Analytical device, analytical method, and analytical 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 analysis device, an analysis method, and an analysis program for analyzing data. [Background technology]

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

[0003] To date, methods have been devised for calculating and computing learning parameters to maximize the accuracy of predicted values. For example, a method called a perceptron outputs a predicted value based on the input feature vector and the linear combination of a weight vector. Neural networks, also known as multiperceptrons, have the ability to solve linearly inseparable problems by stacking multiple perceptrons in layers. Deep learning is a method that introduces new techniques such as dropout into neural networks and has attracted attention as a method capable of achieving high prediction accuracy. As such, machine learning technologies have been developed with the aim of improving prediction accuracy, and their prediction accuracy is beginning to exceed that of humans.

[0004] When implementing machine learning technology in society, there are challenges other than prediction accuracy, such as security, how to update models after delivery, and limitations on the use of finite resources such as memory.

[0005] One security issue is data confidentiality. When creating predictive models using data that includes personal information, such as in the medical or financial fields, it can be difficult to move the data outside of the location where it is stored due to the high level of confidentiality of the data. Generally, machine learning can achieve high prediction accuracy by using large amounts of data for training.

[0006] When training is performed using only data collected at a single location, the model may be usable only in a very localized area due to the small number of data samples, the characteristics of the area where the data was collected, etc. In other words, machine learning technology is needed that makes it possible to create predictive models that can achieve high predictions using the wide variety of data from each location, without having to take the data out of the location.

[0007] Patent Document 1 below discloses a method and system with a federated learning model for healthcare applications. The federated learning system comprises a plurality of 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 of multiple client terminals with a master model on an integration server before each of the multiple client terminals starts learning. Each client terminal performs machine learning on the learning model using data stored at 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 an accuracy threshold, it extracts the client terminal that is causing the accuracy degradation from the client cluster used to create the 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 uses one common model as the initial value and performs learning using data from each location to generate a predictive model. The model parameter information of the generated predictive model is sent to a server. The server then repeatedly executes a process to generate a global predictive model from the model parameter information of the predictive model using coefficients according to the amount of data learned. Ultimately, a global predictive model that achieves high prediction accuracy for data from all locations is generated. [Prior art documents] [Patent documents]

[0010] [Patent Document 1] U.S. Patent Publication No. 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. Summary of the Invention [Problem to be solved by the invention]

[0012] The method and system of Patent Document 1 smooth out collective biases for each end user, and therefore do not take into consideration the characteristics of each end user when generating a prediction model.

[0013] The machine learning system in Patent Document 2 identifies locations that cannot achieve the desired prediction accuracy and performs federated learning again at locations excluding those locations. However, repeatedly performing federated learning means that data that has already been learned once must be re-learned multiple 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 region where the data was collected, a global prediction model with degraded predictive performance may be generated for data from any location.

[0015] The present invention aims to make it possible to generate a prediction model appropriate for each location without moving data stored at multiple locations outside the location. [Means for solving the problem]

[0016] An analysis device according to one aspect of the invention disclosed in the present application is an analysis device capable of communicating with a plurality of learning devices, the analysis device including: 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 a plurality of the features of the plurality of learning devices based on the transformed features for each of the learning devices received by the receiving unit; and a distribution analysis result analyzed by the distribution analysis unit. map information indicating the similarity of feature quantities between each combination of two of the plurality of learning devices, and an output unit that outputs the signal. [Effects of the Invention]

[0017] According to a representative embodiment of the present invention, it is possible to generate a predictive model appropriate for each site without moving data stored at multiple sites outside the site. Problems, configurations, and effects other than those described above will become clear from the description of the following examples. [Brief explanation of the drawings]

[0018] [Figure 1] FIG. 1 is an explanatory diagram showing an example of associative 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 of an example of a functional configuration of the server according to the first embodiment. [Figure 4] FIG. 4 is a block diagram of an example of a functional configuration of a base according to the first embodiment. [Figure 5] FIG. 5 is an explanatory diagram showing a first example of similarity analysis of transformed feature amounts performed by the distribution analysis unit. [Figure 6] FIG. 6 is an explanatory diagram showing a second example of similarity analysis of transformed feature quantities by the distribution analysis unit. [Figure 7] FIG. 7 is an explanatory diagram showing learning example 1 using a server and a base. [Figure 8] FIG. 8 is an explanatory diagram showing a learning example 2 using a server and a base. [Figure 9] FIG. 9 is a flowchart illustrating an example of a procedure of a pre-processing for integrated learning performed by the server. [Figure 10] FIG. 10 is a flowchart illustrating an example of a pre-learning process procedure performed by a base. [Figure 11] FIG. 11 is an explanatory diagram showing a display example 1 of the display screen. [Figure 12] FIG. 12 is an explanatory diagram showing a display example 2 of the display screen. [Figure 13] FIG. 13 is an explanatory diagram showing federated learning method 1, which realizes learning appropriate for each site among a plurality of sites where learning data cannot be sent outside the site. [Figure 14] FIG. 14 is an explanatory diagram showing a federated learning method 2 that realizes learning appropriate for each site among a plurality of sites where learning data cannot be sent outside the site. [Figure 15] FIG. 15 is a block diagram showing an example of the functional configuration of a computer. DETAILED DESCRIPTION OF THE INVENTION

[0019] The analysis device converts the data characteristics of each location without moving the data stored at multiple locations outside the location, moves it outside the location, and analyzes the data characteristics of each location. As a result, the analysis device can present a model building method and grouping appropriate for each location, (A) Display the clustering results using the distance between transformed features, which are obtained by transforming the features contained in the learning data of each base. (B) Clustering results are displayed by calculating scores using federated learning.

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

[0021] <Associative learning> Fig. 1 is an explanatory diagram showing an example of federated learning. Each of multiple locations (e.g., 10 locations S1 to S10 in Fig. 1) that are learning devices holds learning data D1 to D10 (when not distinguishing between them, simply referred to as learning data D), and is prohibited from sending the learning data D1 to D10 outside the locations S1 to S10.

[0022] The server 100 is an analysis device that integrates prediction models M1 to M10 (when not distinguishing between them, simply referred to as the prediction model M) generated at the locations S1 to S10 (when not distinguishing between them, simply referred to as the location S). The 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 may be a trained neural network in which model parameters such as weights and biases are set.

[0023] The bases S1 to S10 are computers that have training data D1 to D10 and generate prediction models M1 to M10 using the training data D1 to D10. Each of the training data D1 to D10 is a combination of training data and correct answer data as input.

[0024] First, the server 100 transmits the base prediction model M0 to the locations S1 to S10. The locations S1 to S10 perform learning using the base prediction model M0 and respective learning data D1 to D10, and generate prediction models M1 to M10.

[0025] Sites S1 to S10 transmit model parameters θ1 to θ10 (when not distinguishing between them, simply referred to as model parameter θ) such as weights and biases of prediction models M1 to M10, respectively, to server 100. Server 100 executes an integration process on the received model parameters θ1 to θ10 to generate integrated prediction model M100.

[0026] The server 100 repeats the update process for the generated integrated prediction model M100 until the generated integrated prediction model M100 achieves the desired prediction accuracy. Note that the sites S1 to S10 may transmit the gradients of the model parameters θ1 to θ10 of the prediction models M1 to M10 to the server 100. The sites S1 to S10 transmit the model parameters θ1 to θ10, such as the weights and biases of the prediction models M1 to M10, to the server 100, respectively.

[0027] The integration process is a process of calculating the average value of the model parameters θ1 to θ10. If the number of samples of the training data D1 to D10 is different, a weighted average may be calculated based on the number of samples of the training data D1 to D10. Furthermore, the integration process may be a process of calculating the average value of each gradient of the model parameters θ1 to θ10 transmitted from each of the locations S1 to S10 instead of the model parameters θ1 to θ10.

[0028] The update process for the integrated prediction model M100 is a process in which the server 100 transmits the integrated prediction model M100 to the sites S1 to S10, the sites S1 to S10 input training data D1 to D10 into the integrated prediction model M100, respectively, and the sites S1 to S10 perform training. The server 100 then transmits the model parameters θ1 to θ10 of the regenerated prediction models M1 to M10 to the server 100, and the server 100 regenerates the integrated prediction model M100. When the generated integrated prediction model M100 achieves a desired prediction accuracy, the federated learning ends. Alternatively, a predetermined number of updates may be used instead of a desired prediction accuracy.

[0029] <Non-independent identical distribution> In general, it is assumed that the data given to machine learning is independently and identically distributed (iid). Independent and identically distributed means that, for example, when you roll a die with numbers 1 to 6, the probability of a certain number appearing is the same, and when you roll the die multiple times, the results will be independent regardless of which number appears on which roll. When this definition breaks down, it becomes non-independent and identically distributed (non-iid).

[0030] Data under non-iid conditions is called data that has undergone data shift or data skew, or non-iid data. For example, data from a hospital in Japan and data from a hospital in the United States are expected to have significantly different distributions of patients' physical condition, 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 variations in characteristics can be classified into four main types depending on the type of variation: 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 objective variables, differences in the conditional probabilities of these two, etc. Although the types of non-iid data are statistically defined, multiple types may occur in actual data, making it difficult to evaluate which types occur at what intensity.

[0033] If training and testing data are separated at a single site so that they are non-iid data, applying a predictive model that can achieve highly accurate predictions when trained to validation data will result in a different distribution, and the same high accuracy will not be achieved, making it impossible to properly verify the accuracy of the predictive model. For this reason, methods such as setting the training and validation data by randomly sampling from a common data set have been used.

[0034] Even if data is iid within a single location, it is highly likely that the data at each location is non-iid. If the data at each location is non-iid, 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 in 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 possible method for determining whether data is non-iid data is to compare the data distributions feature by feature. Furthermore, if a prediction model M1 generated using training data D1 from site S1 is applied to training data D2 from site S2 and the prediction accuracy is significantly lower than when the prediction model M1 was generated, it is possible to confirm that the distributions of the training data D1 from site S1 and the training data D2 from site 2 are different. Furthermore, when a prediction model is generated to determine whether the training data D1 and D2 are from sites S1 and S2, it is possible to estimate that the higher the prediction accuracy, the more different the data distributions of the training data D1 and D2 are.

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

[0037] <Example of hardware configuration of computer (server 100, base S)> FIG. 2 is a block diagram showing an example of the hardware configuration of a computer. The 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, the storage device 202, the input device 203, the output device 204, and the communication IF 205 are connected via a bus 206. The processor 201 controls the computer 200. The storage device 202 serves as a working area for the processor 201. The storage device 202 is a non-transitory or temporary recording medium that stores various programs and data. Examples of the storage device 202 include a read-only memory (ROM), a random access memory (RAM), a hard disk drive (HDD), and a flash memory. The input device 203 inputs data. Examples of the input device 203 include a keyboard, a mouse, a touch panel, a numeric keypad, and a scanner. The output device 204 outputs data. The output device 204 may be, for example, a display or a printer. The communication IF 205 connects to a network and transmits and receives data.

[0038] <Example of functional configuration of computer 200> 3 is a block diagram illustrating an example of a functional configuration of the server 100 according to the first embodiment. The server 100 includes a receiving unit 301, a distribution analysis unit 302, a generating unit 303, and an output unit 304 (a transmitting unit 341 and a display unit 342). Specifically, the receiving unit 301, the distribution analysis unit 302, the generating unit 303, and the output unit 304 realize their functions by, for example, causing the processor 201 to execute a program stored in a storage device illustrated in FIG. 2 or by the communication IF 205.

[0039] The receiving unit 301 receives the transformed features TF1 to TF10 of the locations S1 to S10 (when no distinction is made between them, they are simply referred to as the transformed features TF), the model parameters θ1 to θ10 learned by the locations S1 to S10, and the accuracy verification results of the integrated prediction model M100 via the communication IF 205. The transformed features TF are data obtained by transforming multiple features, which are training data included in the learning data D, according to preset rules.

[0040] The distribution analysis unit 302 analyzes the similarity of the training data D using the transformation features TF1 to TF10 received by the receiving unit 301. In the example of Fig. 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 45 combinations of two locations S.

[0041] Possible methods for calculating the similarity of the training data D include, for example, Euclidean distance or cosine distance between transformed features TF, an index that indicates differences in distribution such as Jensen-Shannon divergence, methods that perform dimensionality reduction such as t-SNE or PCA to convert 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 model M transmitted from one or more locations S. Specifically, for example, for a cluster including only one location S, the generation unit 303 generates the prediction model M based on the model parameters θ (local learning method). Alternatively, the generation unit 303 operates the prediction model M only at the location S without sending it to the server 100. Furthermore, for a cluster including multiple locations S, the generation unit 303 executes an integration process to generate one or more integrated prediction models based on the model parameters θ of the multiple locations S.

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

[0044] The transmitter 341 transmits to the locations S1 to S10 a notification of execution of a learning method selected from the local learning method that generates one prediction model M using learning data D from a single location S, the federated learning method that generates one integrated prediction model M100 using learning data D from multiple locations S, and the individualized federated learning method that generates one or more integrated prediction models using learning data D from multiple locations S, as well as the base prediction model M0.

[0045] There are various learning methods for personalized federated learning. For example, there is a learning method that generates a single integrated predictive model based on multiple training data D, with the focus on improving the convergence of the integrated predictive model by smoothing out the statistical heterogeneity between multiple training data D from multiple locations S. There is also a learning method that generates multiple integrated predictive models based on multiple predictive models, with the focus on improving the performance of the integrated predictive model at either the individual location S or clustered locations S level.

[0046] Since individualized associative learning is a learning method that falls under 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] The transmission unit 341 also transmits the integrated prediction model M100 generated by the generation unit 303 to the location S that generated the prediction model M used to generate the integrated prediction model. The transmission unit 341 also transmits to each location S a continuation instruction indicating whether to continue or terminate the repetition of the federated learning, based on the accuracy verification result executed at each location S.

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

[0049] 4 is a block diagram illustrating an example of a functional configuration of the base station S according to the first embodiment. The 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 by, for example, causing the processor 201 to execute a program stored in the storage device illustrated in FIG. 2 or by the communication IF 205.

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

[0051] After generating the prediction model M, the receiving unit 401 receives the integrated prediction model generated by the generating unit 303 during the associative learning. The receiving unit 401 also receives a continuation instruction from the server 100 indicating whether to continue or terminate the repetition of the associative 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 configured with parameters determined using 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 generates the model parameters θ and the prediction model M by calculating a loss function based on the prediction data and the correct answer data output as a result of inputting the training data into the base prediction model M0 so as to minimize the sum-of-squares error or cross-entropy error.

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

[0054]

number

[0055] The distribution transformation unit 403 transforms the training data D at the site S, performing a transformation process to alter the data to a state where there are no problems even if the data is moved outside the site S, and generates a transformed feature TF. For example, if the training data D consists of 100 samples each having five features, the distribution transformation unit 403 takes the average along the sample direction and transforms the data into a feature vector consisting of the average values ​​for each of the five features, which is used as the transformed feature TF. In addition to the average value, statistics such as maximum value, minimum value, and variance may also be used.

[0056] Furthermore, when the training data is image data, the distribution conversion unit 403 may convert the image data into a feature vector by quantifying multiple feature quantities using a predefined conversion rule. For example, if the image data is image data of an animal, the predefined multiple feature quantities may be the number of ears, limbs, and eyes of the animal. Such conversion rules are held as common rules in the server 100 and the base S.

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

[0058] Note that data transmitted and received by the receiving units 301, 401 and the transmitting units 341, 404 is converted and protected by encryption or the like from a security standpoint. This makes it difficult to analyze the data used for learning from the prediction model M.

[0059] <Similarity analysis of transformed feature TF by distribution analysis unit 302> Fig. 5 is an explanatory diagram showing Example 1 of similarity analysis of transformed features TF by the distribution analysis unit 302. Fig. 5 shows an example in which the server 100 analyzes the similarity between locations S1 to S10 based on the Euclidean distance between the transformed features TF received from the location S. In the Euclidean distance heat map 500, the darker the color, the closer the distance, that is, the higher the similarity in the distribution of the transformed features TF.

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

[0061] 5, the distance is expressed by a heat map 500, that is, the distance is expressed by a shade of color, but the distance may be expressed by a numerical value. Such map information is an example of information regarding the similarity of the feature amounts between the bases S.

[0062] Fig. 6 is an explanatory diagram showing a second similarity analysis example of the transformed feature TF by the distribution analysis unit 302. Fig. 6 shows an example in which the server 100 analyzes the similarity between the locations S1 to S10 by hierarchical clustering of the transformed feature TF received from the location S.

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

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

[0065] Furthermore, the dendrogram 600 can grasp the similarity in distribution of the training data D for the locations S6, S7, and S8, which is difficult to understand from the heat map 500, as cluster C3, and the similarity in distribution of the training data D for the locations S1, S3, S4, S5, S9, and S10 as cluster C12. The cluster containing only the location S2 is designated as cluster C4.

[0066] Here, if the threshold for the distribution similarity of the training data D is set to 1.0, the ten 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 includes only one location S2, is considered to have better accuracy for the prediction model M2 created by a single location compared to the integrated prediction model M100 generated by federated learning, due to the similarity in the distribution of the training data D. The application of federated learning is considered appropriate for clusters C12 and C3, which include two or more locations S.

[0067] However, it is believed that there is an appropriate federated learning method depending on the similarity of the training data D. Sites S6 to S8 form a single cluster at around 0.8, well above the threshold of 1.0, suggesting that federated learning can generate a highly accurate predictive model M. Furthermore, although six sites (sites S1, sites S3 to S5, site S9, and site S10) form a single cluster C12, they can be divided into cluster C2, which includes site S1, site S3, site S5, and site S9, and cluster C1, which includes site S4 and site S10. Thus, when there are large differences in the distribution of training data D, it may be possible to apply individualized federated learning, even within federated learning.

[0068] Based on the results of Euclidean distance and hierarchical clustering, the distribution analysis unit 302 selects one of the following: a local learning method that generates one prediction model M at a single location S; a non-individualized associative learning method that generates one integrated prediction model M100 at multiple locations S; or an individualized associative learning method that generates multiple integrated prediction models at multiple locations S.

[0069] This selection method may be, for example, a method in which the user of server 100 refers to heat map 500 or dendrogram 600 displayed on the screen and makes a selection, or a method in which the user of server 100 sets a threshold value in advance, and distribution analysis unit 302 selects one of the local learning method, the non-individualized associative learning method, and the individualized associative learning method based on the threshold value.

[0070] 6, when the threshold value is set to 1.0, the locations S1 to S10 are classified into clusters C12, C3, and C4. In this case, since only one location, S2, belongs to cluster C4, the distribution analysis unit 302 determines that the learning method for location S2 in cluster C4 is the local learning method.

[0071] Cluster C3 includes multiple locations S6 to S8, and forms a single cluster with a threshold of about 0.8, which is slightly larger than the threshold of 1.0. Therefore, distribution analysis unit 302 determines that the learning method for locations S6 to S8 in cluster C3 is the non-individualized associative learning method.

[0072] Cluster C12 includes multiple locations S1, S3, S4, S5, S9, and S10, and there are multiple clusters (clusters C1 and C2) within cluster C12 that include multiple locations. Clusters C1 and C2 are grouped together near the threshold of 1.0, so there is a large difference in the distribution of training data D. Therefore, distribution analysis unit 302 determines that the training method for locations S1, S3, S4, S5, S9, and S10 in cluster C12 is the individualized associative learning method.

[0073] Furthermore, the distribution analysis unit 302 may set constraints other than the threshold value to select a learning method. For example, the distribution analysis unit 302 may vary the threshold value, and when a preset number of clusters or the number of bases belonging to each cluster is reached, the distribution analysis unit 302 may stop varying the threshold value and determine a learning method for the cluster at that time.

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

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

[0076] The latter clusters C3, C4, and C12 belong to a number of locations of "6." The threshold value continues to decrease. When the threshold value reaches 0.9, the clusters to which cluster C1 belongs are clusters C1, C2, C3, and C4. Because the number of clusters to which cluster C1 belongs is "4," the decrease in the threshold value ends, and the distribution analysis unit 302 determines a learning method for each of 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 to various values ​​such as "n" (n is an integer of 1 or more), "n or more," "more than n," "n or less," "fewer than n," etc. The number of bases within one cluster can also be set to various values ​​such as "m" (m is an integer of 1 or more), "m or more," "more than m," "m or less," "fewer than m," etc.

[0078] In the above example, the set number of clusters and the number of bases in one cluster are separately described, but both the set number of clusters and the number of bases in one cluster may be set as constraints, for example, the set number of clusters is "n" and the number of bases in one cluster is "m." In the above example, the distribution analysis unit 302 decreases the threshold value from the maximum value, but the threshold value may be increased by a predetermined amount from the minimum value (for example, 0.0).

[0079] <Study example> 7 is an explanatory diagram showing learning example 1 between server 100 and base S. In FIG. 7, when cluster C12 is the cluster for which the learning method is to be determined, individualized federated learning is performed to generate one integrated prediction model M700 using learning data D1, D3 to D5, D9, and D10 from cluster C1 consisting of bases S1, S3, S5, and S9 and cluster C2 consisting of bases S4 and S10.

[0080] Each of locations S1, S3 to S5, S9, and S10 receives, via receiver 401, a notification from server 100 indicating that personalized federated learning has been selected. Each of locations S1, S3 to S5, S9, and S10 generates prediction models M1, M3 to M5, M9, and M10 using the respective learning data D1, D3 to D5, D9, and D10 via learner 402. Thereafter, each of locations S1, S3 to S5, S9, and S10 transmits, via transmitter 404 to server 100, model parameters θ1, θ3 to θ5, θ9, and θ10 of the generated prediction models M1, M3 to M5, M9, and M10.

[0081] The server 100 receives the model parameters θ1, θ3 to θ5, θ9, and θ10 via the receiving unit 301. The server 100 executes an integration process using the model parameters θ1, θ3 to θ5, θ9, and θ10 via the generating unit 303 to generate an integrated prediction model M700. The server 100 repeats the update process for the generated integrated prediction model M700 until the generated integrated prediction model M700 achieves the desired prediction accuracy.

[0082] Note that FIG. 7 shows an example in which individualized associative learning is performed to generate one integrated prediction model M700 using the learning data D1, D3 to D5, D9, and D10 of clusters C1 and C2 when cluster C12 is the cluster for which the learning method is to be determined. However, individualized associative learning may be performed to generate one integrated prediction model using the learning data D1, D3, D5, and D9 of cluster C1, and another integrated prediction model using the learning data D4 and D10 of cluster C2.

[0083] Fig. 8 is an explanatory diagram showing learning example 2 by server 100 and site S. Fig. 8 shows an example in which, when cluster C3 is the target cluster for determining the learning method, non-individualized associative learning is performed to generate one integrated prediction model using learning data D3 of cluster C3 consisting of sites S6 to S8.

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

[0085] The server 100 receives the model parameters θ6 to θ8 of the prediction models M6 to M8 via the receiving unit 301. The server 100 generates an integrated prediction model M800 by executing an integration process using the model parameters θ6 to θ8 via the generating unit 303. The server 100 repeats the update process for the generated integrated prediction model M800 until the generated integrated prediction model M800 achieves the desired prediction accuracy.

[0086] Note that in Figures 7 and 8, an example has been described in which the server 100 determines a learning method by referring to the dendrogram 600 and generates a prediction model M or an integrated prediction model, but the server 100 may also determine a learning method by clustering multiple locations S by referring to the heat map 500 and generate a prediction model M or an integrated prediction model.

[0087] <Integration processing of prediction model M> Next, a specific description will be given of the integration process of the prediction model M by the generation unit 303. In the integration process of the prediction model M, model parameters θ t The integrated prediction model M100 to which the integrated prediction model M100 has been applied (t is the number of updates of the integrated prediction model M100) is used as an initial value, and each of the K locations S performs learning using its own learning data D.

[0088] The generation unit 303 then calculates the model parameters θ of the K prediction models M generated by each of the K locations S. k The gradient with respect to g k and obtain the gradient g k Using the sum of the averages of the above, the model parameters θ of the integrated prediction model M100, which is updated for the t+1th time, are calculated as shown in the following formula (2). t+1 N is the total number of samples of all the training data D used for training at K locations S, and N k is the number of samples of the training data D at the location k.

[0089]

number

[0090] Here, in the above equation (2), the model parameter θ k The gradient with respect to g k was used, which is the model parameter θ k This is a security-conscious method to prevent the training data D from being analyzed, and the model parameters θ k Alternatively, the generation unit 303 may use a method other than the above formula (2) to integrate the prediction models M1 to M10, depending on the structure of the prediction model, such as a fully connected layer or a convolutional layer.

[0091] For example, as shown in the following equation (3), each model parameter θ k It can be the average value normalized by the number of samples of the learning data D used for learning, or the model parameters θ k and its gradient g k may also be used.

[0092]

number

[0093] <Integrated learning pre-processing by server 100> 9 is a flowchart showing an example of a procedure for pre-processing for integrated learning by the server 100. It is assumed that the server 100 is capable of communicating with locations S that wish to participate in federated learning. The server 100 receives the transformation feature TF from each location S via the receiving unit 301 (step S901). Next, the server 100 uses the received transformation feature TF via the distribution analysis unit 302 to analyze the similarity of the learning data D (step S902).

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

[0095] Specifically, for example, the server 100 displays, on the display unit 342, the heat map 500, the dendrogram 600, and the training method for the association prediction model M determined in step S903, that is, the method for generating the integrated prediction model M100 and the training method for the prediction model M, as distribution analysis results. The server 100 also notifies each site S of the training method determined in step S903 via the transmission unit 341. This completes the integrated learning pre-processing.

[0096] <Pre-learning processing by base S> 10 is a flowchart showing an example of a pre-learning process procedure performed by the site S. The site S has already acquired, from the server 100, a method and a tool for converting feature quantities, which are teacher data in the learning data D.

[0097] The site S uses a feature conversion method and tool to convert the features that are training data in the training data D, and generates converted features TF (step S1001). Next, the site S transmits the converted features TF to the server 100 (step S1002). Thereafter, the site S transmits the notification of step S904 from the server 100 (step S1003). This completes the pre-learning process.

[0098] In this way, according to the above-described data analysis processing system, it is possible to determine a learning method for the prediction model M appropriate for each site S for the wide variety of learning data D1 to D10 at each site S without moving the learning data D1 to D10 stored at multiple sites S1 to S10 outside the site S, and to generate a prediction model M or integrated prediction models M100, M700, M800 appropriate for each site S.

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

[0100] 11 is an explanatory diagram showing a display example 1 of a display screen. A 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 location list 1111, location analysis results 1112, a location classification start button 1113, and a location classification result confirmation button 1114.

[0101] The server 100 accepts the selection of "Analysis" in the mode field 1103 by user operation, and accepts the pressing of the View clients button 1101. As a result, the server 100 displays a list of locations wishing to participate in federated learning in the location list 1111. When the server 100 accepts the pressing of the location classification start button 1113 by user operation, the server 100 transmits to each location S the methods and tools required for generating transformation features.

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

[0103] When the user checks the location list 1111 and the location analysis result 1112 and the server 100 accepts the pressing of the location classification result confirmation button 1114, it transmits a notification of the learning method assigned to each location S. Here, the learning method for each location S in the location list 1111 may be individually editable directly by user operation.

[0104] 12 is an explanatory diagram showing display 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 location list 1111, a group selection field 1211, a federated learning result 1212, and a federated learning result confirmation button 1213.

[0105] The server 100 accepts the selection of "Feseration" in the mode field 1103 through a user operation and accepts the pressing of the View results button 1102, thereby displaying the federated learning results 1212. The server 100 also accepts the selection of a learning group in the group selection field 1211 through a user operation, thereby displaying the federated learning results 1212 of the site S belonging to the learning group displayed in the site list 1111. If the federated learning has been repeated a predetermined number of times or has achieved a desired prediction accuracy, the server 100 ends the federated learning process by pressing the federated learning result confirmation button 1213 through a user operation. If the user wishes to redo the site selection or redo the learning based on the federated learning results, the user can change the mode field 1103 to "Analysis," return to the display screen 1100, and perform the generation of the transformation feature TF, the determination of the learning method, and the federated learning again.

[0106] As described above, according to the first embodiment, it is possible to generate an integrated prediction model M100 appropriate for each of the locations S for the varied training data D at each of the locations S, without moving the training data D stored in the locations S outside the location S. Furthermore, although the number of locations is 10 in the first embodiment, it may be 9 or less or 11 or more, as long as it is 2 or more. [Example]

[0107] Example 2 will be described. In Example 1, an appropriate learning method for each location S was determined by performing analysis using the transformation feature of each location S. Example 2 is an example in which the distribution analysis unit 302 determines an appropriate learning method for each location S using federated 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 a learning method. In Example 2, differences from Example 1 will be mainly described, and therefore commonalities with Example 1 will not be described.

[0108] For example, if the learning data D of each site S is a stage classification of cancer patients at each hospital, the output of the finally generated integrated prediction model will also be a stage classification. In Example 2, when a learning method appropriate for each site S is determined by federated learning, the output of the finally generated integrated prediction model will be the learning data D of which site S (i.e., the site S to which the learning data D belongs).

[0109] 13 is an explanatory diagram showing federated learning method 1 for realizing appropriate learning for each site S at multiple sites S where training data D cannot be sent outside the site S. When the multiple sites S consist only of sites S1 and S2, site S1 generates a classifier that is a prediction model M1 using a base prediction model M0 and training data D1 and transmits model parameters θ1 of this classifier to the server 100. Similarly, site S2 generates a classifier that is a prediction model M2 using the base prediction model M0 and training data D2 and transmits model parameters θ2 of this classifier to the server 100. The server 100 generates a two-class classification classifier (hereinafter referred to as the integrated classifier) ​​that is an integrated prediction model M100 by federated learning based on the model parameters θ1 and θ2 of the prediction models M1 and M2, which were generated by assigning 0 to the class of the training data D for site S1 and 1 to the class of the training data D for site S2.

[0110] The bases S1 and S2 receive the generated integrated classifier (integrated prediction model M100, which is a two-class classification classifier) ​​from the server 100 and apply it to calculate predicted probabilities. The bases S1 and S2 transmit the calculated predicted probabilities or values ​​obtained by converting the predicted probabilities into propensity scores to the server 100 as converted features. In FIG. 13, the propensity scores are used.

[0111] 13, the server 100 compares a box plot 1301 of the propensity scores for site S1 with a box plot 1302 of the propensity scores for site S2. For example, if site S1 has a high probability of being site S1 and site S2 has a low probability of being site S1, the learning data D1 and D2 for the two sites S1 and S2 are easily distinguishable, meaning that the similarity in their data distributions is low. In this case, the overlapping range 1312 between the box plots 1301 and 1302 becomes narrower, or the number of included samples becomes smaller.

[0112] Conversely, if the probability of being site S1 at site S1 and the probability of being site S1 at site S2 are similar, the learning data D1 and D2 for the two sites S1 and S2 are difficult to distinguish, meaning that the data distributions are highly similar. In this case, the overlapping range 1312 between the boxplots 1301 and 1302 will be wider, or the number of included samples will be greater. By comparing the distributions of predicted probabilities when the integrated prediction model for site classification is applied to sites S1 and S2, the server 100 can evaluate the similarity between the learning data D1 and D2 for both sites S1 and S2.

[0113] Here, the propensity score is a value obtained by a statistical method or balancing method used to estimate causal effects by adjusting covariates in observational studies where random allocation is difficult and various confounds are likely to occur. In Example 2, in the analysis of the characteristics of the learning data D from each site S, the predicted probability may be used as is, or a weighting method using the propensity score, such as an inverse probability weighting (IPW) estimator or a doubly robust estimator, may be used.

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

[0115] 13 shows an example in which the number of bases is two, but the fewer the number of bases, the easier it is for the server 100 to estimate the learning data D of other bases S from the model parameters of the integrated prediction model. For this reason, measures can be taken such as not informing the bases S of the number of participating bases, limiting use to cases in which the number of participating bases is a certain number or more, or using privacy protection methods such as Differential Privacy Stochastic Gradient Descent SGD (DPSGD) or Private Aggregation of Teacher Ensembles (PATE).

[0116] FIG. 14 is an explanatory diagram illustrating federated learning method 2, which realizes appropriate learning for each site S at multiple sites S where training data D cannot be sent outside the site S. FIG. 14 particularly illustrates a graph 1400 for a case where the number of sites S is three or more. When the multiple sites S are sites S1, S2, and S3, site S1 generates a classifier that is a prediction model M1 using the base prediction model M0 and training data D1 and transmits model parameters θ1 of this classifier to the server 100. Similarly, site S2 generates a classifier that is a prediction model M2 using the base prediction model M0 and training data D2 and transmits model parameters θ2 of this classifier to the server 100. Similarly, site S3 generates a classifier that is a prediction model M3 using the base prediction model M0 and training data D3 and transmits model parameters θ3 of this classifier to the server 100.

[0117] The server 100 generates a three-class classification classifier (integrated classifier) ​​that is an integrated prediction model M100 by federated learning based on model parameters θ1 to θ3 of prediction models M1 to M3 that were generated by setting the class of learning data D for location S1 as 0, the class of learning data D for location S2 as 1, and the class of learning data D for location S3 as 2. Generally, a propensity score is a method used when there are two groups, and when there are three or more groups, a generalized propensity score is used.

[0118] As shown in the following formula (4), the generalized propensity score can be expressed using the propensity score PS(k|x) at site k, which is the kth site S among the K sites S, for K different pieces of learning data D at K sites S. Formula (4) below is designed so that the sum of the K generalized propensity scores is 1.

[0119]

number

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

[0121] For example, if the overlap range 1412 is equal to or greater than a threshold, the server 100 determines that the data distributions of the learning data D1 and D2 are similar. If they are determined to be similar, the server 100 assigns the locations S1 and S2 to the same cluster. A similar determination is made for the other overlap ranges 1423 and 1413. As with the federated learning method 1, the similarity of the learning data D1 to D3 may be determined not only based on the size of the overlap ranges 1412, 1423, and 1413, but also on the number of samples included in the overlap ranges 1412, 1423, and 1413.

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

[0123] To avoid this, the server 100 first executes an integrated prediction model generation process using federated learning to generate the final output of the integrated prediction model (e.g., cancer stage). However, the number of updates may be insufficient. Next, the server 100 generates a classification classifier using federated learning, with the model parameter θ of the intermediate layer of the generated integrated prediction model unchanged and the output of the base prediction model M0 as the number of locations S. By performing such federated learning based on transfer learning, it becomes possible to use the output of the intermediate layer that can extract the features of the training data D from each location S. This makes it possible to avoid falling into a local solution when generating a prediction model M at each location S by including only the training data D from one location S.

[0124] As described above, according to the second embodiment, it is possible to evaluate the similarity of the learning data D of the base S by comparing the distributions of the calculated generalized tendency scores. [Example]

[0125] A description will be given of Example 3. Example 3 is an example in which the device configurations of the server 100 and the base S are common to Examples 1 and 2. In Example 3, differences from Examples 1 and 2 will be mainly described, and therefore commonalities with Examples 1 and 2 will not be described.

[0126] 15 is a block diagram showing an example of the functional configuration of a computer 1500 that serves as at least one of the server 100 and the site S. The computer 1500 that serves as at least one of the server 100 and the site S includes a receiving unit 1501, a distribution analysis unit 302, a generating unit 303, a learning unit 402, a distribution conversion unit 403, and an output unit 1504. The receiving unit 1501 functions as the 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 the transmitting units 341 and 404.

[0127] Even if computer 1500 is server 100, it may participate in federated learning as one location using learning data D in computer 1500. Also, for example, in federated learning performed at locations S6 to S8, location S6 may play the role of server 100 when generating an integrated learning model.

[0128] As described above, even if the training data D is non-iid data, it is possible to generate an appropriate prediction model M at each location S by using the similarity relationship between the transformation features TF without determining whether the training data D that cannot be sent outside the location S is non-iid data, and the server 100 can generate an appropriate integrated prediction model through federated learning.

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

[0130] Furthermore, although it is not necessary 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 prediction model, just as in the case of non-iid data.

[0131] The analysis device, which is the server 100 according to the first and second embodiments, can also be configured as follows (1) to (13).

[0132] (1) An analysis device capable of communicating with learning devices that are multiple locations S includes a receiving unit 301 that receives transformed features TF obtained by transforming features (teacher data) included in learning data D held by each of the multiple learning devices according to a predetermined rule, a distribution analysis unit 302 that analyzes the distribution of multiple features of the multiple learning devices based on the transformed features 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 device of (1) above, the output unit 304 outputs information on the similarity of the feature amounts between the learning devices as the distribution analysis result.

[0134] (3) In the analysis device of (1) above, the information regarding the similarity of the features between the learning devices is map information indicating the similarity of the features between each combination of two of the plurality of learning devices.

[0135] (4) In the analysis device of (1) above, the information on the similarity of the feature amounts between the learning devices is a dendrogram 600 that indicates the similarity of the feature amounts between the plurality of learning devices.

[0136] (5) In the analysis device of (1) above, the distribution analysis unit 302 determines, based on the distribution analysis results, the learning method to be applied to each of the multiple learning devices to be one of the following: a first learning method (local learning method) that generates a predictive model M using the learning data held by the learning device; an associative learning method (non-individualized associative learning method) that generates one integrated predictive model M100 through associative learning with other learning devices (cluster C3); and a third learning method (individualized associative learning method) that generates one or more predictive models integrated through associative learning with other learning devices that have similar features to the learning device; and the output unit 304 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.

[0137] (6) In the analysis device of (5) above, the distribution analysis unit 302 determines the 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 distribution analysis results.

[0138] (7) In the analysis device of (6) above, the distribution analysis unit 302 determines the 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 distance between the transformation features TF of the plurality of learning devices.

[0139] (8) In the analysis device of (6) above, the distribution analysis unit 302 determines the 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 transformation feature TF and a threshold value.

[0140] (9) In the analysis device of (8) above, the distribution analysis unit 302 determines the 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 transformation feature TF, a threshold value, and constraint conditions.

[0141] (10) In the analysis device of (9) above, the constraint is the number of learning devices whose features are similar to each other.

[0142] (11) In the analysis device of (9) above, the constraint is the number of sets of the learning devices in which the features are similar to each other.

[0143] (12) In the analysis device of (1) above, the receiving unit 301 receives model parameters of classifiers for identifying the learning devices, which are generated by learning the learning data held by each of the plurality of learning devices; 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 generating unit to each of the plurality of learning devices; the receiving unit receives a classification result by the integrated classifier from each of the plurality of learning devices; and the distribution analysis unit 302 analyzes the plurality of learning devices based on the classification result from each of the learning devices received by the receiving unit 301. The distribution analysis unit 302 determines the learning method to be applied to each of the plurality of learning devices as one of a first learning method (local learning method) that generates a predictive model M using the learning data held by the learning device, a federated learning method (non-individualized federated learning method) that generates one integrated predictive model M100 by federated learning between the learning device and other learning devices (cluster C3) that have similar features to the learning device, and a third learning method (individualized federated learning method) that generates one or more integrated predictive models by federated learning of multiple groups consisting of the learning device and other learning devices that have similar features to the learning device, and the output unit 304 transmits the learning method determined for each of the plurality of learning devices by the distribution analysis unit 302 to each of the plurality of learning devices.

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

[0145] The present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the spirit and scope of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to configurations including all of the described configurations. Furthermore, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Furthermore, the configuration of another embodiment may be added to the configuration of one embodiment. Furthermore, part of the configuration of each embodiment may be added to, deleted from, or replaced with other configurations.

[0146] Furthermore, the aforementioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by having a processor interpret and execute a program that realizes each function.

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

[0148] In addition, the control lines and information lines shown are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines that are necessary for implementation. In reality, it can be considered that almost all components are interconnected. [Explanation of symbols]

[0149] 100 servers 200 computers 301 Receiving unit 302 Distribution analysis section 303 Generation part 304 Output section 313 Distribution Conversion Unit 341 Transmitter 342 Display section 401 Receiving unit 402 Learning Department 403 Distribution Conversion Unit 404 Sending Unit 500 Heatmaps 600 dendrograms 1500 calculator 1501 Receiving unit 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 plurality of feature quantities of the plurality of learning devices based on the transformed feature quantities for each of the learning devices received by the receiving unit; an output unit that outputs, as a distribution analysis result analyzed by the distribution analysis unit, map information indicating the similarity of feature quantities between each combination of two of the plurality of learning devices; An analytical device comprising:

2. The analytical device according to claim 1, the information regarding the similarity of the feature values ​​between the learning devices is a dendrogram indicating the similarity of the feature values ​​between the plurality of learning devices; An analytical device characterized by:

3. The analytical device according to claim 1 , the distribution analysis unit determines, based on the distribution analysis result, a learning method to be applied to each of the plurality of learning devices to one of a first learning method for generating a prediction model using the learning data held by the learning device, a second learning method for generating one prediction model integrated by federated learning with other learning devices, and a third learning method for generating one or more prediction models integrated by federated learning with other learning devices similar to the feature quantities 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. An analytical device characterized by:

4. The analytical device according to claim 3, the distribution analysis unit determines the 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 distribution analysis result; An analytical device characterized by:

5. The analytical device according to claim 4, 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. An analytical device characterized by:

6. The analytical device according to claim 4, 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 transformation feature and a threshold value; An analytical device characterized by:

7. The analytical 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 transformation feature, a threshold, and a constraint condition; An analytical device characterized by:

8. The analytical device according to claim 7, the constraint is the number of learning devices having similar features; An analytical device characterized by:

9. The analytical device according to claim 7, the constraint is the number of sets of the learning devices having similar features; An analytical device characterized by:

10. The analytical device according to claim 3, 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:

11. 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, and model parameters of a classifier that identifies the learning device, the model parameters being generated by learning the learning data held by each of the plurality of learning devices; a distribution analysis unit that analyzes a distribution of the plurality of feature quantities of the plurality of learning devices based on the transformed feature quantities for each of the learning devices received by the receiving unit; an output unit that outputs the distribution analysis result analyzed by the distribution analysis unit; 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 integrated 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 plurality of 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 that have similar features to the learning device, and a third learning method that generates one or more predictive models integrated through federated learning with other learning devices that have similar features to 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. An analytical device characterized by:

12. An analysis method using an analysis device capable of communicating with a plurality of learning devices, comprising: The analysis device 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 quantities of the plurality of learning devices based on the transformed feature quantities for each of the learning devices received by the receiving process; an output process for outputting, as a distribution analysis result obtained by the distribution analysis process, map information indicating the similarity of feature quantities between each combination of two of the plurality of learning devices; An analytical method characterized by carrying out the steps of:

13. 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 quantities of the plurality of learning devices based on the transformed feature quantities for each of the learning devices received by the receiving process; an output process for outputting the distribution analysis result analyzed by the distribution analysis process; An analysis program characterized by executing the above.

Images