category number 3 is assigned to category 3, and category number 4 is assigned to category 6.

[0116] Initial category tag

[0117] When assigning initial category labels to each cell culture data point, firstly, as... Figure 18 As shown by arrow Ar1 in the diagram, matrix W1 is generated by calculating the z-score and normalizing the corresponding data (contribution) in the basis matrix W indicating the clustering results of the corresponding samples.

[0118] Next, as Figure 18 As shown by arrow Ar2 in the diagram, for each sample, the class with the highest z-score is extracted, and a matrix W2 is generated such that for each sample, the class with the highest z-score records data of 1, and for the other classes, data of 0 is recorded. Figure 18 In matrix W2, the color-coded categories are the categories that have the highest z-scores for the corresponding samples. For example, in sample 1, among categories 1 to 7, category 5 has the highest z-score.

[0119] Next, as Figure 18 As shown by arrow Ar3 in the diagram, matrix W3 is generated by combining the categories that have already been merged into one category in determining the optimal number of categories, which is also in matrix W2.

[0120] Since categories 1 and 5 were already merged into category 1 in the determination of the optimal number of categories, in matrix W3, for the sample with the highest z-score among either category 1 or category 5 before the merge, the category with the highest z-score is set as category 1. Similarly, since categories 2, 4, and 7 were already merged into category 2 in the determination of the optimal number of categories, in matrix W3, for the sample with the highest z-score among either category 2, category 4, or category 7 before the merge, the category with the highest z-score is set as category 2.

[0121] Since class number 3 was reassigned to class 3 before merging in the determination of the optimal number of classes, in matrix W3, for the sample with the largest z-score among the samples of class 3 before merging, the class with the largest z-score is set as class 3. Similarly, since class number 4 was assigned to class 6 before merging in the determination of the optimal number of classes, in matrix W3, for the sample with the largest z-score among the samples of class 6 before merging, the class with the largest z-score is set as class 4.

[0122] Next, in matrix W3, the class label indicating the class with the highest z-score for each sample is assigned to the cell culture data of the sample as the initial class label. For example, the class label of class 1 is assigned as the initial class label to the cell culture data of sample 1.

[0123] In this way, after treating the categories to be merged in determining the optimal number of categories as a single category, the category label indicating the category with the largest contribution in each sample is assigned as the initial category label to the cell culture data of the sample. For example, the same category label is assigned as the initial category label to the cell culture data of sample 4, which has the highest z-score in category 1 before merging, and to the cell culture data of sample 1, which has the highest z-score in category 5 before merging.

[0124] • Optimize the effect of initial category labels and number of categories

[0125] As described above, by applying soft clustering to each data point, the optimal number of categories in the joint learning was determined to be 4, and initial category labels from category 1 to category 4 were assigned to the cell culture data of the corresponding samples.

[0126] The applicant extracts nine data points (final culture day data) from multiple cell culture data sets as test data. Server 101 then uses the remaining data (excluding the final culture day data of the sample designated as test data and data from all days of other samples) as training data to perform learning and calculate prediction accuracy. This learning is performed by inputting multiple cell culture data sets with initial class labels assigned through initial class labeling processing and the optimal number of classes (=4) determined through optimal number of classes determination processing.

[0127] The applicant used four types of cumulative administered stimulation values ​​(cumulative values ​​of OKT3 and IC3 / 28 as antibody stimulation, and cumulative values ​​of IL2 and IL7 as cytokine stimulation) as explanatory variables and CD8+ cell count as the target variable to create a linear regression model, ultimately obtaining four predictive models. The applicant used these predictive models to predict CD8+ cell counts in the test data and evaluated the models by comparing them with the true values ​​in the test data. As a comparison target to confirm the effect, the applicant prepared predictive models for a case where the number of classes is 7 (except for the condition that the number of classes and the initial class labels are the same), and similarly evaluated the predictive models.

[0128] Figure 19 This is a diagram illustrating an example of the prediction accuracy of the prediction model when the number of categories is 4 and when the number of categories is 7.

[0129] exist Figure 19 In the diagram, the error rate between the predicted and true values ​​for the nine corresponding items of the test data is indicated by a box plot. Figure 19 In the box plot, the horizontal axis indicates the number of categories, and the vertical axis indicates the error rate.

[0130] When comparing the prediction accuracy of the prediction model with 4 categories (the optimal number of categories) with the prediction accuracy of the prediction model with 7 categories (a non-optimal number of categories), the overall error rate is lower and the prediction accuracy of the prediction model is higher when the number of categories is 4.

[0131] In this way, it becomes possible to improve the prediction accuracy of the prediction model built by joint learning by determining the optimal number of classes in joint learning and assigning the optimal initial class labels to the data used for learning (generating) the prediction model before joint learning.

[0132] <2. Configuration and Operation of Each Device>

[0133] Server configuration

[0134] Figure 20 This is a block diagram illustrating an embodiment of the functional configuration of server 101.

[0135] like Figure 20 As shown, server 101 includes a data acquisition unit 151, a soft clustering unit 152, an optimal number of categories determination unit 153, a labeling unit 154, a learning unit 155, a communication unit 156, and an aggregation unit 157.

[0136] The data acquisition unit 151 acquires a dataset for learning (generating) a prediction model and provides the dataset to the soft clustering unit 152 and the labeling unit 154.

[0137] The soft clustering unit 152, for example, applies jNMF to each data point included in the dataset provided by the data acquisition unit 151, and extracts the contribution of each data point to each category and feature quantity of each record item. Specifically, the soft clustering unit 152 calculates the basis matrix indicating the clustering results and the coefficient matrix indicating the extracted feature quantities.

[0138] The soft clustering unit 152 provides the feature extraction results to the optimal number of categories determination unit 153 and the clustering results to the labeling unit 154.

[0139] The optimal number of categories determination unit 153 determines the optimal number of categories in joint learning based on the clustering results of the soft clustering unit 152, and notifies the learning unit 155 and the communication unit 156 of the determined optimal number of categories. Furthermore, when determining the optimal number of categories, the optimal number of categories determination unit 153 provides the labeling unit 154 with labeling information indicating which category each basic vector (each category before merging) has been merged into.

[0140] The labeling unit 154 assigns initial category labels to each data point included in the dataset provided by the data acquisition unit 151 based on the clustering results of the soft clustering unit 152 and the labeling information provided by the optimal number of categories determination unit 153. The labeling unit 154 then provides the learning unit 155 with a dataset containing the data assigned the initial category labels.

[0141] Learning unit 155 uses the optimal number of categories determined by optimal category number determination unit 153 and the dataset provided by labeling unit 154 as input to perform learning, thereby generating a predictive model corresponding to the optimal number of categories. In other words, learning unit 155 uses multiple data points with the optimal number of categories to perform predictive model learning. Specifically, learning unit 155 generates a predictive model by performing learning while updating the category labels assigned to each data point, thereby reducing the difference between the predicted and true values. Learning unit 155 provides the generated predictive model to communication unit 156.

[0142] Communication unit 156 sends the prediction model provided by learning unit 155 and information indicating the optimal number of categories determined by the optimal number of categories determination unit to the information processing devices 102 of each medical institution. Additionally, communication unit 156 receives prediction models sent from the information processing devices 102 of each medical institution and provides the received prediction models to aggregation unit 157. Furthermore, communication unit 156 sends the prediction models provided from aggregation unit 157 to the information processing devices 102 of each medical institution.

[0143] The aggregation unit 157 aggregates the prediction models retrained by the information processing equipment 102 of various medical institutions, generates new prediction models, and provides the new prediction models to the communication unit 156.

[0144] It should be noted that the administrator (user) of server 101 can be presented with information such as the initial category labels assigned to each data point, and the update status of the category labels during the learning process of learning unit 155. For example, this information is displayed on the display unit of server 101. The administrator can confirm the status of the predictive model by viewing the information.

[0145] Additionally, identification information can be sent to or from the information processing equipment 102 of each medical institution. When merging categories in the information processing equipment 102, the identification information on the information processing equipment 102 indicates which category each base vector (each category before merging) has been merged into.

[0146] When labeling information from server 101 is sent to information processing device 102, information processing device 102 can merge categories based on the labeling information from server 101, for example, by performing clustering in the same manner as when server 101 learns. When labeling information from information processing device 102 is sent from information processing devices 102 of various medical institutions to server 101, the labeling information from information processing device 102 can be presented to the administrator together with the labeling information from server 101, etc.

[0147] Configuration of information processing equipment

[0148] Figure 21 This is a block diagram illustrating an embodiment of the functional configuration of the information processing device 102.

[0149] like Figure 21 As shown, the information processing device 102 includes a data acquisition unit 171, a soft clustering unit 172, a category merging unit 173, a labeling unit 174, a learning unit 175, and a communication unit 176.

[0150] The data acquisition unit 171 acquires a dataset including patient data for retraining the prediction model and provides the dataset to the soft clustering unit 172 and the labeling unit 174.

[0151] The soft clustering unit 172, for example, applies jNMF to each patient data point included in the dataset provided by the data acquisition unit 171, and extracts the contribution of each patient data point to each category and feature quantity of each record item. Specifically, the soft clustering unit 172 calculates the basis matrix indicating the clustering results and the coefficient matrix indicating the extracted feature quantities.

[0152] The soft clustering unit 172 provides the feature extraction results to the category merging unit 173 and the clustering results to the labeling unit 174.

[0153] The category merging unit 173 merges the categories into which each data item is classified by the soft clustering unit 172 until the number of categories becomes the optimal number of categories determined by the server 101. Specifically, the category merging unit 173 performs processing according to the optimal number of categories determined by the server 101 and merges categories by adjusting the threshold of the correlation coefficient, etc., until the number of categories becomes the optimal number of categories determined by the server 101.

[0154] It should be noted that in the information processing device 102, the threshold for the correlation coefficient is set to be no less than a specific value (e.g., 0.5). If, even when merging combinations of categories with a correlation coefficient greater than 0.5, the number of categories is not optimal, retraining is performed using that number of categories.

[0155] The category merging unit 173 informs the learning unit 175 of the optimal number of categories. In addition, the category merging unit 173 provides the labeling unit 174 with flag information indicating which category each basis vector (each category before merging) has been merged into.

[0156] The labeling unit 174 assigns initial category labels to each patient data point included in the dataset provided by the data acquisition unit 171 based on the clustering results of the soft clustering unit 172 and the labeling information provided by the category merging unit 173. The labeling unit 174 then provides the learning unit 175 with a dataset containing the patient data assigned the initial category labels.

[0157] Learning unit 175 uses the optimal number of classes notified by class merging unit 173 and the dataset provided by labeling unit 174 as input to retrain the prediction model provided by communication unit 176. In other words, learning unit 175 performs retraining of the prediction model using multiple patient data points classified into the optimal number of classes. Specifically, learning unit 175 performs retraining while updating the class labels assigned to each patient data point, thereby reducing the difference between the predicted and true values. Learning unit 175 provides the retrained prediction model to communication unit 176.

[0158] Communication unit 176 receives a prediction model and information indicating the optimal number of categories from server 101. Communication unit 176 provides the prediction model to learning unit 175 and the information indicating the optimal number of categories to category merging unit 173. Furthermore, communication unit 176 transmits the prediction model provided by learning unit 175 to server 101.

[0159] Server operation

[0160] Next, refer to Figure 22 The flowchart will describe the processing performed by server 101.

[0161] In step S1, the soft clustering unit 152 applies soft clustering to each cell culture data entry included in the dataset. Through soft clustering, clustering of each sample (extracting the contribution of each cell culture data entry to each category) and feature extraction of each record are performed.

[0162] In step S2, the optimal number of categories determination unit 153 performs optimal number of categories determination processing based on the feature extraction results from the soft clustering unit 152. Through this optimal number of categories determination processing, combinations of categories with high feature similarity are merged into one category, and the optimal number of categories in joint learning is determined. (See later...) Figure 23 Describe the details of the process for determining the optimal number of categories.

[0163] In step S3, the labeling unit 154 performs initial category labeling processing based on the clustering results of the soft clustering unit 152 and the category merging results of the optimal number of categories determination unit 153. Through this initial category labeling processing, initial category labels are assigned to each cell culture data point. (See later...) Figure 24 Describe the details of the initial category tagging process.

[0164] In step S4, the learning unit 155 uses the optimal number of categories determined by the optimal number of categories determination unit 153 and the cell culture data with initial category labels assigned by the labeling unit 154 as input to learn and generate a prediction model corresponding to the optimal number of categories.

[0165] In step S5, the communication unit 156 sends the prediction model to the information processing equipment 102 of each medical institution.

[0166] In step S6, the communication unit 156 receives the prediction model sent from the information processing equipment 102 of each medical institution.

[0167] In step S7, aggregation unit 157 aggregates the prediction models retrained by the information processing devices 102 of each medical institution and generates a new prediction model. Afterward, the process returns to step S5 and repeats the subsequent processing. For example, the prediction model is updated repeatedly until the prediction accuracy of the prediction model becomes higher than a predetermined threshold.

[0168] See Figure 23 The flowchart will describe in Figure 22 The optimal number of categories is determined in step S2.

[0169] In step S21, the optimal number of categories determination unit 153 extracts the feature quantities of each category of the prediction term from the feature quantity extraction results of soft clustering unit 152.

[0170] In step S22, the optimal number of categories determination unit 153 calculates the correlation coefficient of the time variation of the feature quantities between categories.

[0171] In step S23, the optimal number of categories determination unit 153 determines whether there is a combination of categories with a correlation coefficient equal to or greater than a threshold.

[0172] If, in step S23, it is determined that there is a combination of categories with a correlation coefficient equal to or greater than the threshold, in step S24, the optimal number of categories determination unit 153 merges the combination of categories with a correlation coefficient equal to or greater than the threshold into one category.

[0173] In step S25, the optimal number of categories determination unit 153 merges the feature values ​​of each category into one category on average each day. Afterwards, the process returns to step S22 and performs subsequent processing.

[0174] On the other hand, if it is determined in step S23 that there is no combination of categories with a correlation coefficient equal to or greater than the threshold, in step S26, the optimal number of categories determination unit 153 determines the number of categories at this time as the optimal number of categories.

[0175] After that, the process returns to Figure 22 Step S2 is performed, and subsequent processing is carried out.

[0176] See Figure 24 The flowchart will describe in Figure 22The initial category labeling process is performed in step S3.

[0177] In step S41, the labeling unit 154 extracts the category with the highest contribution (z-score) for each sample based on the clustering results of the soft clustering unit 152.

[0178] In step S42, the marking unit 154 treats the categories merged in the optimal category quantity determination process as a single category.

[0179] In step S43, the labeling unit 154 assigns an initial category label to the cell culture data of each sample. Specifically, the labeling unit 154 assigns a category label to the merged category that contributes the most to the sample's cell culture data.

[0180] After that, the process returns to Figure 22 Step S3 is performed, and subsequent processing is carried out.

[0181] Operation of information processing equipment

[0182] Next, refer to Figure 25 The flowchart will describe the processing performed by the information processing device 102. For example, when the information processing device 102 receives the prediction model and information indicating the optimal number of categories sent from the server 101, it begins... Figure 25 The processing.

[0183] In step S61, the soft clustering unit 172 applies soft clustering to each cell culture data (patient data) included in the dataset. Through soft clustering, clustering of the corresponding samples (extraction of the contribution of the corresponding patient data to the corresponding category) and extraction of feature quantities for the corresponding record are performed.

[0184] In step S62, the category merging unit 173 performs category merging processing based on the feature extraction results from the soft clustering unit 172. Through category merging processing, combinations of categories with high similarity in feature values ​​are merged into one category until the number of categories reaches the optimal number. (See later...) Figure 26 Describe the details of the category merging process.

[0185] In step S63, the labeling unit 174 performs initial category labeling processing based on the clustering results of the soft clustering unit 172 and the category merging results of the category merging unit 173. Through this initial category labeling processing, initial category labels are assigned to each patient data point. (See later...) Figure 27 Describe the details of the initial category tagging process.

[0186] In step S64, the learning unit 175 uses the optimal number of categories determined by the server 101 and the data with initial category labels assigned by the labeling unit 174 as input to retrain the prediction model.

[0187] In step S65, the communication unit 176 transmits the retrained prediction model to the server 101.

[0188] See Figure 26 The flowchart will describe in Figure 25 The category merging process is performed in step S62.

[0189] In step S81, the category merging unit 173 extracts the feature quantities of each category of the prediction term from the feature quantity extraction results of the soft clustering unit 172.

[0190] In step S82, the category merging unit 173 determines whether the number of categories matches the optimal number of categories.

[0191] If the number of categories determined in step S82 does not match the number of optimal categories, in step S83, the category merging unit 173 calculates the correlation coefficient of the time change of the characteristic quantities between the categories.

[0192] In step S84, the category merging unit 173 merges combinations of categories with a correlation coefficient equal to or greater than a threshold into one category.

[0193] In step S85, the category merging unit 173 merges the feature values ​​of each category in a single category on average per day. Afterward, the process returns to step S82 and performs subsequent processing.

[0194] If, in step S82, it is determined that the number of classes matches the optimal number of classes, the process returns to... Figure 25 Step S62, and then perform subsequent processing.

[0195] See Figure 27 The flowchart will describe Figure 25 The initial category labeling process is performed in step S63.

[0196] In step S101, the labeling unit 174 extracts the category with the highest contribution (z-score) for each sample based on the clustering results of the soft clustering unit 172.

[0197] In step S102, the marking unit 174 treats the categories merged in the category merging process as a single category.

[0198] In step S103, the labeling unit 174 assigns an initial category label to the patient data of each sample. Specifically, the labeling unit 174 assigns the category label of the category that has the highest contribution to the sample after merging to the patient data of the sample.

[0199] After that, the process returns to Figure 25 Step S63, and then perform subsequent processing.

[0200] As described above, in the learning system of this technique, soft clustering is applied to multiple datasets used for learning the predictive model, extracting feature quantities for each category of the predictive item, and determining the optimal number of categories (optimal number of categories) for performing the predictive model learning based on the feature quantities. Furthermore, in the learning system of this technique, soft clustering is applied to multiple datasets for learning the predictive model, and together with the feature quantities for each category of the predictive item, the contribution of each data point to each category is extracted, and category labels are assigned to each segment of the multiple datasets based on the results of category merging performed when determining the optimal number of categories and the contribution quantities.

[0201] Before joint learning, it became possible to improve the prediction accuracy of prediction models built through joint learning because the number of categories into which each piece of data is classified and the initial class labels assigned to each piece of data were optimized for learning (generating) or retraining the prediction model.

[0202] <3. Variations>

[0203] When performing joint learning on data using hierarchical clustering, the goal is to classify each data point into appropriate groups and assign optimal group labels to each data point before the predictive model learns. Groups are lower-level categories, and data classified into the same group are classified into the same category. Group labels indicate which group the data point with the assigned group label has been classified into.

[0204] Server 101 assigns the best group label to each data point by applying soft clustering to each data point.

[0205] Figure 28 This is a diagram illustrating another embodiment of a process including soft clustering performed by server 101.

[0206] Before generating the predictive model, server 101 performs soft clustering, initial number of categories determination, and initial category labeling, similar to [see [link]]. Figure 4 The described situation. For example... Figure 28 As indicated in #51, server 101 assigns a group label to each piece of data during the initial category labeling.

[0207] Figure 29This is a diagram used to explain the details of the processing of initial data tags and group tags.

[0208] When assigning initial category and group labels to each data entry, it is similar to seeing [see also...] Figure 18 The described scenario involves generating matrix W1 based on the basis matrix W, and generating matrix W2 based on matrix W1.

[0209] Next, as Figure 29 As indicated by arrow Ar11 in the diagram, in matrix W2, for each sample, the class label of the class with the highest z-score is assigned to the sample's cell culture data. For example, the group label of class 5 (group 5) is assigned to sample 1.

[0210] Next, as Figure 29 As shown by arrow Ar3 in the diagram, matrix W3 is generated by merging multiple categories that were combined into one category in determining the optimal number of categories into another category in matrix W2. Furthermore, in matrix W3, for each sample, the category label of the category with the highest z-score is assigned to the cell culture data of that sample. For example, the category label of category 1 is assigned to the cell culture data of sample 1.

[0211] In this way, before considering the categories to be merged in determining the optimal number of categories as a single category, the group label of the category (group) that contributes the most to each sample is assigned to the cell culture data of the sample. Subsequently, after considering the categories to be merged in determining the optimal number of categories as a single category, the category label of the category that contributes the most to each sample is assigned to the sample.

[0212] As described above, by performing soft clustering on multiple cell culture data, initial category labels from category 1 to category 4 are assigned to each cell culture data, and group labels from group 1 to group 7 are assigned to each cell culture data.

[0213] The applicant extracts nine data points (final culture day data) from multiple cell culture datasets as test data. Server 101 then uses the remaining data (excluding the final culture day data of the sample extracted as test data and data from all days of other samples) as training data to perform learning and obtain prediction accuracy. This learning is performed by inputting multiple cell culture datasets that have been assigned initial class labels through initial class labeling and group labels through group labeling, along with the optimal number of classes (=4) determined through optimal number of classes determination.

[0214] The applicant used four types of cumulative applied stimulation values ​​(cumulative values ​​of OKT3 and IC3 / 28 applied as antibody stimulation, and cumulative values ​​of IL2 and IL7 applied as cytokine stimulation) as explanatory variables and CD8+ cell count as the target variable to create a linear regression model, ultimately obtaining four predictive models. The applicant used the predictive models to predict the CD8+ cell count in the test data and evaluated the predictive models by comparing them with the true values ​​in the test data. As a comparison target to confirm the effect, the applicant prepared a predictive model for a case where group labels were randomly assigned to each cell culture data point (with all other conditions being the same), and evaluated the predictive models similarly.

[0215] Figure 30 This is a diagram illustrating an example of the prediction accuracy of a prediction model when group labels are assigned based on clustering results and when random group labels are assigned.

[0216] exist Figure 30 In the diagram, the box plot indicates the error rate between the predicted and true values ​​calculated for the corresponding nine test data points. Figure 30 In the box plot, the horizontal axis indicates what kind of group label is assigned to each cell culture data point, and the vertical axis indicates the error rate.

[0217] When comparing the prediction accuracy of a prediction model with group labels based on clustering results (optimal group labels) to that with random group labels (non-optimal group labels), the prediction model with optimal group labels generally has a lower error rate and higher prediction accuracy.

[0218] In this way, by assigning optimal group labels to the data used for learning (generating) the predictive model before joint learning, it becomes possible to improve the predictive accuracy of the predictive model built through joint learning.

[0219] <Example of Computer Configuration>

[0220] The aforementioned series of processes can be performed by hardware or by software. In the case where the processes are performed by software, the program constituting the software is installed from the program recording medium into a computer contained in dedicated hardware, or into a general-purpose personal computer, etc.

[0221] Figure 31 This is a block diagram illustrating an embodiment of the hardware configuration of a computer that performs the above-described series of processes via a program.

[0222] The central processing unit (CPU) 501, read-only memory (ROM) 502 and random access memory (RAM) 503 are interconnected via bus 504.

[0223] Furthermore, the input / output interface 505 is connected to the bus 504. Input units 506, including a keyboard, mouse, etc., and output units 507, including a display, speakers, etc., are connected to the input / output interface 505. Additionally, storage units 508, including hard disks, non-volatile memory, etc., communication units 509, including network interfaces, etc., and drivers 510 that drive removable media 511 are connected to the input / output interface 505.

[0224] In the computer configured as described above, the above series of processes are performed by, for example, loading the program stored in the storage unit 508 into the RAM 503 via the input / output interface 505 and the bus 504 through the CPU 501 and executing the program.

[0225] The program executed by CPU 501 is recorded on, for example, removable medium 511, or provided via wired or wireless transmission media such as a local area network, the Internet, or digital broadcasting, and is installed in storage unit 508.

[0226] It should be noted that a program executed by a computer may be a program in which processing is performed sequentially in the order described in this specification, or a program in which processing is performed in parallel or at necessary time intervals, such as when a call is made.

[0227] In this specification, a system means a group of multiple components (devices, modules (parts), etc.), and it is irrelevant whether all components are housed in the same housing. Therefore, multiple devices housed in separate housings and connected via a network, as well as a device in which multiple modules are housed in a single housing, are all systems.

[0228] The effects described in this specification are illustrative and not limiting, and other effects may also be provided.

[0229] The implementation of this technology is not limited to the above-described implementation, and various modifications are possible without departing from the spirit of this technology.

[0230] For example, this technology can be configured using cloud computing, where a function is shared and jointly processed by multiple devices via a network.

[0231] Furthermore, each step described in the flowchart above can be performed by a single device or by multiple devices in a distributed manner.

[0232] Furthermore, in cases where a step includes multiple types of processing, the multiple types of processing included in that step can be performed by a single device or by multiple devices in a distributed manner.

[0233] <Example of Configuration Combination>

[0234] This technology can also be configured as follows. (1)

[0236] An information processing method includes: applying soft clustering to multiple data sets used for learning a prediction model to extract feature quantities for each category of predictions relative to the prediction model; and

[0237] Based on these features, the optimal number of categories for learning the predictive model is determined. (2)

[0239] According to the information processing method in (1), where, The optimal number of categories is determined by merging multiple categories into one category based on the similarity of feature quantities between categories. (3)

[0241] According to the information processing method in (2), where, Categories whose correlation coefficients of the time changes of feature quantities between categories are equal to or greater than a threshold are merged into one category. (4)

[0243] According to the information processing method in (3), it further includes: Soft clustering is applied to multiple datasets to extract the contribution of each dataset to its corresponding category, along with feature values; and Based on the merging results and contributions of the categories, a category label indicating the category is assigned to each of the multiple data points. (5)

[0245] According to the information processing method in (4), it further includes: Based on contribution, assign a group label to each of the multiple data points, indicating the group as a lower-level category. (6)

[0247] The information processing method according to any one of (3) to (5) further includes: Present the user with tagging information that indicates which category each category has been merged into before the merge. (7)

[0249] According to any one of the information processing methods in (1) to (6), where, The prediction model is learned through joint learning. (8)

[0251] According to the information processing method in (7), where, Feature quantities are extracted by applying soft clustering to multiple data points used to generate the predictive model. (9)

[0253] According to the information processing method of (8), it further includes: Through the first information processing device Extracting feature quantities, Determine the optimal number of categories, and A predictive model is generated by learning from multiple datasets; and Through the second information processing device The prediction model is retrained using multiple data points categorized into the optimal number of categories determined by the first information processing device. (10)

[0255] According to the information processing method of (9), it further includes: The first information processing device aggregates the prediction models that have been retrained by multiple second information processing devices. (11)

[0257] According to any one of the information processing methods in (1) to (10), where, Soft clustering includes jNMF. (12)

[0259] An information processing system, comprising: Soft clustering units apply soft clustering to multiple data points used for learning the prediction model to extract feature quantities for each category relative to the prediction terms of the prediction model; and The optimal number of categories determination unit determines the optimal number of categories for learning the prediction model based on the number of features.

[0260] Reference Symbol List

[0261] 101 Server

[0262] 102 Information processing equipment

[0263] 151 Data Acquisition Unit

[0264] 152 soft clustering units

[0265] 153 Optimal Category Quantity Determination Unit

[0266] 154 marker units

[0267] 155 Learning Units

[0268] 156 Communication Units

[0269] 157 Aggregation Units

[0270] 171 Data Acquisition Unit

[0271] 172 Soft clustering units

[0272] 173 Category Merging Units

[0273] 174 marker units

[0274] 175 Learning Units

[0275] 176 Communication Units

Claims

1. An information processing method, comprising: Soft clustering is applied to multiple data sets used for learning the prediction model to extract the feature quantity of each category relative to the prediction terms of the prediction model; as well as The optimal number of categories for learning the prediction model is determined based on the aforementioned feature values.

2. The information processing method according to claim 1, wherein, The optimal number of categories is determined by merging multiple categories into one category based on the similarity of feature quantities between categories.

3. The information processing method according to claim 2, wherein, Categories whose correlation coefficients for the time changes of the aforementioned feature quantities are equal to or greater than a threshold are merged into a single category.

4. The information processing method according to claim 3, further comprising: The soft clustering is applied to multiple data points to extract the contribution of each data point to the corresponding category, together with the feature value. as well as Based on the merging results of the categories and the contribution, a category label indicating the category is assigned to each of the multiple data entries.

5. The information processing method according to claim 4, further comprising: Based on the contribution, a group label indicating a lower-level group of categories is assigned to each of the multiple data entries.

6. The information processing method according to claim 3, further comprising: The user is presented with tagging information indicating which category each category has been merged into before the merge.

7. The information processing method according to claim 1, wherein, The prediction model is learned through joint learning.

8. The information processing method according to claim 7, wherein, The feature quantities are extracted by applying the soft clustering to multiple data points used to generate the prediction model.

9. The information processing method according to claim 8, further comprising: Through the first information processing device Extract the feature quantity, Determine the optimal number of categories, and The predictive model is generated by learning from multiple sets of the aforementioned data; and Through the second information processing device The prediction model is retrained using multiple data points categorized into the optimal number of categories determined by the first information processing device.

10. The information processing method according to claim 9, further comprising: The first information processing device aggregates the prediction models that have been retrained by multiple second information processing devices.

11. The information processing method according to claim 1, wherein, The soft clustering includes jNMF.

12. An information processing system, comprising: A soft clustering unit applies soft clustering to multiple data points used for learning a prediction model to extract feature quantities for each category of the prediction terms relative to the prediction model. as well as An optimal number of categories determination unit determines the optimal number of categories for learning the prediction model based on the feature quantity.

Images