Analysis device, analysis method, and analysis program

Abstract

To enable an appropriate predictive model to be generated for each base without moving data present in a plurality of bases outside the bases.SOLUTION: An analysis device capable of communicating with a plurality of learning devices includes: a receiving unit that receives a converted feature amount obtained by converting a feature amount included in learning data possessed by each of the plurality of learning devices according to a predetermined rule; a distribution analysis unit that analyzes distribution of the plurality of feature amounts of the plurality of learning devices based on the converted feature amount of each of the learning devices received by the receiving unit; and an output unit that outputs distribution analysis results analyzed by the distribution analysis unit.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an analytical device, an analytical method, and an analytical program for analyzing data. [Background technology]

[0002] Machine learning is one of the technologies that realize AI (Artificial Intelligence). Machine learning technology consists of a learning process and a prediction process. First, the learning process calculates learning parameters so that the error between the predicted value obtained from the input feature vector and the actual value (true value) is minimized. Next, the prediction process calculates new predicted values ​​from data that was not used in learning (hereafter referred to as test data).

[0003] To date, methods have been devised to calculate and compute learning parameters that maximize the accuracy of predicted values. For example, a technique called a perceptron outputs predicted values ​​based on the result of a linear combination of the input feature vector and the weight vector. Neural networks, also known as multi-perceptrons, have the ability to solve linearly inseparable problems by stacking multiple perceptrons in a multi-layered structure. Deep learning is a technique that introduces new technologies such as dropout into neural networks and has attracted attention as a method that can achieve high prediction accuracy. In this way, machine learning techniques have been developed with the aim of improving prediction accuracy, and their prediction accuracy is now showing capabilities that surpass those of humans.

[0004] When implementing machine learning technology in society, there are challenges beyond predictive accuracy. For example, security, methods for updating models after delivery, and limitations on the use of finite resources such as memory.

[0005] One security challenge is data confidentiality. In fields such as healthcare and finance, where predictive models are created using data containing personal information, the high level of data confidentiality can make it difficult to move the data outside of the storage location. Generally, machine learning can achieve high predictive accuracy by using large amounts of data for training.

[0006] When training a model using only data acquired from a single location, the resulting model may be limited to a very localized scope due to the small sample size and the characteristics of the region where the data was collected. In other words, there is a need for machine learning techniques that enable the creation of predictive models that achieve high accuracy with the diverse data from various locations without having to transfer data from one location to another.

[0007] Patent Document 1 below discloses a method and system with a federated learning model for healthcare applications. This federated learning system consists of multiple end-user edge devices, one or more federated learner update repositories, and one or more clouds. Each edge device has a federated learner model configured to send tensors to the federated learner update repository. The cloud has a federated learner model configured to send tensors to the federated learner update repository. The federated learner update repository has a backend configuration configured to send model updates to the edge devices and the cloud.

[0008] Patent Document 2 below discloses a machine learning system that performs federated learning. This machine learning system synchronizes the learning model on each client terminal with the master model on the integration server before each of the multiple client terminals starts learning. Each client terminal performs machine learning on the learning model using data stored in the medical institution and sends the learning results to the integration server. The integration server divides the multiple client terminals into multiple client clusters and integrates the learning results for each client cluster to create a master model candidate. The integration server evaluates the inference accuracy of each master model candidate, and if it detects a master model candidate that falls below the accuracy threshold, it extracts the client terminals that are causing the accuracy deterioration from the client cluster used to create that master model candidate.

[0009] Non-patent document 1 discloses a practical method for federated learning of deep networks based on iterative model averaging. This federated learning method uses one common model as an initial value and trains it on each data point at each location to generate a predictive model. The model parameter information of the generated predictive model is sent to a server. The server repeatedly performs a process to generate a global predictive model from the model parameter information of the predictive model using coefficients corresponding to the amount of data trained. Finally, a global predictive model is generated that achieves high predictive accuracy for data from all locations. [Prior art documents] [Patent Documents]

[0010] [Patent Document 1] U.S. Patent Publication 2020 / 0293887 [Patent Document 2] International Publication No. 2021 / 059607 [Non-patent literature]

[0011] [Non-Patent Document 1] H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson and Blaise Aguera y Arcas, “Communication-efficient learning of deep networks from decentralized data”, In Artificial Intelligence and Statistics, pp. 1273-1282, 2017. [Overview of the project] [Problems that the invention aims to solve]

[0012] The method and system described in Patent Document 1 do not take into account the characteristics of each end user in order to smooth out collective bias for each end user and generate a predictive model.

[0013] The machine learning system described in Patent Document 2 identifies locations that cannot achieve the desired prediction accuracy and performs federated learning again with the locations excluding those locations. However, repeatedly performing federated learning means that data that has already been learned is retrained many times, which is redundant.

[0014] In federated learning techniques such as those described in Non-Patent Document 1, depending on the small number of data samples at each location and the variability in data characteristics, such as the characteristics of the regions where the data was collected, a global prediction model with degraded prediction performance may be generated for data from any location.

[0015] The present invention aims to enable the generation of appropriate predictive models for each location without moving data from multiple locations outside of those locations. [Means for solving the problem]

[0016] An analysis apparatus that is an aspect of the invention disclosed in the present application is an analysis apparatus capable of communicating with a plurality of learning apparatuses, and includes: a receiving unit that receives conversion feature amounts obtained by converting feature amounts included in learning data possessed by each of the plurality of learning apparatuses according to a predetermined rule; a distribution analysis unit that analyzes the distribution of the plurality of feature amounts of the plurality of learning apparatuses based on the conversion feature amounts for each learning apparatus received by the receiving unit; and an output unit that outputs a distribution analysis result analyzed by the distribution analysis unit.

Advantages of the Invention

[0017] According to a typical embodiment of the present invention, it is possible to realize generation of an appropriate prediction model for each site without moving data at a plurality of sites outside the sites. Problems, configurations, and effects other than those described above will be clarified by the description of the following examples.

Brief Description of the Drawings

[0018] [Figure 1] FIG. 1 is an explanatory diagram showing an example of federated learning. [Figure 2] FIG. 2 is a block diagram showing an example of the hardware configuration of a computer. [Figure 3] FIG. 3 is a block diagram showing an example of the functional configuration of a server according to Example 1. [Figure 4] FIG. 4 is a block diagram showing an example of the functional configuration of a site according to Example 1. [Figure 5] FIG. 5 is an explanatory diagram showing Example 1 of similarity analysis of conversion feature amounts by a distribution analysis unit. [Figure 6] FIG. 6 is an explanatory diagram showing Example 2 of similarity analysis of conversion feature amounts by a distribution analysis unit. [Figure 7] FIG. 7 is an explanatory diagram showing Example 1 of learning by a server and a site. [Figure 8] FIG. 8 is an explanatory diagram showing Example 2 of learning by a server and a site. [Figure 9] FIG. 9 is a flowchart showing an example of a preprocessing procedure for integrated learning by a server. [Figure 10] Figure 10 is a flowchart showing an example of a pre-training processing procedure at a research site. [Figure 11] Figure 11 is an explanatory diagram showing example 1 of the display screen. [Figure 12] Figure 12 is an explanatory diagram showing example 2 of the display screen. [Figure 13] Figure 13 is an explanatory diagram illustrating federated learning method 1, which enables appropriate learning at multiple locations where training data cannot be transferred outside of the location. [Figure 14] Figure 14 is an explanatory diagram illustrating federated learning method 2, which enables appropriate learning at each of multiple locations where training data cannot be transferred outside of the location. [Figure 15] Figure 15 is a block diagram showing an example of the functional configuration of a computer. [Modes for carrying out the invention]

[0019] The analysis device transforms the data characteristics of each location without moving the data from multiple locations outside of those locations, then moves it outside of those locations and performs analysis of the data characteristics of each location. This allows the analysis device to present appropriate model construction methods and grouping processes for each location. (A) Display the clustering results using the distance between transformed features obtained by transforming the features contained in the training data of each location, (B) Display clustering results by calculating scores using associative learning.

[0020] (A) will be explained in Example 1, and (B) will be explained in Example 2. [Examples]

[0021] <Associative Learning> Figure 1 is an explanatory diagram illustrating an example of associative learning. Each of the multiple learning devices (in Figure 1, for example, 10 devices S1 to S10) holds learning data D1 to D10 (if these are not distinguished, they are simply referred to as learning data D), and it is assumed that the learning data D1 to D10 cannot be shared outside of the devices S1 to S10.

[0022] Server 100 is an analysis device that integrates prediction models M1 to M10 (or simply prediction model M) generated at sites S1 to S10 (or simply site S if they are not distinguished). Server 100 has a base prediction model (hereinafter referred to as the base prediction model) M0. The base prediction model M0 may be an untrained neural network or a trained neural network with model parameters such as weights and biases set.

[0023] Centers S1 to S10 are computers that possess training data D1 to D10 and generate predictive models M1 to M10 using training data D1 to D10. Each of the training data D1 to D10 is a combination of input training data and ground truth data.

[0024] First, server 100 sends the base prediction model M0 to locations S1 through S10. Each of locations S1 through S10 uses the base prediction model M0 and their respective training data D1 through D10 to train and generate their own prediction models M1 through M10.

[0025] Each of the locations S1 to S10 sends model parameters θ1 to θ10 (or simply model parameters θ if they are not distinguished) such as weights and biases of the prediction models M1 to M10 to the server 100. The server 100 performs integration processing on the received model parameters θ1 to θ10 and generates an integrated prediction model M100.

[0026] Server 100 repeats the process of updating the integrated prediction model M100 until the generated integrated prediction model M100 achieves the desired prediction accuracy. Alternatively, locations S1 to S10 may send the gradients of the model parameters θ1 to θ10 of the prediction models M1 to M10 to Server 100. Each of locations S1 to S10 sends the model parameters θ1 to θ10, such as weights and biases, of the prediction models M1 to M10 to Server 100.

[0027] The integration process is the process of calculating the average value of the model parameters θ1 to θ10. If the number of samples in the training data D1 to D10 differs, a weighted average may be calculated based on the number of samples in the training data D1 to D10. Alternatively, the integration process may calculate the average value of the gradients of the model parameters θ1 to θ10 sent from each location S1 to S10, instead of the model parameters θ1 to θ10.

[0028] The process for updating the integrated prediction model M100 involves the server 100 sending the integrated prediction model M100 to locations S1 through S10, each location S1 through S10 inputting training data D1 through D10 into the integrated prediction model M100 to perform training, sending the model parameters θ1 through θ10 of the regenerated prediction models M1 through M10 to the server 100, and the server 100 regenerating the integrated prediction model M100. When the generated integrated prediction model M100 achieves the desired prediction accuracy, the federated learning process ends. Alternatively, the process may end after a predetermined number of updates instead of the desired prediction accuracy.

[0029] <Non-independent identical distribution> Generally, it is assumed that the data provided in machine learning is independently and identically distributed (iid). Independent and identically distributed means, for example, that when rolling a die numbered 1 through 6, the probability of each number appearing is the same, and when rolling the die multiple times, the results will be obtained independently regardless of which number appears on which roll. A state where this definition is broken is called non-independent and identically distributed (non-iid).

[0030] Data under non-IID conditions is also called non-IID data, or data where data shift or data skew has occurred. For example, data from hospitals in Japan and hospitals in the United States are expected to have significantly different distributions of patients' physical conditions, race, income, etc., due to differences in insurance systems. Such data can be said to be non-IID data.

[0031] Non-iid data with such variability in characteristics can be broadly classified into four types depending on the nature of the variability: covariate shift (covariate shift or feature distribution skew), concept shift (same features or different features, etc.), label shift (target shift, label distribution skew, prior probability shift, etc.), and concept drift (same label or different label, etc.).

[0032] These are defined by differences in the distribution of features, differences in the distribution of the target variable, and differences in the conditional probabilities of these two. Although the types of non-iid data are statistically defined, multiple types may occur in actual data, making it difficult to evaluate which type occurs and to what intensity.

[0033] If training and test data are separated at a single location so that they are non-iid data, a predictive model that can achieve highly accurate predictions during training will not produce similarly accurate prediction results when applied to validation data because the distributions are different. This makes it impossible to properly validate the accuracy of the predictive model. Therefore, methods such as setting the training and validation data by randomly sampling from a common dataset have been employed.

[0034] Even if the data within a single location is iid data, it is highly likely that the data at each location is non-iid data relative to each other. When the data at each location is non-iid data, it is expected that the model parameters θ1 to θ10 of the appropriate prediction models M1 to M10 at each location will differ significantly. In other words, even if updates are performed multiple times during the integration process, there is a high possibility that the model parameters of the integrated prediction model M100 will not converge to appropriate values.

[0035] One way to determine whether data is non-iid is to compare the data distributions feature by feature. Alternatively, if applying a predictive model M1 generated using training data D1 from site S1 to training data D2 from site S2 results in a significantly lower prediction accuracy than when model M1 was generated, it can be confirmed that the distributions of training data D1 from site S1 and training data D2 from site S2 are different. Furthermore, when generating predictive models to determine whether the training data D1 and D2 are from sites S1 and S2, a higher prediction accuracy suggests a greater difference in the data distributions of training data D1 and D2.

[0036] However, in a federated learning environment, training data D cannot be sent outside of site S. Therefore, it is not possible to directly compare the data distribution between sites S. For this reason, although methods have been considered such as identifying sites that cannot achieve the desired prediction accuracy after federated learning, as shown in Patent Document 2, and then performing federated learning again with the sites excluding those sites, repeatedly performing federated learning means relearning data that has already been trained many times, which is redundant. Below, we will specifically describe a generation method for generating an appropriate prediction model according to the characteristics of the training data D of each site S before federated learning.

[0037] <Example of hardware configuration for a computer (Server 100, Branch S)> Figure 2 is a block diagram showing an example of a computer hardware configuration. Computer 200 includes a processor 201, a storage device 202, an input device 203, an output device 204, and a communication interface (communication IF) 205. The processor 201, storage device 202, input device 203, output device 204, and communication IF 205 are connected by a bus 206. The processor 201 controls computer 200. The storage device 202 serves as the work area for the processor 201. The storage device 202 is also a non-temporary or temporary recording medium that stores various programs and data. Examples of storage devices 202 include ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), and flash memory. The input device 203 inputs data. Examples of input devices 203 include a keyboard, mouse, touch panel, numeric keypad, and scanner. The output device 204 outputs data. Output devices 204 include, for example, displays and printers. The communication interface 205 connects to a network and sends and receives data.

[0038] <Example of a functional configuration of computer 200> Figure 3 is a block diagram showing an example of the functional configuration of server 100 according to Embodiment 1. Server 100 includes a receiving unit 301, a distribution analysis unit 302, a generation unit 303, and an output unit 304 (transmission unit 341 and display unit 342). Specifically, the receiving unit 301, the distribution analysis unit 302, the generation unit 303, and the output unit 304 realize their functions, for example, by causing the processor 201 to execute a program stored in the storage device shown in Figure 2, or by the communication IF 205.

[0039] The receiver 301 receives, via the communication IF 205, the transformed features TF1 to TF10 (or simply transformed features TF if they are not distinguished), the model parameters θ1 to θ10 learned by the sites S1 to S10, and the accuracy verification results of the integrated prediction model M100. Transformed features TF are transformed data obtained by transforming multiple features, which are training data included in the learning data D, according to pre-set rules.

[0040] The distribution analysis unit 302 uses the transformed features TF1 to TF10 received by the receiving unit 301 to perform similarity analysis of the training data D. In the example in Figure 1, there are 10 locations S1 to S10, so the distribution analysis unit 302 calculates the similarity of the training data D for each of the 45 combinations of locations S, two at a time.

[0041] Methods for calculating the similarity of the training data D include, for example, the Euclidean distance or cosine distance between transformed features TF, metrics representing differences in distribution such as Jensen-Shannon divergence, dimensionality reduction techniques such as t-SNE and PCA that transform vectors into three dimensions or less, hierarchical clustering methods such as Ward's method, and non-hierarchical clustering methods such as K-Means.

[0042] The generation unit 303 generates one or more prediction models based on the model parameters θ of the prediction models M transmitted from one or more locations S. Specifically, for example, for clusters containing only one location S, the generation unit 303 generates a prediction model M based on its model parameters θ (local learning method). Alternatively, it operates the prediction model M only at the location S without sending it to the server 100. Furthermore, for clusters containing multiple locations S, the generation unit 303 performs an integration process to generate one or more integrated prediction models based on their model parameters θ.

[0043] The output unit 304 includes a transmission unit 341 and a display unit 342.

[0044] The transmitting unit 341 sends to sites S1 to S10 an execution notification of a learning method selected from among a local learning method that generates one prediction model M using learning data D from a single site S, a federated learning method that generates one integrated prediction model M100 using learning data D from multiple sites S, and an individualized federated learning method that generates one or more integrated prediction models using learning data D from multiple sites S, along with the base prediction model M0.

[0045] Personalized Federated Learning (PFI) encompasses various learning methods. For example, one method focuses on smoothing out statistical heterogeneity between multiple training data sets D from multiple locations S to improve the convergence of the integrated predictive model, generating a single integrated predictive model based on multiple training data sets D. Another method focuses on improving the performance of the integrated predictive model at any level, such as each individual location S or clustered locations S, generating multiple integrated predictive models based on multiple predictive models.

[0046] Since individualized associative learning is a learning method included in associative learning, associative learning that is not individualized associative learning will be referred to as "non-individualized associative learning," and associative learning methods that are not individualized associative learning methods will be referred to as "non-individualized associative learning methods."

[0047] Furthermore, the transmission unit 341 transmits the integrated prediction model M100 generated by the generation unit 303 to the site S that generated the prediction model M used to generate the integrated prediction model. The transmission unit 341 also transmits a continuation instruction to each site S indicating whether to continue or terminate the federated learning iteration based on the accuracy verification results performed at each site S.

[0048] The display unit 342 displays data on a display, which is an example of an output device 204. For example, the display unit 342 displays the distribution analysis results from the distribution analysis unit 302.

[0049] Figure 4 is a block diagram showing an example of the functional configuration of base station S according to Embodiment 1. Base station S includes a receiving unit 401, a learning unit 402, a distribution conversion unit 403, and a transmitting unit 404. Specifically, the receiving unit 401, the learning unit 402, the distribution conversion unit 403, and the transmitting unit 404 realize their functions, for example, by having the processor 201 execute a program stored in the storage device shown in Figure 2, or by the communication IF 205.

[0050] The receiving unit 401 receives an execution notification of a learning method selected from a local learning method that generates one prediction model M at a single site S, a federated learning method that generates one integrated prediction model M100 at multiple sites S, and an individualized federated learning method that generates multiple integrated prediction models at multiple sites S, as well as the base prediction model M0.

[0051] Furthermore, after the generation of the prediction model M, the receiving unit 401 receives the integrated prediction model generated by the generation unit 303 during federated learning. The receiving unit 401 also receives a continuation instruction from the server 100 indicating whether to continue or terminate the repeated federated learning.

[0052] The learning unit 402 generates a prediction model M by providing training data D to a prediction model or base prediction model M0, which is composed of parameters determined by random initial values. Specifically, for example, the learning unit 402 obtains the base prediction model M0 from the server 100 and provides the training data D to the base prediction model M0. The learning unit 402 determines that the training data D1 is a combination of training data and ground truth data. Therefore, based on the prediction data output as a result of inputting the training data into the base prediction model M0 and the ground truth data, the learning unit 402 generates model parameters θ and a prediction model M by calculating a loss function such that the sum of squared errors or cross-entropy errors are minimized.

[0053] A predictive model M trained on the feature vector x, which is the training data within the training data D, is expressed by the output y (the predicted data), the model parameter θ, and the function h, as shown in equation (1) below. It is also possible to create a base predictive model M0 using equation (1) below.

[0054]

number

[0055] The distribution transformation unit 403 transforms the training data D at base S and performs a transformation process to change the data so that it can be moved outside of base S without problems, thereby generating transformed features TF. For example, if the training data D contains 100 samples, each with 5 features, the distribution transformation unit 403 takes the average in the sample direction and transforms it into a feature vector consisting of the average values ​​for each of the 5 features, which is then used as the transformed features TF. In addition to the mean, statistical measures such as the maximum value, minimum value, or variance may also be used.

[0056] Furthermore, if the training data is image data, the distribution transformation unit 403 may transform the image data into a feature vector by quantifying multiple features using predefined transformation rules. These predefined features could be, for example, the number of ears, limbs, and eyes of an animal if the image data is of an animal. Such transformation rules are stored as common rules in the server 100 and the base station S.

[0057] The transmission unit 404 transmits the transformed feature quantity TF generated by the distribution transformation unit 403, the model parameters θ, and the accuracy verification results of the integrated prediction model to the server 100.

[0058] Furthermore, the data transmitted and received by the receiving units 301, 401 and the transmitting units 341, 404 are converted and protected by encryption or other means for security reasons. This makes it difficult to analyze the data used for training the predictive model M.

[0059] <Similarity analysis of transformed feature quantities TF by distribution analysis unit 302> Figure 5 is an explanatory diagram showing example 1 of similarity analysis of transformed feature quantities TF by the distribution analysis unit 302. Figure 5 is an example in which the server 100 analyzes the similarity between locations S1 to S10 using the Euclidean distance between the transformed feature quantities TF received from location S. In the Euclidean distance heatmap 500, the darker the color, the closer the distance, meaning that the distribution of the transformed feature quantities TF is highly similar.

[0060] In this example, it can be seen that the distribution of transformed features TF for sites S1, S3, S5, and S9 is similar. It can also be seen that the distribution of transformed features TF for sites S4 and S10 is similar. Furthermore, it can be seen that the distribution of transformed features TF for sites S6 and S8 is similar. Based on this similarity in the distribution of transformed features TF, server 100 determines, for example, that sites S1, S3, S5, and S9 are clusters with similar distributions of transformed features TF, and that sites S6 and S8 are clusters with similar distributions of transformed features TF.

[0061] In Figure 5, the distance is represented using a heatmap of 500, meaning the intensity of the color indicates distance. However, distance can also be represented numerically. This map information is an example of information regarding the similarity of features between locations S.

[0062] Figure 6 is an explanatory diagram showing example 2 of similarity analysis of transformed feature quantities TF by the distribution analysis unit 302. Figure 6 is an example in which the server 100 analyzes the similarity of sites S1 to S10 by hierarchical clustering of transformed feature quantities TF received from site S.

[0063] Dendrogram 600 is created by hierarchical clustering of the transformed features TF. The horizontal axis of Dendrogram 600 represents the array of locations S according to the hierarchical clustering, and the vertical axis represents the distance calculated from the transformed features TF. The method of calculating the distance differs depending on the method of measuring the distance between clusters in the hierarchical clustering. Dendrogram 600 is also an example of information regarding the similarity of features between locations S.

[0064] According to Dendrogram 600, sites S1, S3, S5, and S9 belong to the same cluster early on (Cluster C1). Sites S4 and S10 also belong to the same cluster early on (Cluster C2). Sites S6 and S8 also belong to the same cluster early on.

[0065] Furthermore, the dendrogram 600 allows us to understand the similarity in the distribution of training data D for sites S6, S7, and S8, which is difficult to discern from the heatmap 500, as cluster C3, and the similarity in the distribution of training data D for sites S1, S3, S4, S5, S9, and S10, as cluster C12. The cluster for site S2 alone is designated as cluster C4.

[0066] Here, if we set the threshold for the similarity of the distributions of the training data D to 1.0, the 10 locations S1 to S10 can be divided into three clusters C12, C3, and C4. Of the three clusters C12, C3, and C4, cluster C4, which contains only one location S2, is expected to have better accuracy when using a single prediction model M2 compared to the integrated prediction model M100 generated by federative learning, based on the similarity of the distributions of the training data D. For clusters C12 and C3, which contain two or more locations S, the application of federative learning is considered appropriate.

[0067] However, it is thought that there is an appropriate associative learning method depending on the similarity of the training data D. Since sites S6 to S8 form a single cluster with a threshold of around 0.8, comfortably above 1.0, it is thought that an accurate predictive model M can be generated using associative learning. In addition, although sites S1, S3 to S5, S9, and S10 form a single cluster C12, they can be divided into cluster C2, which includes sites S1, S3, S5, and S9, and cluster C1, which includes sites S4 and S10. Thus, when there are large differences in the distribution of the training data D, it is also possible to apply individualized associative learning within the field of associative learning.

[0068] The distribution analysis unit 302 selects one of the following methods based on the results of Euclidean distance and hierarchical clustering: a local learning method that generates one prediction model M for a single site S; a non-individualized associative learning method that generates one integrated prediction model M100 for multiple sites S; or an individualized associative learning method that generates multiple integrated prediction models for multiple sites S.

[0069] This selection method could be, for example, a user of server 100 making a selection after referring to the heatmap 500 or dendrogram 600 displayed on the screen, or the user of server 100 could set a threshold in advance, and the distribution analysis unit 302 could select one of the local learning method, the non-individualized associative learning method, or the individualized associative learning method based on the threshold.

[0070] For example, in Figure 6, if the threshold is set to 1.0, sites S1 to S10 are classified into clusters C12, C3, and C4. In this case, since cluster C4 has only one site S2, the distribution analysis unit 302 determines that the learning method for site S2 within cluster C4 is the local learning method.

[0071] Furthermore, cluster C3 includes multiple locations S6 to S8, and forms a single cluster with a threshold of approximately 0.8, with a margin of safety above 1.0. Therefore, the distribution analysis unit 302 determines the learning method for locations S6 to S8 within cluster C3 to be a non-individualized associative learning method.

[0072] Furthermore, cluster C12 includes multiple locations S1, S3, S4, S5, S9, and S10, and within cluster C12, there are multiple clusters containing multiple locations (clusters C1 and C2). Clusters C1 and C2 form a single cluster around a threshold of 1.0, indicating a significant difference in the distribution of the training data D. Therefore, the distribution analysis unit 302 determines the learning method for locations S1, S3, S4, S5, S9, and S10 within cluster C12 to be an individualized associative learning method.

[0073] Furthermore, the distribution analysis unit 302 may also set constraints other than the threshold to select a learning method. For example, the distribution analysis unit 302 may vary the threshold and, when it reaches a predetermined number of clusters or the number of affiliated locations within a cluster, terminate the threshold variation and determine a learning method for the clusters at that point.

[0074] For example, the distribution analysis unit 302 sets the initial value of the threshold in the dendrogram 600 to the maximum value (for example, 2.0) and decreases it by a predetermined amount (for example, 0.1) from the maximum value. If the constraint condition is that the number of set clusters is "3", then when the threshold reaches 1.4, there are two clusters: cluster C4 and the cluster consisting of sites S1, S3~S10, so the number of clusters is "2". Therefore, the decrease in the threshold continues. When the threshold reaches 1.2, the clusters become clusters C12, C3, and C4. Since the number of clusters has reached "3", the decrease in the threshold ends, and the distribution analysis unit 302 determines the learning method for each of the clusters C12, C3, and C4 at this point.

[0075] Furthermore, if the constraint is that the number of locations within a cluster is "4 or less", then when the threshold reaches 1.4, there will be two clusters: cluster C4 and cluster S1, S3-S10, with the latter cluster having "9" locations. The threshold continues to decrease. When the threshold reaches 1.2, the clusters will be clusters C3, C4, and C12.

[0076] The latter of clusters C3, C4, and C12 has a total of "6" member locations. The threshold continues to decrease. When the threshold reaches 0.9, the clusters are clusters C1, C2, C3, and C4. Since cluster C1 has a total of "4" member locations, the threshold decreases, and the distribution analysis unit 302 determines the learning method for each of the clusters C1, C2, C3, and C4 at this point.

[0077] The above example is just one example, and the number of clusters can be set in various ways, such as "n" (where n is an integer greater than or equal to 1), "greater than or equal to n", "more than n", "less than or equal to n", and "less than n". Similarly, the number of locations within a cluster can also be set in various ways, such as "m" (where m is an integer greater than or equal to 1), "greater than or equal to m", "more than m", "less than or equal to m", and "less than m".

[0078] Furthermore, in the example described above, the explanation was divided into an example for the number of configured clusters and an example for the number of locations within one cluster. However, it is also possible to set both the number of configured clusters and the number of locations within one cluster as constraints, for example, the number of configured clusters being "n" and the number of locations within one cluster being "m". Also, in the example described above, the distribution analysis unit 302 decreased the threshold from the maximum value, but it is also possible to increase the threshold by a predetermined amount from the minimum value (for example, 0.0).

[0079] <Learning Example> Figure 7 is an explanatory diagram showing learning example 1 using server 100 and site S. In Figure 7, when cluster C12 is the cluster to be determined for the learning method, an example of performing individualized associative learning is shown in which learning data D1, D3~D5, D9, and D10 from cluster C1 consisting of sites S1, S3, S5, and S9 and cluster C2 consisting of sites S4 and S10 are used to generate a single integrated predictive model M700.

[0080] Each of the base stations S1, S3-S5, S9, and S10 receives a notification from the server 100 via the receiving unit 401 indicating that personalized associative learning has been selected. Each of the base stations S1, S3-S5, S9, and S10 generates prediction models M1, M3-M5, M9, and M10 using their respective learning data D1, D3-D5, D9, and D10 via the learning unit 402. Subsequently, each of the base stations S1, S3-S5, S9, and S10 transmits the model parameters θ1, θ3-θ5, θ9, and θ10 of the generated prediction models M1, M3-M5, M9, and M10 to the server 100 via the transmitting unit 404.

[0081] Server 100 receives model parameters θ1, θ3-θ5, θ9, and θ10 via the receiving unit 301. Server 100 uses the generation unit 303 to perform integration processing using the model parameters θ1, θ3-θ5, θ9, and θ10 to generate the integrated prediction model M700. Server 100 repeats the update process of the generated integrated prediction model M700 until it achieves the desired prediction accuracy.

[0082] In Figure 7, an example is shown in which personalized associative learning is performed to generate one integrated predictive model M700 using the training data D1, D3-D5, D9, and D10 from clusters C1 and C2, when cluster C12 is the cluster to be used for determining the learning method. However, it is also possible to perform personalized associative learning by generating one integrated predictive model using the training data D1, D3, D5, and D9 from cluster C1, and then generating one integrated predictive model using the training data D4 and D10 from cluster C2.

[0083] Figure 8 is an explanatory diagram showing learning example 2 using server 100 and site S. In Figure 8, when cluster C3 is the cluster to be used to determine the learning method, an example of non-individualized associative learning is performed to generate a single integrated predictive model using the learning data D3 of cluster C3, which consists of sites S6 to S8.

[0084] Each of the locations S6 to S8 receives a notification from the server 100 via the receiving unit 401 indicating that non-individualized associative learning has been selected. Each of the locations S6 to S8 generates a prediction model M6 to M8 using its respective training data D6 to D8 via the learning unit 402. Subsequently, each of the locations S6 to S8 transmits the generated prediction model M6 to M8 to the server 100 via the transmitting unit 404.

[0085] Server 100 receives model parameters θ6 to θ8 of prediction models M6 to M8 via the receiving unit 301. Server 100 uses the generation unit 303 to perform integration processing using the model parameters θ6 to θ8 to generate the integrated prediction model M800. Server 100 repeats the update process of the generated integrated prediction model M800 until it achieves the desired prediction accuracy.

[0086] In addition, while Figures 7 and 8 illustrate an example where the server 100 determines the learning method by referring to the dendrogram 600 and generates a prediction model M or an integrated prediction model, the server 100 may also determine the learning method by referring to the heatmap 500, clustering multiple locations S, and generating a prediction model M or an integrated prediction model.

[0087] <Integration process of predictive model M> Next, the integration process of the prediction model M by the generation unit 303 will be explained in detail. In the integration process of the prediction model M, the model parameters θ are added to the base prediction model M0 through learning at K locations S (where K is an integer greater than or equal to 1). t Using the integrated predictive model M100 (where t is the number of updates to the integrated predictive model M100) as the initial value, each of the K locations S learns using its own training data D.

[0088] Then, the generation unit 303 generates the model parameters θ of the K prediction models M that each of the K bases S has generated. k Gradient g k Obtain the gradient g k Using the sum of the averages, the model parameter θ of the integrated prediction model M100 at the t+1th update is as shown in equation (2) below. t+1 This generates the following: N is the total number of samples in the training data D used for training at K locations S. k This is the number of samples in the training data D at site k.

[0089]

number

[0090] Here, in the above formula (2), the gradient g k with respect to the model parameter θ k is used. This is a method that takes security considerations so that the learning data D cannot be analyzed from the model parameter θ k , and model parameters θ k such as encoding and encryption may also be used. Further, the generation unit 303 may integrate the prediction models M1 to M10 in a method different from the above formula (2) according to the structure of the prediction model such as a fully connected layer or a convolutional layer.

[0091] For example, as shown in the following formula (3), it may be the average value normalized by the number of samples of the learning data D used for learning each model parameter θ k , or the model parameter θ k and its gradient g k in units of batches or epochs in the learning process of the prediction model M may also be used.

[0092]

Number

[0093] <Pre - processing for integrated learning by server 100> FIG. 9 is a flowchart showing an example of the pre - processing procedure for integrated learning by server 100. It is assumed that server 100 can communicate with the site S wishing to participate in federated learning. Server 100 receives the transformed feature amount TF from each site S through the receiving unit 301 (step S901). Next, server 100 uses the received transformed feature amount TF to execute an analysis of the similarity of the learning data D by the distribution analysis unit 302 (step S902). [[ID=3,6]]

[0094] Then, server 100 determines the learning method for each site S based on the similarity of the distribution of the learning data D as shown in FIGS. 5 and 6 by the distribution analysis unit 302 (step S903). After that, server 100 outputs the distribution analysis result by the distribution analysis unit 302 through the output unit 304 (step S904).

[0095] Specifically, for example, the server 100 displays the distribution analysis results, including the heatmap 500, the dendrogram 600, and the learning method for the federated prediction model M determined in step S903, that is, the method for generating the integrated prediction model M100 and the learning method for the prediction model M, via the display unit 342. The server 100 also notifies each site S of the learning method determined in step S903 via the transmission unit 341. This completes the integrated learning preprocessing.

[0096] <Pre-processing by site S> Figure 10 is a flowchart showing an example of the pre-training processing procedure performed by base station S. Base station S is assumed to have already obtained the transformation methods and tools for the features, which are the training data in training data D, from server 100.

[0097] Station S uses a feature transformation method and tools to transform the features that are the training data in the training data D, and generates transformed features TF (step S1001). Next, Station S sends the transformed features TF to Server 100 (step S1002). After this, Station S receives the notification in step S904 from Server 100 (step S1003). This completes the pre-training processing.

[0098] Thus, according to the data analysis processing system described above, it is possible to determine the appropriate training method for the prediction model M for each location S, without moving the training data D1 to D10 located at multiple locations S1 to S10 outside of the locations S, and to generate an appropriate prediction model M or integrated prediction models M100, M700, and M800 for each location S.

[0099] <Display screen example> Next, we will describe an example of a display screen shown on a display, which is an example of an output device 204 of the computer 200.

[0100] Figure 11 is an explanatory diagram showing example 1 of the display screen. The display screen 1100 is displayed, for example, on the display of the server 100. The display screen 1100 includes a View clients button 1101, a View results button 1102, a mode field 1103, a site list 1111, site analysis results 1112, a site classification start button 1113, and a site classification result confirmation button 1114.

[0101] Server 100 accepts the user's input, specifically the selection of "Analysis" in the mode field 1103 and the press of the View clients button 1101. As a result, Server 100 displays the list of locations 1111 that wish to participate in federated learning in the location list 1111. When Server 100 accepts the user's input, specifically the press of the Start Location Classification button 1113, Server 100 sends the methods and tools necessary for generating transformed features to each location S.

[0102] When transformed features are generated at each site S, they are sent to the server 100, and the server 100 performs analysis of the received transformed features using the distribution transformation unit 403. After this, when the server 100 receives a press of the View results button 1102 by the user, the server 100 displays the site analysis results 1112. At the same time, the assignment of each learning method to each site S is also reflected in the site list 1111.

[0103] When the user confirms the site list 1111 and the site analysis results 1112, and the server 100 receives confirmation of the site classification result confirmation button 1114, it sends a notification of the learning method assigned to each site S. Here, the learning method for each site S in the site list 1111 may be individually edited directly by the user.

[0104] Figure 12 is an explanatory diagram showing example 2 of the display screen. The display screen 1200 is displayed, for example, on the display of the server 100. The display screen 1200 includes a View clients button 1101, a View results button 1102, a mode field 1103, a site list 1111, a group selection field 1211, federated learning results 1212, and a federated learning results confirmation button 1213.

[0105] Server 100 accepts the user's input to select "Feseration" in the mode field 1103 and displays the federated learning results 1212 when the user presses the "View results" button 1102. Server 100 also accepts the user's input to select a learning group in the group selection field 1211 and displays the federated learning results 1212 for the bases S belonging to the learning group displayed in the base list 1111. If the iteration process in federated learning has completed a specified number of times or achieved the desired prediction accuracy, Server 100 terminates the federated learning process when the user presses the "Confirm Federation Learning Results" button 1213. If, after receiving the federated learning results, the user wishes to re-select bases or repeat the learning process, they can change the mode field 1103 to "Analysis," return to the display screen 1100, and perform the transformation feature TF generation, learning method determination, and federated learning again.

[0106] Thus, according to Example 1, it is possible to generate an appropriate integrated predictive model M100 for each location S, based on the diverse training data D at each location S, without moving the training data D located at multiple locations S outside of the locations S. Furthermore, although the number of locations was set to 10 in Example 1, it may be 9 or less, or 11 or more, as long as there are 2 or more locations. [Examples]

[0107] Example 2 will now be described. In Example 1, the appropriate learning method for each location S was determined by analyzing the transformed features of each location S. Example 2 is an example in which the distribution analysis unit 302 determines the appropriate learning method for each location S using associative learning. Note that the integrated prediction model described in Example 2 is not the integrated prediction model generated by the generation unit 303 in Example 1, but an integrated prediction model for determining the learning method. In Example 2, the explanation will focus on the differences from Example 1, so the explanation of the similarities with Example 1 will be omitted.

[0108] For example, if the training data D for each base S is the stage classification of cancer patients in each hospital, the output of the final integrated predictive model will similarly be the stage classification. In Example 2, when the appropriate training method for each base S is determined by associative learning, the output of the final integrated predictive model will be which base S's training data D it represents (i.e., the base S to which the training data D belongs).

[0109] Figure 13 is an explanatory diagram illustrating a federated learning method 1 that enables appropriate learning for each of multiple locations S where the training data D cannot be sent outside of location S. If the multiple locations S consist only of locations S1 and S2, location S1 generates a classifier, which is the prediction model M1, using the base prediction model M0 and the training data D1, and sends the model parameters θ1 of this classifier to the server 100. Similarly, location S2 generates a classifier, which is the prediction model M2, using the base prediction model M0 and the training data D2, and sends the model parameters θ2 of this classifier to the server 100. Based on the model parameters θ1 and θ2 of the prediction models M1 and M2, which were generated with the class where the training data D is from location S1 being 0 and the class where the training data D is from location S2 being 1, the server 100 generates a two-class classification classifier (hereinafter referred to as the integrated classifier), which is the integrated prediction model M100, through federated learning.

[0110] Locations S1 and S2 receive the generated integrated classifier (integrated predictive model M100, which is a two-class classification classifier) ​​from server 100, apply it, and calculate the prediction probability. Locations S1 and S2 send the calculated prediction probability, or the value obtained by converting the prediction probability into a propensity score, to server 100 as a converted feature. In Figure 13, this is shown as a propensity score.

[0111] Server 100 compares the box plot 1301 of the propensity score for site S1 with the box plot 1302 of the propensity score for site S2, as shown in graph 1300 in Figure 13. For example, if the probability of being site S1 is high at site S1 and the probability of being site S1 is low at site S2, it means that the training data D1 and D2 from the two sites S1 and S2 are easily distinguishable, that is, the similarity of the data distributions is low. In this case, the overlap range 1312 of box plots 1301 and 1302 becomes narrower, or the number of samples included decreases.

[0112] Conversely, if the probability of a location being location S1 is similar to the probability of a location being location S1 at location S2, then the training data D1 and D2 from the two locations S1 and S2 are difficult to distinguish, meaning that the data distributions are highly similar. In this case, the overlap range 1312 of box plots 1301 and 1302 becomes wider, or the number of samples included increases. Server 100 can evaluate the similarity of the training data D1 and D2 from both locations S1 and S2 by comparing the distribution of predicted probabilities when applying the integrated prediction model for location classification at locations S1 and S2.

[0113] Here, propensity score refers to a value obtained by a balance adjustment statistical method or statistical method used to estimate causal effects by adjusting for covariates in observational studies where random assignment is difficult and various confounding factors are likely to occur. In Example 2, in the characterization of the training data D at each site S, the predicted probability may be used as is, or a weighting method using propensity scores, such as the IPW (Inverse Probability Weighting) estimator or a double-robust estimator, may be used.

[0114] In the example shown in Figure 13, server 100 determines the similarity between training data D1 and D2 based on the size of the overlapping range 1312 of box plots 1301 and 1302. For example, if the overlapping range 1312 is greater than or equal to a threshold, server 100 determines that training data D1 and D2 have similar data distributions. If similarity is determined, server 100 assigns sites S1 and S2 to the same cluster. In addition to the size of the overlapping range 1302, the similarity of training data D1 and D2 may also be determined using the number of samples included in the overlapping range 1302.

[0115] Furthermore, while Figure 13 shows an example with two locations, the fewer locations there are, the easier it becomes for server 100 to estimate the training data D of other locations S from the model parameters of the integrated prediction model. For this reason, measures such as not informing locations S of the number of participating locations, limiting its use to cases where the number of participating locations exceeds a certain number, or using privacy protection techniques such as Differential Privacy Stochastic Gradient Descent SGD (DPSGD) or Private Aggregation of Teacher Ensembles (PATE) should be taken.

[0116] Figure 14 is an explanatory diagram illustrating federated learning method 2, which enables appropriate learning for each of multiple locations S where the training data D cannot be sent outside of location S. Figure 14 is a graph 1400 in particular when there are three or more locations. When the multiple locations S are locations S1, S2, and S3, location S1 generates a classifier which is the prediction model M1 using the base prediction model M0 and the training data D1, and sends the model parameters θ1 of this classifier to the server 100. Similarly, location S2 generates a classifier which is the prediction model M2 using the base prediction model M0 and the training data D2, and sends the model parameters θ2 of this classifier to the server 100. Similarly, location S3 generates a classifier which is the prediction model M3 using the base prediction model M0 and the training data D3, and sends the model parameters θ3 of this classifier to the server 100.

[0117] Server 100 generates a 3-class classifier (integrated classifier), which is an integrated prediction model M100, by federative learning, based on the model parameters θ1 to θ3 of the prediction models M1 to M3, which are generated with class 0 representing the class where the training data D is at base S1, class 1 representing the class where the training data D is at base S2, and class 2 representing the class where the training data D is at base S3. Generally, propensity scores are used when data is divided into two groups, and when there are three or more groups, a generalized propensity score is used.

[0118] The generalized propensity score can be expressed using the propensity score PS(k|x) at the k-th location S (location k) for K different training data D at K locations S, as shown in equation (4) below. Equation (4) is designed so that the sum of the K generalized propensity scores equals 1.

[0119]

number

[0120] In the example shown in Figure 14, server 100 determines the similarity of training data D1 and D2 based on the size of the overlapping range 1412 of box plots 1401 and 1402, the similarity of training data D2 and D3 based on the size of the overlapping range 1423 of box plots 1402 and 1403, and the similarity of training data D1 and D3 based on the size of the overlapping range 1313 of box plots 1401 and 1403.

[0121] For example, if the overlapping range 1412 is greater than or equal to a threshold, server 100 determines that the training data D1 and D2 have similar data distributions. If similarity is determined, server 100 assigns sites S1 and S2 to the same cluster. The same determination is made for the other overlapping ranges 1423 and 1413. Similar to associative learning method 1, the similarity of training data D1 to D3 may be determined not only by the size of the overlapping ranges 1412, 1423, and 1413, but also by the number of samples included in the overlapping ranges 1412, 1423, and 1413.

[0122] In this embodiment, propensity scores calculated using two-class and three-class classifiers generated in advance by federative learning were used. When generating two-class and three-class classifiers by federative learning, each site S generates a prediction model M using training data D. Since site S generates a prediction model M using only its own training data D, the prediction model M is prone to falling into local optima. When server 100 performs integration processing on the model parameters θ of the prediction model M, which can predict only the training data D, the model parameters of the integrated prediction model shift in a direction that diverges with respect to the training data D held by the sites S that participated in the federative learning. In other words, an integrated prediction model with low prediction accuracy is generated.

[0123] To avoid this, first, as the output of the final integrated prediction model (for example, cancer stage), server 100 executes the integrated prediction model generation process in federated learning. However, the number of updates may be insufficient. Next, keeping the model parameters θ of the intermediate layer of the generated integrated prediction model unchanged, and using the output of the base prediction model M0 as the number of locations S, server 100 generates a class classifier using federated learning. By performing federated learning based on such transfer learning, it becomes possible to use the output of the intermediate layer that can extract features of the training data D of each location S, and it is possible to avoid falling into a local minimum when generating the prediction model M at each location S, due to only including the training data D of one location S.

[0124] Thus, according to Example 2, by comparing the distribution of the calculated generalized propensity scores, it becomes possible to evaluate the similarity of the training data D at site S. [Examples]

[0125] Example 3 will now be described. Example 3 is an example in which the equipment configuration of server 100 and site S is standardized compared to Examples 1 and 2. In Example 3, the explanation will focus on the differences from Examples 1 and 2, and the similarities with Examples 1 and 2 will be omitted.

[0126] Figure 15 is a block diagram showing an example of the functional configuration of a computer 1500 responsible for at least one of server 100 or site S. The computer 1500 responsible for at least one of server 100 or site S includes a receiving unit 1501, a distribution analysis unit 302, a generation unit 303, a learning unit 402, a distribution transformation unit 403, and an output unit 1504. The receiving unit 1501 functions as receiving units 301 and 401. The output unit 1504 includes a transmitting unit 1541 and a display unit 342. The transmitting unit 1541 functions as transmitting units 341 and 404.

[0127] Even if computer 1500 is server 100, it may participate in federative learning as one site using the training data D within computer 1500. Also, for example, in federative learning conducted at sites S6 to S8, site S6 may play the role of server 100 when generating the integrated learning model.

[0128] As explained above, even if the training data D is non-iid data, it is possible to generate an appropriate prediction model M at each site S using the similarity relationship between the transformed features TF, without having to determine whether the training data D that cannot be left site S is non-iid data or not. Through federative learning, server 100 can generate an appropriate integrated prediction model.

[0129] Furthermore, since it is not necessary to determine whether the training data D is non-iid data or not, the server 100 can provide the base station S with an integrated prediction model that satisfies the requirement that the training data D cannot be taken outside of base station S.

[0130] Furthermore, it is unnecessary to determine whether the training data D is non-iid data or not. Even if the training data D is iid data, the server 100 can generate an appropriate integrated predictive model, just as it would in the case of non-iid data.

[0131] Furthermore, the analytical device, which is the server 100 in the above-described Examples 1 and 2, can also be configured as shown in (1) to (13) below.

[0132] (1) An analysis device capable of communicating with learning devices which are multiple locations S includes a receiving unit 301 that receives transformed feature quantities TF obtained by transforming the feature quantities (training data) contained in the learning data D of each of the multiple learning devices according to predetermined rules, a distribution analysis unit 302 that analyzes the distribution of multiple feature quantities of the multiple learning devices based on the transformed feature quantities TF for each learning device received by the receiving unit 301, and an output unit 304 that outputs the distribution analysis results analyzed by the distribution analysis unit 302.

[0133] (2) In the analysis apparatus described in (1) above, the output unit 304 outputs information regarding the similarity of the feature quantities between the learning devices as the distribution analysis result.

[0134] (3) In the analysis apparatus described in (1) above, the information relating to the similarity of the feature quantities between the learning devices is map information showing the similarity of the feature quantities between each combination of two of the learning devices from the plurality of learning devices.

[0135] (4) In the analysis apparatus described in (1) above, the information relating to the similarity of the feature quantities between the learning devices is a dendrogram 600 showing the similarity of the feature quantities between the plurality of learning devices.

[0136] (5) In the analysis apparatus described in (1) above, the distribution analysis unit 302 determines, based on the distribution analysis results, which learning method to apply to each of the plurality of learning devices to one of the following learning methods: a first learning method (local learning method) that generates a prediction model M using the learning data held by the learning device; a federated learning method (non-individualized federated learning method) that generates one integrated prediction model M100 through federated learning between other learning devices (cluster C3); and a third learning method (individualized federated learning method) that generates one or more prediction models integrated through federated learning between other learning devices whose features are similar to those of the learning device. The output unit 304 then transmits to each of the plurality of learning devices the learning method determined for each of the plurality of learning devices by the distribution analysis unit 302.

[0137] (6) In the analysis apparatus described in (5) above, the distribution analysis unit 302 determines, based on the distribution analysis results, which learning method to apply to each of the plurality of learning devices to be one of the first learning method, the second learning method, and the third learning method.

[0138] (7) In the analysis apparatus described in (6) above, the distribution analysis unit 302 determines which learning method to apply to each of the plurality of learning devices is one of the first learning method, the second learning method, and the third learning method, based on the distance between the transformed feature quantities TF of the plurality of learning devices.

[0139] (8) In the analysis apparatus described in (6) above, the distribution analysis unit 302 determines which learning method to apply to each of the plurality of learning devices is one of the first learning method, the second learning method, and the third learning method, based on the transformed feature quantity TF and the threshold.

[0140] (9) In the analysis apparatus described in (8) above, the distribution analysis unit 302 determines which learning method to apply to each of the plurality of learning devices is one of the first learning method, the second learning method, and the third learning method, based on the transformed feature quantity TF, the threshold, and the constraint conditions.

[0141] (10) In the analysis apparatus described in (9) above, the constraint is the number of learning devices whose features are similar.

[0142] (11) In the analysis apparatus of (9) above, the constraint is the number of sets of learning devices whose features are similar.

[0143] (12) In the analysis apparatus of (1) above, the receiving unit 301 receives model parameters of a classifier that identifies a learning device, which is generated by training each of the plurality of learning devices with the learning data; the distribution analysis unit integrates the classifiers of the learning devices based on the model parameters of the classifiers of the learning devices received by the receiving unit to generate an integrated classifier; the transmitting unit transmits the model parameters of the integrated classifier generated by the generation unit to each of the plurality of learning devices; the receiving unit receives the identification results from the integrated classifier from each of the plurality of learning devices; and the distribution analysis unit 302 integrates the classifiers of the learning devices based on the identification results from each of the learning devices received by the receiving unit 301 to generate an integrated classifier. The learning method to be applied to each of the learning devices is determined to be one of the following: a first learning method (local learning method) that generates a prediction model M using the learning data held by the learning device; a federated learning method (non-individualized federated learning method) that generates one integrated prediction model M100 through federated learning between the learning device and other learning devices similar to the features of the learning device (cluster C3); and a third learning method (individualized federated learning method) that generates one or more integrated prediction models through federated learning of multiple groups consisting of other learning devices similar to the features of the learning device. The output unit 304 then transmits the learning method determined for each of the multiple learning devices by the distribution analysis unit 302 to each of the multiple learning devices.

[0144] (13) In the analysis apparatus described in (5) above, the receiving unit 301 receives the model parameters θ of the prediction model M generated by the learning device based on the learning method determined by the distribution analysis unit 302 from the learning device, and has a generating unit 303 that generates a prediction model based on the model parameters from the learning device received by the receiving unit 301.

[0145] It should be noted that the present invention is not limited to the embodiments described above, but includes various modifications and equivalent configurations within the spirit of the attached claims. For example, the embodiments described above are described in detail to make the present invention easier to understand, and the present invention is not necessarily limited to having all of the described configurations. Furthermore, some of the configurations of one embodiment may be replaced with those of another embodiment. Furthermore, some of the configurations of one embodiment may be added to those of another embodiment. Furthermore, some of the configurations of each embodiment may be added, deleted, or replaced with other configurations.

[0146] Furthermore, each of the aforementioned configurations, functions, processing units, and processing means may be implemented in hardware, for example, by designing them as integrated circuits, or they may be implemented in software by having a processor interpret and execute programs that realize each function.

[0147] Information such as programs, tables, and files that implement each function can be stored in memory, hard disks, SSDs (Solid State Drives), or on recording media such as IC (Integrated Circuit) cards, SD cards, and DVDs (Digital Versatile Discs).

[0148] Furthermore, the control lines and information lines shown are those deemed necessary for explanation purposes and do not necessarily represent all control lines and information lines required for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of Symbols]

[0149] 100 servers 200 Computers 301 Receiver 302 Distribution analysis section 303 Generation part 304 Output section 313 Distribution Transformation Unit 341 Transmitter 342 Display section 401 Receiver Room 402, Learning Department 403 Distribution Transformation Unit 404 Transmitter 500 Heatmaps 600 Dendrograms 1500 calculator 1501 Receiver 1504 Output section 1541 Transmitter

Claims

1. An analysis device capable of communicating with a plurality of learning devices, a receiving unit that receives transformed features obtained by transforming features included in learning data held by each of the plurality of learning devices according to a predetermined rule; a distribution analysis unit that analyzes a distribution of the feature amounts of the plurality of learning devices based on the converted feature amount for each of the learning devices received by the receiving unit; an output unit that outputs a distribution analysis result analyzed by the distribution analysis unit; An analytical device comprising:

2. The analysis device according to claim 1 , the output unit outputs information regarding a similarity of the feature amounts between the learning devices as the distribution analysis result.

1. An analytical device comprising:

3. The analysis device according to claim 1 , the information regarding the similarity of the feature between the learning devices is map information indicating, for each combination of two of the plurality of learning devices, the similarity of the feature between the combinations; 1. An analytical device comprising:

4. The analysis device according to claim 1 , the information regarding the similarity of the feature between the learning devices is a dendrogram indicating the similarity of the feature between the plurality of learning devices.

1. An analytical device comprising:

5. The analysis device according to claim 1 , the distribution analysis unit determines, based on a result of the distribution analysis, a learning method to be applied to each of the plurality of learning devices to one of a first learning method of generating a predictive model using the learning data held by the learning device, a second learning method of generating one predictive model integrated by federated learning with other learning devices, and a third learning method of generating one or more predictive models integrated by federated learning with other learning devices similar to the feature amount of the learning device; the output unit transmits to each of the plurality of learning devices the learning method determined for each of the plurality of learning devices by the distribution analysis unit.

1. An analytical device comprising:

6. The analysis device according to claim 5 , the distribution analysis unit determines a learning method to be applied to each of the plurality of learning devices to be one of the first learning method, the second learning method, and the third learning method based on a result of the distribution analysis; 1. An analytical device comprising:

7. The analysis device according to claim 6, the distribution analysis unit determines a learning method to be applied to each of the plurality of learning devices to be one of the first learning method, the second learning method, and the third learning method based on distances between the transformation features of the plurality of learning devices; 1. An analytical device comprising:

8. The analysis device according to claim 6, the distribution analysis unit determines a learning method to be applied to each of the plurality of learning devices to be one of the first learning method, the second learning method, and the third learning method based on the conversion feature and a threshold value; 1. An analytical device comprising:

9. The analysis device according to claim 8, the distribution analysis unit determines a learning method to be applied to each of the plurality of learning devices to be one of the first learning method, the second learning method, and the third learning method based on the converted feature, a threshold value, and a constraint condition; 1. An analytical device comprising:

10. The analysis device according to claim 9, the constraint condition is the number of the learning devices having similar features; 1. An analytical device comprising:

11. The analysis device according to claim 9, The constraint condition is the number of sets of the learning devices having similar features.

1. An analytical device comprising:

12. The analysis device according to claim 1 , the receiving unit receives model parameters of a classifier for identifying the learning device, the model parameters being generated by learning the learning data held by each of the plurality of learning devices; a generation unit that generates an integrated classifier by integrating the classifiers of the learning devices based on the model parameters of the classifiers of the learning devices received by the receiving unit; a transmission unit that transmits the model parameters of the integrated classifier generated by the generation unit to each of the plurality of learning devices; the receiving unit receives a classification result by the integrative classifier from each of the plurality of learning devices; the distribution analysis unit determines, based on the identification results from each of the learning devices received by the receiving unit, a learning method to be applied to each of the multiple learning devices, from among a first learning method that generates a predictive model using the learning data held by the learning device, a federated learning method that generates one predictive model integrated through federated learning with other learning devices similar to the feature amount of the learning device, and a third learning method that generates one or more predictive models integrated through federated learning with other learning devices similar to the feature amount of the learning device; the output unit transmits to each of the plurality of learning devices the learning method determined for each of the plurality of learning devices by the distribution analysis unit.

1. An analytical device comprising:

13. The analysis device according to claim 5 , the receiving unit receives, from the learning device, model parameters of the prediction model generated by the learning device based on the learning method determined by the distribution analysis unit; a generation unit that generates a prediction model based on the model parameters from the learning device received by the receiving unit; An analytical device comprising:

14. An analysis method using an analysis device capable of communicating with a plurality of learning devices, comprising: The analysis device comprises: a receiving process for receiving transformed features obtained by transforming features included in learning data held by each of the plurality of learning devices according to a predetermined rule; a distribution analysis process of analyzing a distribution of the plurality of feature amounts of the plurality of learning devices based on the converted feature amount for each of the learning devices received by the receiving process; an output process for outputting a distribution analysis result analyzed by the distribution analysis process; An analysis method comprising the steps of:

15. A processor of an analysis device capable of communicating with a plurality of learning devices, a receiving process for receiving transformed features obtained by transforming features included in learning data held by each of the plurality of learning devices according to a predetermined rule; a distribution analysis process of analyzing a distribution of the plurality of feature amounts of the plurality of learning devices based on the converted feature amount for each of the learning devices received by the receiving process; an output process for outputting a distribution analysis result analyzed by the distribution analysis process; An analysis program characterized by executing the above.

Images