JP7527627B2 - Distributed data processing device, terminal, distributed data processing program, terminal control program, distributed data processing method, and terminal control method - Google Patents

Description

The present invention relates to a distributed data processing device, a terminal, a distributed data processing program, a terminal control program, a distributed data processing method, and a terminal control method.

In recent years, the networking of society has progressed rapidly, and an environment for collecting data in various sectors of society has been established. For example, various institutions such as medical institutions and financial institutions independently collect and hold large amounts of data. Against this background, it is hoped that productivity can be improved by using technology that integrates this data and analyzes it as big data while protecting personal information, trade secrets, etc. One example of such technology is the service user information sharing system disclosed in Patent Document 1.

This service user information sharing system includes a provided information acquisition terminal and a sharing device. The provided information acquisition terminal is a user terminal capable of transmitting a history of identification information, and a terminal capable of adding information related to services used by the user to the identification information received from the user's terminal and the provideable information set by the user, and transmitting the information. The sharing device includes a reception history storage means and an association means. The reception history storage means stores information transmitted from the provided information acquisition terminal. The association means associates, based on the history of identification information transmitted from the user's terminal, the reception information of a common user extracted from the reception information stored in the reception history storage means as information of a common user.

JP 2016-038776 A

However, the above-mentioned service user information sharing system deletes data that should be protected and integrates data that has a format that complies with a standard. For this reason, the service user information sharing system needs to establish a data standard, taking into consideration the purpose of integrating and analyzing data received from multiple terminals, and integrate the data after confirming that the data has been converted into a format that complies with that standard.

Data collaboration analysis is one example of a technique for integrating and analyzing multiple data sets without using standards. However, data collaboration analysis outputs an integrated analysis model with a multi-stage structure that uses abstracted data. Such integrated analysis models have relatively low interpretability and explainability, and it is often difficult for users to understand the reasons for each result of the integrated analysis.

The present invention has been made in consideration of the above circumstances, and aims to provide a distributed data processing device, a terminal, a distributed data processing program, a terminal control program, a distributed data processing method, and a terminal control method that can generate an integrated analysis model that has relatively high interpretability and/or explainability.

One aspect of the present invention is a distributed data processing device that includes: a receiving unit that performs, for each of the terminals, a process of receiving abstracted anchor data generated by performing an abstraction process on anchor data held by the terminal; and a process of receiving abstracted partial data generated by performing an abstraction process on partial data including a part of data included in overall data and a part of features included in the overall data, which are held separately on the multiple terminals; a generating unit that generates integrated data by integrating the multiple abstracted partial data using the abstracted anchor data received for each of the terminals, generates a first integrated analysis model that outputs first label data corresponding to the content indicated by the elements included in the integrated data based on the integrated data, generates integrated anchor data by integrating the multiple anchor data using the abstracted anchor data received for each of the terminals, inputs the integrated anchor data into the first integrated analysis model, and generates second label data corresponding to the content indicated by the elements included in the integrated anchor data; and a transmitting unit that transmits the second label data to the terminals.

In addition, in the distributed data processing device described above, the receiving unit executes, for each terminal, the process of receiving the abstraction anchor data, a process of receiving first abstraction anchor data generated by applying an abstraction process to first anchor data held by the terminal, and a process of receiving second abstraction anchor data generated by applying an abstraction process to second anchor data held by the terminal and which is the same as the first anchor data, and the generating unit generates the integrated data by integrating multiple pieces of abstraction partial data using the first abstraction anchor data received for each terminal, and generates the integrated anchor data by integrating multiple pieces of anchor data using the second abstraction anchor data received for each terminal.

In addition, in the distributed data processing device described above, the receiving unit executes, for each terminal, the process of receiving the abstraction anchor data, a process of receiving first abstraction anchor data generated by applying an abstraction process to first anchor data held by the terminal, and a process of receiving second abstraction anchor data generated by applying an abstraction process to second anchor data held by the terminal and different from the first anchor data, and the generating unit generates the integrated data by integrating multiple pieces of abstraction partial data using the first abstraction anchor data received for each terminal, and generates the integrated anchor data by integrating multiple pieces of anchor data using the second abstraction anchor data received for each terminal.

In the distributed data processing device described above, the receiving unit executes, for each terminal, a process of receiving the abstracted anchor data represented by a square matrix or a rectangular matrix, which is generated by performing an abstraction process on the anchor data represented by a square matrix or a rectangular matrix, and a process of receiving the abstracted partial data represented by a square matrix or a rectangular matrix, which is generated by performing an abstraction process on the partial data represented by a square matrix or a rectangular matrix.

In the distributed data processing device described above, the receiving unit executes, for at least one of the terminals, a process of receiving the abstracted anchor data, which is generated by applying an abstraction process to the anchor data represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing, and which is represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing, and a process of receiving the abstracted partial data, which is generated by applying an abstraction process to the partial data represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing, and which is represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing.

One aspect of the present invention is a terminal including: a transmitting unit that transmits abstracted anchor data generated by performing an abstraction process on anchor data, and abstracted partial data generated by performing an abstraction process on partial data including a part of data included in overall data held in a plurality of terminals and a part of features included in the overall data, to a distributed data processing device; a receiving unit that inputs integrated anchor data generated by integrating a plurality of anchor data using the abstracted anchor data received by the distributed data processing device for each terminal into a first integrated analysis model that outputs first label data corresponding to the contents indicated by elements included in integrated data generated by integrating a plurality of the abstracted partial data based on the integrated data, and receives second label data corresponding to the contents indicated by elements included in the integrated anchor data from the distributed data processing device; and a generating unit that generates a second integrated analysis model using the anchor data and the second label data.

One aspect of the present invention is a distributed data processing program that causes a computer to execute, for each terminal, a process of receiving abstracted anchor data generated by performing an abstraction process on anchor data held by a terminal, and a process of receiving abstracted partial data generated by performing an abstraction process on partial data including a portion of data included in overall data and a portion of features included in the overall data, which are held separately on a plurality of the terminals; a generation function of generating integrated data by integrating a plurality of the abstracted partial data using the abstracted anchor data received for each terminal, generating a first integrated analysis model that outputs first label data corresponding to the content indicated by the elements included in the integrated data based on the integrated data, generating integrated anchor data by integrating a plurality of the anchor data using the abstracted anchor data received for each terminal, inputting the integrated anchor data into the first integrated analysis model to generate second label data corresponding to the content indicated by the elements included in the integrated anchor data, and a transmission function of transmitting the second label data to the terminal.

One aspect of the present invention is a terminal control program that executes the following functions: a transmission function that transmits, to a distributed data processing device, abstracted anchor data generated by performing an abstraction process on anchor data, and abstracted partial data generated by performing an abstraction process on partial data including a portion of data included in overall data held in a plurality of terminals and a portion of features included in the overall data; a reception function that inputs integrated anchor data generated by integrating a plurality of anchor data using the abstraction anchor data received by the distributed data processing device for each terminal into a first integrated analysis model that outputs first label data corresponding to the contents indicated by elements included in integrated data generated by integrating a plurality of the abstracted partial data based on the integrated data, and receives second label data corresponding to the contents indicated by elements included in the integrated anchor data from the distributed data processing device; and a generation function that generates a second integrated analysis model using the anchor data and the second label data.

One aspect of the present invention is a method of information processing performed by a distributed data processing device, which is a computer device, under the control of software, in which the distributed data processing device executes, for each terminal, a process of receiving abstracted anchor data generated by applying an abstraction process to anchor data held by a terminal, and a process of receiving abstracted partial data generated by applying an abstraction process to partial data including a portion of data included in overall data held by a plurality of the terminals and a portion of features included in the overall data, the distributed data processing device generates integrated data by integrating a plurality of the abstracted partial data using the abstracted anchor data received for each terminal, generates a first integrated analysis model that outputs first label data corresponding to content indicated by elements included in the integrated data based on the integrated data, generates integrated anchor data by integrating a plurality of the anchor data using the abstracted anchor data received for each terminal, inputs the integrated anchor data into the first integrated analysis model to generate second label data corresponding to content indicated by elements included in the integrated anchor data , and the distributed data processing device transmits the second label data to the terminal.

One aspect of the present invention is a method of information processing performed by a terminal that is a computer device under the control of software, in which the terminal transmits, to a distributed data processing device, abstraction anchor data generated by applying an abstraction process to anchor data and abstraction partial data generated by applying an abstraction process to partial data including a portion of data included in overall data held by a plurality of terminals and a portion of features included in the overall data, the terminal uses the abstraction anchor data received by the distributed data processing device for each terminal to output first label data corresponding to content indicated by elements included in integrated data generated by integrating a plurality of the abstraction partial data based on the integrated data, the terminal inputs integrated anchor data generated by integrating a plurality of the anchor data using the abstraction anchor data received by the distributed data processing device for each terminal, and receives second label data corresponding to content indicated by elements included in the integrated anchor data from the distributed data processing device, and the terminal generates a second integrated analysis model using the anchor data and the second label data.

The present invention provides a distributed data processing device, a terminal, a distributed data processing program, a terminal control program, a distributed data processing method, and a terminal control method that can generate an integrated analysis model that has relatively high interpretability and/or explainability.

FIG. 1 is a diagram illustrating an example of an integrated analysis model generation system according to a first embodiment of the present invention. 4 is a sequence diagram showing an example of processing executed by a terminal and a distributed data processing device according to the first embodiment of the present invention; FIG. 13 is a diagram showing an example of an integrated analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the content indicated by elements contained in integrated data generated by integrating four abstracted partial data received from four institutions based on the integrated data. FIG. 4 is a diagram showing an example of a result of classifying artificial data using the integrated analysis model shown in FIG. 3 . FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 6 is a diagram showing an example of a result of classifying artificial data using the analysis model shown in FIG. 5 . FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 8 is a diagram showing an example of a result of classifying artificial data using the analysis model shown in FIG. 7 . FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 10 is a diagram showing an example of a result of classifying artificial data using the analysis model shown in FIG. 9 . FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 12 is a diagram showing an example of a result of classifying artificial data using the analysis model shown in FIG. 11 . A figure showing an example of a second integrated analysis model generated by a terminal using second label data generated by a distributed data processing device related to the first embodiment of the present invention based on four abstracted partial data received from each of the four institutions. FIG. 14 is a diagram showing an example of a result of classifying artificial data using the second integrated analysis model shown in FIG. 13 . 3A to 3C are diagrams illustrating an example of a company identification code, financial data, an industry label, and a rating according to the first embodiment of the present invention. FIG. 13 is a diagram showing an example of an integrated analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the content indicated by elements contained in integrated data generated by integrating four abstracted partial data received from each of four institutions based on the integrated data. FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. FIG. 13 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from one institution, based on the partial data. A figure showing an example of a second integrated analysis model generated by a terminal using second label data generated by a distributed data processing device related to the first embodiment of the present invention based on four abstracted partial data received from each of the four institutions. FIG. 13 is a diagram showing an example of a submatrix indicating partial data according to the third embodiment of the present invention. FIG. 13 is a diagram showing an example of a submatrix indicating partial data according to another embodiment of the present invention.

[First embodiment]
An example of an integrated analysis model generation system according to the first embodiment will be described with reference to FIG. 1. FIG. 1 is a diagram showing an example of an integrated analysis model generation system according to the first embodiment of the present invention. As shown in FIG. 1, the integrated analysis model generation system 1 includes a terminal 111, a terminal 112, a terminal 121, ..., a terminal 1cd (c: a natural number equal to or greater than 1 and less than n, d: a natural number equal to or greater than 1 and less than m, where c x d is 2 or greater), and a distributed data processing device 20. The terminals 111, ..., the terminal 1cd, and the distributed data processing device 20 are all connected to a network NW and can communicate with each other. The network NW is, for example, the Internet, a Wide Area Network (WAN), an intranet, or a Local Area Network (LAN).

Terminals 111, ..., terminal 1cd are computers, for example servers, managed by different organizations, institutions, etc. The institutions referred to here are, for example, medical institutions, corporate rating agencies. As shown in FIG. 1, terminal 111 comprises a generation unit 1111, a transmission unit 1211, and a reception unit 1311. Similarly, terminals 112, ..., terminal 1cd comprise a generation unit 1112, a transmission unit 1212, and a reception unit 1312, ..., generation unit 11cd, transmission unit 12cd, and reception unit 13cd, respectively. Also, as shown in FIG. 1, distributed data processing device 20 comprises a reception unit 210, a generation unit 220, and a transmission unit 230.

Terminal 111, terminal 112, terminal 121, ..., terminal 1cd each have partial data represented by submatrix _X1,1 , partial data represented by submatrix _X1,2 , partial data represented by submatrix _X2,1 , ..., partial data represented by submatrix Xc _,d . These submatrices are square or rectangular matrices, and are submatrices of matrix X with n rows and m columns (n: the number of data included in the entire data, m: the number of features included in the entire data) that represents the entire data. Therefore, the relationship between matrix X and submatrix _X1,1 , submatrix _X1,2 , submatrix _X2,1 , ..., submatrix Xc _,d is expressed, for example, by the following formula (1). Also, submatrix Xi _,j (i: a natural number from 1 to c, j: a natural number from 1 to d) included in the right side of formula (1) is a matrix with n _i rows and m _j columns.

The partial data described above each contain a portion of the features contained in the overall data, and are different data from each other.

Furthermore, matrix Y having row vectors corresponding to the contents indicated by each row vector of matrix X is expressed, for example, by the following formula (2). Matrix Y is a matrix with n rows and l columns (l: a natural number equal to or greater than 1). The submatrix Y _i with r _i rows and m _j columns included on the right side of formula (2) has the same number of rows as the submatrix X _i,j included on the right side of formula (1). In the following description, data indicating the contents corresponding to each row of matrix Y may be referred to as first label data.

Terminal 111, terminal 112, terminal 121, ..., terminal 1cd respectively possess anchor data represented by submatrix X ^anc _1,1 , anchor data represented by submatrix X ^anc _1,2 , anchor data represented by submatrix X ^anc _2,1 , ..., anchor data represented by submatrix X ^anc _c,d . These anchor data serve as first anchor data used to generate integrated data described later, and second anchor data used to generate integrated anchor data described later. In other words, the first anchor data and the second anchor data are the same.

Moreover, these submatrices are square or rectangular matrices, and are submatrices of the matrix ^Xanc with r rows and m _columns (r: a natural number less than n) representing all anchor data. Therefore, the relationship between the matrix ^Xanc and the submatrix ^Xanc1,1 _, the submatrix ^Xanc1,2 , _{the submatrix Xanc2,1} ^, ..., the submatrix ^Xancc _,d is expressed, for example, by the following formula (3). Furthermore, the ^submatrix Xanc _i,j included in formula (3) is a matrix with r _i rows and m _j columns. Note that the submatrix ^Xanc _:,j included in formula (3) is a submatrix with r rows and m _j columns obtained by combining the submatrix ^Xanc1 _,j , the submatrix ^Xanc2 _,j , ..., the submatrix ^Xancc _,j in the column direction.

The above-mentioned anchor data is data that is distributed commonly to terminal 111, terminal 112, terminal 121, ..., terminal 1cd, and is data that does not indicate information that should be kept secret. Furthermore, it is preferable that the anchor data held by terminal 111 is data that is close to the partial data held by terminal 111. Similarly, it is preferable that the anchor data held by terminal 112, ..., and terminal 1cd is data that is close to the partial data held by terminal 112, ..., and terminal 1cd, respectively.

Here, when the anchor data is close to the partial data, it means that the range of values that the anchor data contains is close to the range of values that the partial data contains. Alternatively, when the anchor data is close to the partial data, it means that the attributes that the anchor data contains are close to the attributes that the partial data contains. However, the anchor data may be data that is generated randomly. Alternatively, the anchor data may be data that is generated by each of terminal 111, terminal 112, terminal 121, ..., terminal 1cd, and that is exchanged between terminal 111, terminal 112, terminal 121, ..., terminal 1cd.

Anchor data is, for example, a handwritten number image depicting relatively clearly written handwritten numbers, a medical image depicting a lesion specific to a specific disease, medical data obtained by a specific examination, or data related to a company's financial affairs. Medical images are, for example, X-ray images, CT (Computed Tomography) images, MRI (Magnetic Resonance Imaging) images, PET (Positron Emission Tomography) images, and ultrasound images. Medical data is, for example, a numerical value indicating the triglyceride level obtained by a blood test.

Furthermore, the feature indicated by the anchor data and the feature indicated by the partial data are, for example, numbers depicted in the handwritten number image or a pathology depicted in the medical image when the anchor data or the partial data is the above-mentioned handwritten number image or medical image. In this case, the feature amount is a pixel included in the handwritten number image or a pixel included in the medical image. Alternatively, the feature indicated by the anchor data and the feature indicated by the partial data are values outside the normal range when the anchor data or the partial data is the above-mentioned medical data. In this case, the feature amount is a numerical value included in the medical data. Alternatively, the feature indicated by the anchor data and the feature indicated by the partial data are a corporate rating when the anchor data or the partial data is the above-mentioned data relating to the financial aspects of the company. In this case, the feature amount is data relating to the financial aspects of the company.

Note that terminal 111 holds the anchor data and partial data, for example, in a storage medium that terminal 111 can access. Furthermore, this storage medium only needs to be accessible from terminal 111, and may or may not be included in terminal 111. Terminals 112, 121, ..., terminal 1cd each hold the anchor data and partial data in a storage medium in a similar manner.

The generation unit 1111 generates abstracted partial data represented by submatrix W _1,1 by applying a unique abstraction process represented by function f _1,1 to partial data represented by submatrix X _1,1 . Similarly, the generation unit 11ij generates abstracted partial data represented by submatrix W _i,j by applying a unique abstraction process represented by function f _i,j to partial data represented by submatrix X _i,j , as shown in the following formula (4). The submatrix W _i,j is a submatrix with n _i rows and p _i,j columns.

The abstracted partial data represented by the submatrix W _1,1 is numerical data obtained by abstracting the feature amounts included in the partial data represented by the submatrix X _1,1 using a unique abstraction process represented by a function f _1,1 . Similarly, the abstracted partial data represented by the submatrix W _i,j is numerical data obtained by abstracting the feature amounts included in the partial data represented by the submatrix X _{i,j using a unique abstraction process represented by a function f i, j} .

The generating unit 1111 generates abstraction anchor data represented by submatrix W ^anc _1,1 by applying a unique abstraction process represented by function f _1,1 to submatrix X ^anc _:,1 . Similarly, the generating unit 11 ij generates abstraction anchor data represented by submatrix W anc _i,j by applying a unique abstraction process represented by function f _i,j to submatrix X ^anc _:,j as shown in the following formula (5). The submatrix W ^{anc i} _,j is a submatrix with r rows and p _i,j columns. Furthermore, these abstraction anchor data correspond to the first abstraction anchor data and the second abstraction anchor data.

The abstracted anchor data represented by the submatrix W ^anc _1,1 is data indicating a linear combination of feature amounts selected from the feature amounts included in the anchor data represented by the submatrix X ^anc _1,1 by a unique abstraction process represented by the function f _1,1 . Similarly, the abstracted anchor data represented by the submatrix W ^anc _i,j is numerical data obtained by abstracting the feature amounts included in the anchor data represented by the submatrix X ^anc _i,j by a unique abstraction process represented by the function f _i,j .

The above-mentioned abstraction process uses a linear or nonlinear dimensionality reduction method that uses an objective function. Dimension reduction methods are broadly divided into supervised dimensionality reduction methods and unsupervised dimensionality reduction methods. Examples of supervised dimensionality reduction methods include Linear Discriminant Analysis and Local Fisher Discriminant Analysis. Examples of unsupervised dimensionality reduction methods include Principal Component Analysis, Local Linear Embedding, Local Tangent Space Alignment, and Locality Preserving Projections. Alternatively, the above-mentioned abstraction process uses a partial structure of deep learning.

The transmitting unit 1211 transmits the abstracted partial data represented by the submatrix _W1,1 and the abstracted anchor data represented by the submatrix W ^anc _1,1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting units 1212, ..., and 12cd transmit the abstracted partial data represented by the submatrix _W1,2 and the abstracted anchor data represented by the submatrix W ^anc _c,d , ..., the abstracted partial data represented by the submatrix _W1,2 and the abstracted anchor data represented by the submatrix W ^anc _c,d to the distributed data processing device 20 via the network NW.

Further, the transmitting unit 1211, the transmitting unit 1212, ..., the transmitting unit 121d transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting unit 12i1, the transmitting unit 12i2, ..., the transmitting unit 12id transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW.

The receiving unit 210 receives abstraction partial data represented by submatrix _W1,1 and abstraction anchor data represented by ^submatrix _Wanc1,1 from the terminal 111 via the network NW. Similarly, the receiving unit 210 receives abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wancc _,d , ..., abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wancc _,d from the terminals 112, ..., terminal 1cd via the network NW.

The receiving unit 210 also receives data indicating the submatrix _Y1 from each of the transmitting units 1211, 1212, ..., 121d via the network NW. Similarly, the receiving unit 210 receives data indicating the submatrix _Y1 from each of the transmitting units 12i1, 12i2, ..., 12id via the network NW.

The generation unit 220 generates a submatrix W anc i with r rows and p i columns by row-wise combining the abstraction anchor data represented by the submatrix W ^anc _i,1 , ..., the abstraction anchor data represented by the submatrix W ^{anc i} _{, d} , as shown in the following _equation (6).

Then, the generating unit 220 generates an integration function g _i that is used to generate integrated data, which will be described later, using the submatrix W ^anc _i .

Specifically, the generating unit 220 uses the temporary merging function _e1 to convert the submatrix W ^anc ₁ into temporary merging data represented by the submatrix W ^anc ₁ '. Similarly, the generating unit 220 uses the temporary merging functions _e2 , ..., temporary merging function e _c to convert the submatrix W ^anc ₂ , ..., submatrix W ^anc _c into temporary merging data represented by the submatrix W ^anc ₂ ', ..., temporary merging data represented by the submatrix W ^anc _c ', respectively. The temporary merging functions _e1 , ..., temporary merging function e _c are linear or nonlinear functions.

Next, the generating unit 220 calculates the provisional integration functions e ₁ , ..., e c that reduce the difference between the provisional integration data represented by the submatrix W ^anc ₁ ', ..., and the provisional integration data represented by the submatrix W ^anc _{c '} as the integration functions g1, ..., and gc. It is also preferable that the generating unit 220 calculates the provisional integration functions e ₁ , ..., and e ^c that reduce the difference between the provisional integration data represented by the submatrix W ^anc ₁ ', ..., and the provisional integration data represented by the submatrix W anc _{c '} as the integration functions _g1 , ..., _{and gc .}

The problem of calculating the above-mentioned integration functions g ₁ , ..., g _c is reduced to, for example, a minimization problem. In particular, when the provisional integration functions are linear functions, this problem is reduced to a generalized total least squares (TLS) problem.

On the other hand, the generation unit 220 generates a submatrix W i with n i rows and p i columns by combining the abstracted partial data represented by submatrices W _i,1 , ..., and the abstracted partial data _represented by submatrix W _{i ,d} in the _row direction, as shown in the following equation (7).

The generating unit 220 generates integrated data represented by a submatrix Z _i having n _i rows and q columns from the submatrix W _i by using an integration function g _i as shown in the following equation (8).

The generating unit 220 generates a matrix Z with n rows and q columns by combining the submatrices Z ₁ , . . . , Z _c in the column direction as shown in the following equation (9).

The generation unit 220 generates a first integrated analysis model h that outputs first label data, which corresponds to the contents indicated by elements included in the integrated data represented by the submatrix _Z1 , ..., the integrated data represented by the submatrix _Zc , and is data indicating the contents corresponding to each row of the matrix Y represented by the above-mentioned formula (2), based on the integrated data. The relationship between the matrix Z, the first integrated analysis model h, and the matrix Y is expressed by the following formula (10).

The first integrated analysis model h is an analysis model having a multi-stage structure generated using abstracted anchor data and abstracted partial data, and is an analysis model with relatively low interpretability and explainability. Interpretability is an index that gives a global explanation of the analysis model, and indicates the difficulty for a user to understand why the analysis results output by the analysis model are generated. Explainability is an index that gives a local explanation of the analysis model, and indicates the degree of influence that each explanatory variable input to the analysis model has on the analysis results by the analysis model.

The generation unit 220 generates integrated anchor data represented by a submatrix Z anc i with r rows and p columns from the submatrix W ^anc _i represented by the above-mentioned equation (6) by using the integration function g _i , as shown in the following equation ⁽¹¹ ₎ .

The generating unit 220 inputs the integrated anchor data represented by the submatrix Z ^anc _i to the above-mentioned first integrated analysis model h as shown in the following formula (12), and generates second label data corresponding to the contents indicated by the elements included in the integrated anchor data. The second label data is data indicating the contents corresponding to each row of the matrix Y ^anc _i on the left side of formula (12).

The transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc ₁ to the terminals 111, 112, ..., and 11d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc ₂ to the terminals 121, 122, ..., and 12d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc _c to the terminals 1c1, 1c2, ..., and 1cd. Note that the same applies to the second label data indicating the contents corresponding to each row of the matrix Y ^anc ₃ , ..., and the second label data indicating the contents corresponding to each row of the matrix Y ^anc _c-1 .

The receiving unit 1311, the receiving unit 1312, ..., the terminal 131d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc ₁ from the distributed data processing device 20. Similarly, the receiving unit 1321, the receiving unit 1322, ..., the receiving unit 132d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc ₂ from the distributed data processing device 20. Similarly, the receiving unit 13c1, the receiving unit 13c2, ..., the receiving unit 13cd each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc _c from the distributed data processing device 20. Note that these processes are similar for the second label data indicating the contents corresponding to each row of the matrix Y ^anc ₃ , ..., the second label data indicating the contents corresponding to each row of the matrix Y ^anc _c-1 .

Then, the generator 1111, the generator 1112, ..., the generator 111d generate a second integrated analysis model _t1 using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _the matrix ^Yanc1 . Similarly, the generator 1121, the generator 1122, ..., the generator 112d generate a second integrated analysis model t2 using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _the matrix ^Yanc2 . Similarly, the generator 11c1, the generator 11c2, ..., the generator 11cd generate a second integrated analysis model _tc using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _{the matrix} ^Yancc . These processes are similar for the other generators and are expressed by the following formula (13).

The second integrated analytical model _ti is an analytical model generated by modifying the first integrated analytical model h, which was generated by taking into consideration partial data including features not held by the terminals 1i1, 1i2, ..., 1id that generated the second integrated analytical model ti. Therefore, the second integrated analytical model _ti is an analytical model with relatively high interpretability and/or explainability.

Next, an example of processing executed by terminal 111, terminal 112, terminal 121, ..., terminal 1cd, and distributed data processing device 20 will be described with reference to FIG. 2. FIG. 2 is a sequence diagram showing an example of processing executed by the terminals and distributed data processing device according to the first embodiment of the present invention. In the description using FIG. 2, terminal 111 will be taken as an example, but the same applies to terminal 112, terminal 121, ..., terminal 1cd.

In step S10, the generation unit 1111 generates abstracted anchor data by performing an abstraction process on the anchor data.

In step S20, the generation unit 1111 generates abstracted partial data by performing an abstraction process on the partial data.

In step S30, the transmission unit 1211 transmits the abstraction anchor data and the abstraction partial data.

In step S40, the receiving unit 210 receives the abstraction anchor data and the abstraction partial data.

In step S50, the generation unit 220 generates integrated data by integrating multiple abstraction partial data using multiple abstraction anchor data.

In step S60, the generation unit 220 generates a first integrated analysis model that outputs first label data corresponding to the contents indicated by the elements included in the integrated data based on the integrated data.

In step S70, the generation unit 220 generates integrated anchor data by integrating multiple anchor data using multiple abstracted anchor data.

In step S80, the generation unit 220 inputs the integrated anchor data into the first integrated analysis model to generate second label data corresponding to the content indicated by the elements included in the integrated anchor data.

In step S90, the transmission unit 230 transmits the second label data.

In step S100, the generation unit 1111 generates a second integrated analysis model using the anchor data and the second label data.

The above describes the integrated analysis model generation system 1 according to the first embodiment. The distributed data processing device 20 generates integrated data by integrating multiple abstracted partial data using the abstracted anchor data received for each terminal, and generates a first integrated analysis model that outputs first label data corresponding to the content indicated by the elements included in the integrated data based on the integrated data. Next, the distributed data processing device 20 generates integrated anchor data by integrating multiple anchor data using the abstracted anchor data received for each terminal, and inputs the integrated anchor data into the first integrated analysis model to generate second label data corresponding to the content indicated by the elements included in the integrated anchor data. Then, the distributed data processing device 20 transmits the second label data to the terminal 111, etc. For example, the terminal 111 generates a second integrated analysis model using the anchor data and the second label data.

As a result, the integrated analysis model generation system 1 can generate an integrated analysis model that has relatively high interpretability and/or explainability.

In addition, the distributed data processing device 20 uses the same anchor data when generating integrated data and when generating integrated anchor data.

As a result, the integrated analysis model generation system 1 can generate an integrated analysis model that has relatively high interpretability and/or explainability while reducing the load of processing anchor data.

Next, we will explain the application of the integrated analysis model generation system according to the first embodiment to a first example, with reference to Figures 3 to 14.

The first example is a case where 3,200 pieces of artificial data, each of which has feature values "1" to "20," are classified into two classes based on feature values "1" and "11." This artificial data is an example of the overall data described above. In addition, in the first example, feature values "2" to "10" and feature values "12" to "20" are random numbers, and are features that are not necessary for classifying the 3,200 pieces of data into two classes.

Feature quantities "1" to "10" of the 1st to 1600th data are held by institution A1. Feature quantities "11" to "20" of the 1st to 1600th data are held by institution A2. Feature quantities "1" to "10" of the 1601st to 3200th data are held by institution B1. Feature quantities "11" to "20" of the 1601st to 3200th data are held by institution B2.

Figure 3 is a diagram showing an example of an integrated analysis model that outputs first label data corresponding to the contents indicated by elements contained in integrated data generated by a distributed data processing device according to a comparative example of the present invention by integrating four non-abstracted partial data received from each of four institutions, based on the integrated data. Figure 4 is a diagram showing an example of the results of classifying artificial data using the integrated analysis model shown in Figure 3. In Figure 4, the vertical axis indicates the value of feature "1", and the horizontal axis indicates the value of feature "11". Also, in Figure 4, data belonging to one class is indicated by black circles, and data belonging to the other class is indicated by white circles.

The decision tree shown in FIG. 3 is an integrated analysis model generated based on the partial data held by institution A1, the partial data held by institution A2, the partial data held by institution B1, and the partial data held by institution B2. As shown in FIG. 3, this decision tree conceptually represents an integrated analysis model that classifies artificial data based on feature value "1" and feature value "11". Therefore, as shown in FIG. 4, the decision tree shown in FIG. 3 can appropriately classify data included in areas where feature value "1" is -0.5 or more and feature value "11" is -0.5 or more from data included in areas other than these areas. However, the process executed by the decision tree shown in FIG. 3 cannot be actually executed because it requires an infeasible process of sharing partial data that is not abstracted.

Figure 5 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. Figure 6 is a diagram showing an example of the results of classifying artificial data using the analysis model shown in Figure 5. In Figure 6, the vertical axis shows the value of feature "1", and the horizontal axis shows the value of feature "11". Also, in Figure 6, data belonging to one class is shown with black circles, and data belonging to the other class is shown with white circles.

The decision tree shown in Figure 5 is an integrated analysis model generated based on partial data held by institution A1. As shown in Figure 5, this decision tree is a decision tree that classifies artificial data based on at least one of feature quantities "1" to "10," and does not take feature quantity "11" into consideration. Therefore, as shown in Figure 6, the decision tree shown in Figure 5 classifies data into data included in an area where feature quantity "1" is -0.5 or more and data included in an area other than that area, and is unable to perform classification based on the line where feature quantity "11" is -0.5 as the boundary.

Figure 7 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. Figure 8 is a diagram showing an example of the results of classifying artificial data using the analysis model shown in Figure 7. In Figure 8, the vertical axis indicates the value of feature "1", and the horizontal axis indicates the value of feature "11". Also, in Figure 8, data belonging to one class is indicated by black circles, and data belonging to the other class is indicated by white circles.

The decision tree shown in FIG. 7 is an integrated analysis model generated based on partial data held by institution A2. As shown in FIG. 7, this decision tree is a decision tree that classifies artificial data based on at least one of feature amounts "11" to "20", and does not take feature amount "1" into consideration. Therefore, as shown in FIG. 8, the decision tree shown in FIG. 7 classifies data into data included in the area where feature amount "11" is -0.5 or more and data included in the area other than said area, and is unable to perform classification using the line where feature amount "1" is -0.5 as the boundary. For this reason, as shown in FIG. 8, black circles are scattered among white circles in the area where feature amount "11" is -0.5 or more.

Figure 9 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. Figure 10 is a diagram showing an example of the results of classifying artificial data using the analysis model shown in Figure 9. In Figure 10, the vertical axis shows the value of feature "1", and the horizontal axis shows the value of feature "11". Also, in Figure 10, data belonging to one class is shown with black circles, and data belonging to the other class is shown with white circles.

The decision tree shown in FIG. 9 is an integrated analysis model generated based on partial data held by institution B1. As shown in FIG. 9, this decision tree is a decision tree that classifies artificial data based on at least one of feature quantities "2" to "10," and does not take into account feature quantities "1" and "11." Therefore, as shown in FIG. 10, the decision tree shown in FIG. 9 cannot perform classification based on the line where feature quantity "1" is -0.5 as a boundary, or the line where feature quantity "11" is -0.5 as a boundary. For this reason, black circles are scattered among the white circles, as shown in FIG. 10.

Fig. 11 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. Fig. 12 is a diagram showing an example of the results of classifying artificial data using the analysis model shown in Fig. 11. In Fig. 12, the vertical axis indicates the value of feature "1", and the horizontal axis indicates the value of feature "11". Also, in Fig. 12, data belonging to one class is indicated by black circles, and data belonging to the other class is indicated by white circles.

The decision tree shown in FIG. 11 is an integrated analysis model generated based on partial data held by institution B2. As shown in FIG. 11, this decision tree is a decision tree that classifies artificial data based on at least one of features "11" to "20", and does not take feature "1" into consideration. Therefore, as shown in FIG. 12, the decision tree shown in FIG. 11 classifies data into data included in the area where feature "11" is -0.5 or more and data included in the area other than that area, and cannot perform classification using the line where feature "1" is -0.5 as the boundary.

Fig. 13 is a diagram showing an example of a second integrated analysis model generated by a terminal using second label data generated by a distributed data processing device according to a first embodiment of the present invention based on the four abstracted partial data received from each of the four institutions. Fig. 14 is a diagram showing an example of the results of classifying artificial data using the second integrated analysis model shown in Fig. 13. In Fig. 14, the vertical axis indicates the value of feature "1", and the horizontal axis indicates the value of feature "11". Furthermore, in Fig. 14, data belonging to one class is shown with black circles, and data belonging to the other class is shown with white circles.

The decision tree shown in FIG. 13 is a second integrated analysis model generated based on the partial data held by institution A1, the partial data held by institution A2, the partial data held by institution B1, and the partial data held by institution B2. As shown in FIG. 13, this decision tree is a decision tree that classifies artificial data based on feature amount "1" and feature amount "11". Therefore, as shown in FIG. 14, like the decision tree shown in FIG. 3, the decision tree shown in FIG. 13 is able to appropriately classify data included in areas where feature amount "1" is -0.5 or more and feature amount "11" is -0.5 or more from data included in areas other than these areas. In addition, the decision tree shown in FIG. 13 has relatively high interpretability and/or explainability.

Next, we will explain the application of the integrated analysis model generation system according to the first embodiment to a second example, with reference to Figures 15 to 21.

Figure 15 shows an example of a company identification code, financial data, industry label, and rating according to the first embodiment of the present invention.

In the second example, each data and feature is working capital/total assets, retained earnings/total assets, pre-tax and interest earnings/total assets, market capitalization/book value of total debt, sales/total assets, or industry label for each company with a unique company identification code, and each company is classified according to its rating. The contents shown in the table in FIG. 15 are an example of the overall data mentioned above. In the second example, pre-tax and interest earnings/total assets and market capitalization/book value of total debt are particularly important in accurately classifying each company according to its rating.

Of the data shown in FIG. 15, the data of companies assigned with company identification codes "ID0001" through "ID3000" is training data used to generate the analytical model or the second analytical model. On the other hand, of the data shown in FIG. 15, the data of companies assigned with company identification codes "ID3001" through "ID3932" is used as test data for the analytical model or the second analytical model.

"aa0001", ..., "aa3932" in the second column from the left in Figure 15 represent the working capital/total assets of companies assigned the company identification codes "ID0001", ..., "ID3932", respectively. "bb0001", ..., "bb3932" in the third column from the left in Figure 15 represent the retained earnings/total assets of companies assigned the company identification codes "ID0001", ..., "ID3932", respectively. "cc0001", ..., "cc3932" in the fourth column from the left in Figure 15 represent the pre-tax earnings/total assets of companies assigned the company identification codes "ID0001", ..., "ID3932", respectively.

The "dd0001", ..., and "dd3932" in the fifth column from the left in FIG. 15 represent the market capitalization of stocks/book value of all liabilities of the companies to which the company identification codes "ID0001", ..., and "ID3932" are assigned, respectively. The "ee0001", ..., and "ee3932" in the sixth column from the left in FIG. 15 represent the sales/total assets of the companies to which the company identification codes "ID0001", ..., and "ID3932" are assigned, respectively. The numbers in the seventh column from the left in FIG. 15 represent the industry labels that indicate the industries of the companies to which the company identification codes "ID0001", ..., and "ID3932" are assigned, respectively. The letters in the eighth column from the left in FIG. 15 represent the ratings of the companies to which the company identification codes "ID0001", ..., and "ID3932" are assigned, respectively, when evaluated from a financial perspective.

In the second example, institution A1 holds data showing working capital/total assets, data showing retained earnings/total assets, and data showing earnings before interest and tax/total assets for each company assigned company identification codes "ID0001" through "ID1500." Institution A2 holds data showing market capitalization/book value of total liabilities, data showing sales/total assets, and data showing industry label for each company assigned company identification codes "ID0001" through "ID1500."

Institution B1 holds data showing working capital/total assets, retained earnings/total assets, and earnings before interest and tax/total assets for each company assigned company identification codes "ID1501" through "ID3000". Institution B2 holds data showing market capitalization/book value of total liabilities, sales/total assets, and industry label for each company assigned company identification codes "ID1501" through "ID3000".

Figure 16 is a diagram showing an example of an integrated analysis model that outputs first label data corresponding to the contents indicated by elements contained in integrated data generated by a distributed data processing device according to a comparative example of the present invention by integrating four non-abstracted partial data received from each of four institutions, based on the integrated data. The decision tree shown in Figure 16 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings, mainly based on earnings before interest and tax/total assets and market capitalization/book value of all liabilities, and therefore can classify these companies with relatively high accuracy. However, the process executed by the decision tree shown in Figure 16 cannot be actually executed because it requires an infeasible process of sharing non-abstracted partial data.

Figure 17 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. The decision tree shown in Figure 17 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings based on at least one of working capital/total assets, retained earnings/total assets, and earnings before interest and tax/total assets, and therefore may not be able to classify these companies with high accuracy.

Figure 18 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. The decision tree shown in Figure 17 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings based on at least one of market capitalization/book value of total debt, sales/total assets, and industry label, and therefore may not be able to classify these companies with high accuracy.

Figure 19 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents indicated by partial data received from an institution based on the partial data. The decision tree shown in Figure 19 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings based on at least one of working capital/total assets, retained earnings/total assets, and earnings before interest and tax/total assets, and therefore may not be able to classify these companies with high accuracy.

Figure 20 is a diagram showing an example of an analysis model in which a distributed data processing device according to a comparative example of the present invention outputs first label data corresponding to the contents of partial data received from an institution based on the partial data. The decision tree shown in Figure 20 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings based on at least one of market capitalization/book value of total debt, sales/total assets, and industry label, so it may not be possible to classify these companies with high accuracy.

Figure 21 is a diagram showing an example of a second integrated analysis model generated by a terminal using second label data generated by a distributed data processing device according to the first embodiment of the present invention based on four abstracted partial data received from each of four institutions. The decision tree shown in Figure 21 is an analysis model that classifies each of the companies shown in Figure 15 according to their ratings, mainly based on earnings before interest and tax/total assets and market capitalization/book value of all liabilities, and therefore can classify these companies with relatively high accuracy. In addition, the decision tree shown in Figure 21 has relatively high interpretability and/or explainability.

[Second embodiment]
Next, an example of an integrated analysis model generation system according to the second embodiment will be described. The integrated analysis model generation system according to the second embodiment differs from the integrated analysis model generation system according to the first embodiment in that different anchor data are used for the first anchor data and the second anchor data. Therefore, in the second embodiment, the differences from the first embodiment will be mainly described, the same symbols and mathematical symbols as in the first embodiment will be used, and descriptions of matters that overlap with the first embodiment will be omitted as appropriate.

Terminal 111, terminal 112, terminal 121, ..., terminal 1cd respectively possess anchor data represented by submatrix X ^anc-a _1,1 , anchor data represented by submatrix X ^anc-a _1,2 , anchor data represented by submatrix X ^anc-a _2,1 , ..., anchor data represented by submatrix X ^anc-a _c,d . These anchor data are first anchor data used to generate integrated data.

Moreover, these submatrices are square or rectangular matrices, and are submatrices of the matrix X ^{anc-a with} r _a rows and m columns (r _a : a natural number less than n) representing all anchor data. Therefore, the relationship between the matrix X ^anc-a and the submatrix X ^anc-a _1,1 , the submatrix X ^anc-a _1,2 , the submatrix X ^anc-a _2,1 , ..., the submatrix X ^anc-a _c,d is expressed by, for example, the following formula (14). Moreover, the submatrix X ^anc-a _i,j included in formula (14) is a matrix with r _a,i rows and m _j columns. Note that the submatrix X ^anc-a _:,j included in equation (14) is a submatrix with r a rows _and m _j columns obtained by combining the submatrix X ^anc-a _1,j , the submatrix X ^anc-a _2,j , ..., the submatrix X ^anc- _ac,j in the column direction.

Furthermore, terminal 111, terminal 112, terminal 121, ..., terminal 1cd respectively possess anchor data represented by submatrix X ^anc-b _1,1 , anchor data represented by submatrix X ^anc-b _1,2 , anchor data represented by submatrix X ^anc-b _2,1 , ..., anchor data represented by submatrix X ^anc-b _c,d . These anchor data are second anchor data used to generate integrated anchor data and are different from the first anchor data.

Moreover, these submatrices are square or rectangular matrices, and are submatrices of the matrix X ^anc-b with r _b rows and m columns (r _a : a natural number less than n) representing all anchor data. Therefore, the relationship between the matrix X ^anc-b and the submatrix X ^anc-b _1,1 , the submatrix X ^anc-b _1,2 , the submatrix X ^anc-b _2,1 , ..., the submatrix X ^anc-b _c,d is expressed by, for example, the following formula (15). Moreover, the submatrix X ^anc-b _i,j included in formula (15) is a matrix with r _b,i rows and m _j columns. Note that the submatrix X ^anc-b _:,j included in equation (15) is a submatrix with r b rows and m _j columns obtained by combining the submatrix X ^anc-b _1,j , the submatrix X ^anc-b _2,j , ..., the submatrix X ^anc-b _{c ,j} in the column direction.

The generation unit 1111 generates abstracted partial data represented by a submatrix _W1,1 by applying a unique abstraction process represented by a function _f1,1 to partial data represented by a submatrix _X1,1 . Similarly, the generation unit 11ij generates abstracted partial data represented by a submatrix Wi _,j by applying a unique abstraction process represented by a function fi _, j to partial data represented by a submatrix Xi _,j , as shown in the above-mentioned formula (4).

The generation unit 1111 generates abstraction anchor data represented by submatrix W ^anc-a _1,1 by performing a unique abstraction process represented by function f _1,1 on submatrix X ^anc-a _:,1 . Similarly, the generation unit 11ij generates abstraction anchor data represented by submatrix W anc-a i,j by performing a unique abstraction process represented by function f _i,j on submatrix X ^anc-a _{: ,j} as shown in the following formula (16). The submatrix W ^{anc-a i} _,j is a submatrix with r _a rows and p _i,j columns. Furthermore, these abstraction anchor data are first abstraction anchor data.

The generation unit 1111 generates abstraction anchor data represented by submatrix W ^anc-b _1,1 by performing a unique abstraction process represented by function f _1,1 on submatrix X ^anc-b _:,1 . Similarly, the generation unit 11ij generates abstraction anchor data represented by submatrix W anc-b i,j by performing a unique abstraction process represented by function f _i,j on submatrix X ^anc-b _{:, j} as shown in the following formula (17). The submatrix W ^{anc-b i} _,j is a submatrix with r _b rows and p _i,j columns. Furthermore, these abstraction anchor data are second abstraction anchor data.

The transmitting unit 1211 transmits the abstracted partial data represented by the submatrix W _1,1 and the abstracted anchor data represented by the submatrix W ^anc-a _1,1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting units 1212, ..., transmitting unit 12cd transmit the abstracted partial data represented by the submatrix W _1,2 and the abstracted anchor data represented by the submatrix W ^anc-a _c,d , ..., the abstracted partial data represented by the submatrix W _1,2 and the abstracted anchor data represented by the submatrix W ^anc-a _c,d to the distributed data processing device 20 via the network NW.

The transmitting unit 1211 transmits the abstracted partial data represented by the submatrix W _1,1 and the abstracted anchor data represented by the submatrix W ^anc-b _1,1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting units 1212, ..., transmitting unit 12cd transmit the abstracted partial data represented by the submatrix W _1,2 and the abstracted anchor data represented by the submatrix W ^anc-b _c,d , ..., the abstracted partial data represented by the submatrix W _1,2 and the abstracted anchor data represented by the submatrix W ^anc-b _c,d to the distributed data processing device 20 via the network NW.

Further, the transmitting unit 1211, the transmitting unit 1212, ..., the transmitting unit 121d transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting unit 12i1, the transmitting unit 12i2, ..., the transmitting unit 12id transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW.

The receiving unit 210 receives abstraction partial data represented by submatrix _W1,1 and abstraction anchor data represented by submatrix ^Wanc-a1,1 _from the terminal 111 via the network NW. Similarly, the receiving unit 210 receives abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wanc-a1 _,d , ..., abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wanc-a1 _,d from the terminals 112, ..., terminal 1cd via the network NW.

The receiving unit 210 receives abstraction partial data represented by submatrix W _1,1 and abstraction anchor data represented by submatrix W ^anc-b _1,1 from the terminal 111 via the network NW. Similarly, the receiving unit 210 receives abstraction partial data represented by submatrix W _1,2 and abstraction anchor data represented by submatrix W ^anc-b _c,d , ..., abstraction partial data represented by submatrix W _1,2 and abstraction anchor data represented by submatrix W ^anc-b _c,d from the terminals 112, ..., terminal 1cd via the network NW.

The receiving unit 210 also receives data indicating the submatrix _Y1 from each of the transmitting units 1211, 1212, ..., 121d via the network NW. Similarly, the receiving unit 210 receives data indicating the submatrix _Y1 from each of the transmitting units 12i1, 12i2, ..., 12id via the network NW.

The generation unit 220 generates a submatrix W anc-ai with r a rows and p i columns by row-wise combining the abstraction anchor data represented by the submatrix W ^anc-ai _,1 , ..., the _{abstraction anchor} data represented by the submatrix W ^{anc- ai} _{,d ,} as shown in the following equation (18).

In addition, the generation unit 220 generates a submatrix W ^anc- bi i with r a rows and p i _columns by row-wise combining the abstraction anchor data represented by the submatrix W anc-bi _,1 , ..., the abstraction anchor data represented by the _submatrix W ^{anc-bi ,} _{d ,} as shown in the following equation (19).

Then, the generating unit 220 generates an integration function g _i used to generate integrated data, which will be described later, using the submatrix W ^anc-a _i .

On the other hand, the generation unit 220 generates a submatrix W i with n i rows and p i columns by combining the abstracted partial data represented by the submatrix W _i,1 , ..., the abstracted partial data _represented by the submatrix W _{i ,d} in the row direction, as shown in the above _- mentioned equation (7).

The generating unit 220 generates integrated data represented by a submatrix Z _i with n _i rows and p columns from the submatrix W _i by using the integration function g _i as shown in the above-mentioned equation (8).

The generation unit 220 generates a first integrated analysis model h that outputs first label data, which corresponds to the contents indicated by elements included in the integrated data represented by the submatrix _Z1 , ..., the integrated data represented by the submatrix _Zc , and is data indicating the contents corresponding to each row of the matrix Y represented by the above-mentioned formula (2), based on the integrated data. The relationship between the matrix Z, the first integrated analysis model h, and the matrix Y is represented by the above-mentioned formula (10).

The generation unit 220 generates integrated anchor data represented by the submatrix Z anc-bi with r _b rows and q columns from the submatrix W ^anc- _bi represented by the above-mentioned equation (19) by using the integration function g _i , as shown in the following equation ⁽²⁰⁾ _.

The generation unit 220 inputs the integrated anchor data represented by the submatrix Z ^anc-bi _into the above-mentioned first integrated analysis model h as shown in the following formula (21), and generates second label data corresponding to the contents indicated by the elements included in the integrated anchor data. The second label data is data indicating the contents corresponding to each row of the matrix Y ^anc-bi _on the left side of formula (21).

The transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc-b ₁ to the terminals 111, 112, ..., and 11d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc-b ₂ to the terminals 121, 122, ..., and 12d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc-b _c to the terminals 1c1, 1c2, ..., and 1cd. Note that the same applies to the second label data indicating the contents corresponding to each row of the matrix Y ^anc-b ₃ , ..., and the second label data indicating the contents corresponding to each row of the matrix Y ^anc-b _c-1 .

The receiving unit 1311, the receiving unit 1312, ..., the terminal 131d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc-b ₁ from the distributed data processing device 20. Similarly, the receiving unit 1321, the receiving unit 1322, ..., the receiving unit 132d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc -b ₂ from the distributed data processing device 20. Similarly, the receiving unit 13c1, the receiving unit 13c2, ..., the receiving unit 13cd each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc-b _c from the distributed data processing device 20. Note that these processes are similar for the second label data indicating the contents corresponding to each row of the matrix Y ^anc-b ₃ , ..., the second label data indicating the contents corresponding to each row of the matrix Y ^anc-b _c-1 .

Then, the generator 1111, the generator 1112, ..., the generator 111d generate a second integrated analysis model _t1 using the second anchor data represented by the submatrix ^Xanc-b and the second label data indicating the contents corresponding to each row of the matrix ^Yanc-b1 _. Similarly, the generator 1121, the generator 1122, ..., the generator 112d generate a second integrated analysis model _t2 using the second anchor data represented by the submatrix ^Xanc-b and the second label data indicating the contents corresponding to each row of the matrix ^Yanc-b2 _. Similarly, the generator 11c1, the generator 11c2, ..., the generator 11cd generate a second integrated analysis model _tc using the second anchor data represented by the submatrix Xanc ^-b and the second label data indicating the contents corresponding to each row of the matrix ^Yanc-bc _. These processes are similar for the other generators and are expressed by the following formula (22).

The above describes the integrated analytical model generation system 1 according to the second embodiment. The integrated analytical model generation system 1 according to the second embodiment executes the same processing as the integrated analytical model generation system 1 according to the first embodiment, except that the first anchor data and the second anchor data are different. As a result, the integrated analytical model generation system 1 can generate an integrated analytical model with relatively high interpretability and/or explainability.

The distributed data processing device 20 according to the second embodiment also uses different anchor data when generating integrated data and when generating integrated anchor data. This allows the integrated analysis model generation system 1 to use both anchor data that is preferable for generating integrated data and anchor data that is preferable for generating integrated anchor data, thereby generating an integrated analysis model with at least one of higher interpretability and explainability.

[Third embodiment]
Next, an example of an integrated analysis model generation system according to a third embodiment will be described. The integrated analysis model generation system according to the third embodiment differs from the integrated analysis model generation system according to the first embodiment in that the integrated analysis model generation system represents partial data in a matrix representing the entire data, and at least one of the rows and columns of adjacent partial matrices may be different. Therefore, in the third embodiment, the differences from the first embodiment will be mainly described, the same symbols and mathematical symbols as in the first embodiment will be used, and descriptions of matters that overlap with the first embodiment will be omitted as appropriate.

Terminal 111, terminal 112, terminal 121, ..., terminal 1cd each have partial data represented by submatrix _X1,1 , partial data represented by submatrix _X1,2 , partial data represented by submatrix _X2,1 , ..., partial data represented by submatrix Xc _,d . In addition, at least one of the rows and columns of adjacent submatrices may differ between these submatrices. Therefore, submatrix Xi _,j is expressed by the following formula (23) including index set _I1 , index set _I2 , ..., index set _Ic related to data division and index set _J1 , index set _J2 , ..., index set _Jc related to feature division.

Moreover, the index set related to the above-mentioned data division has the properties expressed by the following formulas (24), (25), and (26). Formula (24) indicates that the contents corresponding to each row of the submatrix X _i,j do not correspond to any row of other submatrices X _i',j . Formula (25) indicates that the index set I ₁ , index set I ₂ , ..., or index set I _c related to the data division are assigned to all of the n pieces of data included in the entire data. Formula (26) indicates that the number of elements included in the index set I _i is equal to the number of rows n _i of the submatrix X _i,j .

Moreover, the index sets related to the above-mentioned feature division have the properties expressed by the following formulas (27) and (28). Formula (27) indicates that index set _J1 , index set _J2 , ..., or index set _Jdi related to feature division is assigned to all or a part of the m features included in the entire data. Formula (28) indicates that the number of elements included in index set _Ji is equal to the number _mj of columns of submatrix _Xi,j .

Figure 22 is a diagram showing an example of a partial matrix showing partial data according to the third embodiment of the present invention. The six-row, eight-column squares shown in Figure 22 represent each element of a six-row, eight-column rectangular matrix representing the entire data.

In the example shown in Fig. 22, the submatrix _X1,1 is a rectangular matrix with three rows and four columns, as indicated by the index set _I1 related to data division and the index set _J1,1 related to feature division. Also, the white circles shown in Fig. 22 indicate that the partial data represented by the submatrix _X1,1 is held by the first department of the first institution.

The submatrix _X1,2 is a square matrix with three rows and three columns, as indicated by the index set _I1 for data division and the index set _J1,2 for feature division. The asterisks in Fig. 22 indicate that the partial data represented by the submatrix _X1,2 is held by the second department of the first institution.

The submatrix _X2,1 is a square matrix with three rows and three columns, as indicated by the index set _I2 related to data division and the index set _J2,1 related to feature division. The clover symbol in Fig. 22 indicates that the partial data represented by the submatrix _X2,1 is held by the first department of the second institution.

The submatrix _X2,2 is a rectangular matrix with three rows and two columns, as indicated by the index set _I2 for data division and the index set _J2,2 for feature division. The heart mark in Fig. 22 indicates that the partial data represented by the submatrix _X2,2 is held by the second department of the second institution.

The submatrix _X2,3 is a square matrix with three rows and three columns, as indicated by the index set _I2 related to data division and the index set _J2,3 related to feature division. The heart mark in Fig. 22 indicates that the partial data represented by the submatrix _X2,2 is held by a third department of a second institution.

In the example shown in Figure 22, of the elements in the rectangular matrix of six rows and eight columns that represents the entire data, elements in the first row, eight columns, the second row, eight columns, and the third row, eight columns are missing.

Also, matrix Y having row vectors corresponding to the contents indicated by each row vector of matrix X is expressed, for example, by the above-mentioned formula (2).

Terminal 111, terminal 112, terminal 121, ..., terminal 1cd respectively possess anchor data represented by submatrix X ^anc _1,1 , anchor data represented by submatrix X ^anc _1,2 , anchor data represented by submatrix X ^anc _2,1 , ..., anchor data represented by submatrix X ^anc _c,d . These anchor data serve as first anchor data used to generate integrated data described later, and second anchor data used to generate integrated anchor data described later. In other words, the first anchor data and the second anchor data are the same.

The generation unit 1111 generates abstracted partial data represented by a submatrix _W1,1 by applying a unique abstraction process represented by a function _f1,1 to partial data represented by a submatrix _X1,1 . Similarly, the generation unit 11ij generates abstracted partial data represented by a submatrix Wi _,j by applying a unique abstraction process represented by a function fi _, j to partial data represented by a submatrix Xi _,j , as shown in the above-mentioned formula (4).

The generating unit 1111 generates abstraction anchor data represented by submatrix W ^anc _1,1 by applying a unique abstraction process represented by function f _1,1 to submatrix X ^anc _:,1 . Similarly, the generating unit 11ij generates abstraction anchor data represented by submatrix W ^anc _i,j by applying a unique abstraction process represented by function f _i, j to submatrix X ^anc _:,j as shown in the above-mentioned formula (5). The submatrix W ^anc _i,j is a submatrix with r rows and p _i,j columns. Furthermore, these abstraction anchor data correspond to the first abstraction anchor data and the second abstraction anchor data.

The transmitting unit 1211 transmits the abstracted partial data represented by the submatrix _W1,1 and the abstracted anchor data represented by the submatrix W ^anc _1,1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting units 1212, ..., and 12cd transmit the abstracted partial data represented by the submatrix _W1,2 and the abstracted anchor data represented by the submatrix W ^anc _c,d , ..., the abstracted partial data represented by the submatrix _W1,2 and the abstracted anchor data represented by the submatrix W ^anc _c,d to the distributed data processing device 20 via the network NW.

Further, the transmitting unit 1211, the transmitting unit 1212, ..., the transmitting unit 121d transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW. Similarly, the transmitting unit 12i1, the transmitting unit 12i2, ..., the transmitting unit 12id transmit data indicating the submatrix _Y1 to the distributed data processing device 20 via the network NW.

The receiving unit 210 receives abstraction partial data represented by submatrix _W1,1 and abstraction anchor data represented by ^submatrix _Wanc1,1 from the terminal 111 via the network NW. Similarly, the receiving unit 210 receives abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wancc _,d , ..., abstraction partial data represented by submatrix _W1,2 and abstraction anchor data represented by submatrix ^Wancc _,d from the terminals 112, ..., terminal 1cd via the network NW.

The receiving unit 210 also receives data indicating the submatrix _Y1 from each of the transmitting units 1211, 1212, ..., 121d via the network NW. Similarly, the receiving unit 210 receives data indicating the submatrix _Y1 from each of the transmitting units 12i1, 12i2, ..., 12id via the network NW.

The generation unit 220 generates a submatrix W anc i with r rows and p i columns by row-wise combining the abstraction anchor data represented by the submatrices W ^anc _i,1 , ..., and the abstraction anchor data represented by the submatrices W ^{anc i} _,di , as _shown in the following _equation (29).

Then, the generating unit 220 generates an integration function g _i that is used to generate integrated data, which will be described later, using the submatrix W ^anc _i .

On the other hand, the generation unit 220 generates a submatrix W i with n i rows and p i columns by combining the abstracted partial data represented by the submatrices W _i,1 , ..., and the abstracted partial data _represented by the submatrices W _{i ,di} in the row direction, as shown in the following _equation (30).

The generating unit 220 generates integrated data represented by a submatrix Z _i with n _i rows and p columns from the submatrix W _i by using the integration function g _i as shown in the above-mentioned equation (8).

The generation unit 220 generates a first integrated analysis model h that outputs first label data, which corresponds to the contents indicated by elements included in the integrated data represented by the submatrix _Z1 , ..., the integrated data represented by the submatrix _Zc , and is data indicating the contents corresponding to each row of the matrix Y represented by the above-mentioned formula (2), based on the integrated data. The relationship between the matrix Z, the first integrated analysis model h, and the matrix Y is represented by the above-mentioned formula (10).

As shown in the above-mentioned equation (11), the generation unit 220 generates integrated anchor data represented by the submatrix Z ^anc _i with r rows and p columns from the submatrix W ^anc _i represented by the above-mentioned equation (6) by using the integration function g _i .

As shown in the above-mentioned formula (12), the generation unit 220 inputs the integrated anchor data represented by the submatrix Z ^anc _i into the above-mentioned first integrated analysis model h, and generates second label data corresponding to the contents indicated by the elements included in the integrated anchor data.

The transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc ₁ to the terminals 111, 112, ..., and 11d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc ₂ to the terminals 121, 122, ..., and 12d. Similarly, the transmitting unit 230 transmits second label data indicating the contents corresponding to each row of the matrix Y ^anc _c to the terminals 1c1, 1c2, ..., and 1cd. Note that the same applies to the second label data indicating the contents corresponding to each row of the matrix Y ^anc ₃ , ..., and the second label data indicating the contents corresponding to each row of the matrix Y ^anc _c-1 .

The receiving unit 1311, the receiving unit 1312, ..., the terminal 131d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc ₁ from the distributed data processing device 20. Similarly, the receiving unit 1321, the receiving unit 1322, ..., the receiving unit 132d each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc ₂ from the distributed data processing device 20. Similarly, the receiving unit 13c1, the receiving unit 13c2, ..., the receiving unit 13cd each receive second label data indicating the contents corresponding to each row of the matrix Y ^anc _c from the distributed data processing device 20. Note that these processes are similar for the second label data indicating the contents corresponding to each row of the matrix Y ^anc ₃ , ..., the second label data indicating the contents corresponding to each row of the matrix Y ^anc _c-1 .

Then, the generator 1111, the generator 1112, ..., the generator 111d generate a second integrated analysis model _t1 using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _the matrix ^Yanc1 . Similarly, the generator 1121, the generator 1122, ..., the generator 112d generate a second integrated analysis model t2 using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _the matrix ^Yanc2 . Similarly, the generator 11c1, the generator 11c2, ..., the generator 11cd generate a second integrated analysis model _tc using the second anchor data represented by the submatrix ^Xanc and the second label data indicating the contents corresponding to each row of _{the matrix} ^Yancc . These processes are similar for the other generators and are expressed by the above-mentioned formula (13).

The above describes the integrated analysis model generation system 1 according to the third embodiment. The integrated analysis model generation system 1 according to the third embodiment represents partial data within a matrix representing the entire data, and executes the same processing as the integrated analysis model generation system 1 according to the first embodiment, except that adjacent partial matrices may have different rows and/or columns. This allows the integrated analysis model generation system 1 to generate an integrated analysis model with relatively high interpretability and/or explainability.

The distributed data processing device 20 according to the third embodiment also performs the above-mentioned processing on overall data that represents partial data and in which adjacent submatrices may differ in at least one of the rows and columns. As a result, the integrated analysis model generation system 1 can generate an integrated analysis model that has relatively high interpretability and/or explainability even when the partial data is held separately by multiple institutions, such as in the example shown in FIG. 22.

In the third embodiment, the submatrix indicating partial data is a square matrix or a rectangular matrix, but this is not limiting. FIG. 23 is a diagram showing an example of a submatrix indicating partial data according to another embodiment of the present invention. The white circle shown in FIG. 23 indicates that an element is present at that location. On the other hand, the horizontal line shown in FIG. 23 indicates that an element is missing from that location. As shown in FIG. 23, the submatrix indicating partial data may be a square matrix with some elements missing, or a rectangular matrix with some elements missing.

In this case, the abstracted anchor data is generated by performing an abstraction process on the anchor data represented by a matrix with some of the elements of a square matrix missing or a matrix with some of the elements of a rectangular matrix missing, and is represented by a matrix with some of the elements of a square matrix missing or a matrix with some of the elements of a rectangular matrix missing.

In this case, the abstracted partial data is generated by performing an abstraction process on the partial data represented by a matrix with some of the elements of a square matrix missing or a matrix with some of the elements of a rectangular matrix missing, and is represented by a matrix with some of the elements of a square matrix missing or a matrix with some of the elements of a rectangular matrix missing.

At least some of the functions of the terminals 111, ..., 1cd, and the distributed data processing device 20 may be realized by hardware including a circuit section (circuitry) executing a program. The hardware referred to here is, for example, a CPU (Central Processing Unit), an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a GPU (Graphics Processing Unit). The above-mentioned program is stored in a storage device having a storage medium. The storage medium referred to here is, for example, a HDD (Hard Disk Drive), a flash memory, a ROM (Read Only Memory), or a DVD (Digital Versatile Disc). Furthermore, the above-mentioned program may be a differential program that realizes some of the functions of the terminals 111, ..., 1cd, and the distributed data processing device 20.

The above describes an embodiment of the present invention with reference to the drawings. However, the integrated analysis model generation system 1 is not limited to the above-described embodiment, and various modifications, substitutions, combinations, or design changes can be made without departing from the spirit of the present invention.

1... Integrated analysis model generation system, 11, ..., 1n... Terminal, 111, ..., 11n... Transmitter, 121, ..., 12n... Receiver, 131, ..., 13n... Generator, 20... Distributed data processing device, 210... Receiver, 220... Generator, 230... Transmitter

Claims (10)

a receiving unit that executes, for each of the terminals, a process of receiving abstracted anchor data generated by applying an abstraction process to anchor data held by the terminal, and a process of receiving abstracted partial data generated by applying an abstraction process to partial data including a part of data included in overall data and a part of feature amounts included in the overall data, the partial data being divided and held by the multiple terminals;
a generation unit that generates integrated data by integrating a plurality of the abstract partial data using the abstraction anchor data received for each of the terminals, generates a first integrated analysis model that outputs first label data corresponding to content indicated by elements included in the integrated data based on the integrated data, generates integrated anchor data by integrating a plurality of the anchor data using the abstraction anchor data received for each of the terminals, and inputs the integrated anchor data into the first integrated analysis model to generate second label data corresponding to content indicated by elements included in the integrated anchor data;
a transmitting unit that transmits the second label data to the terminal;
A distributed data processing device comprising: The receiving unit executes, as a process of receiving the abstraction anchor data, a process of receiving first abstraction anchor data generated by applying an abstraction process to first anchor data held by the terminal, and a process of receiving second abstraction anchor data generated by applying an abstraction process to second anchor data held by the terminal and which is the same as the first anchor data, for each of the terminals;
The generation unit generates the integrated data by integrating a plurality of the abstraction partial data using the first abstraction anchor data received for each of the terminals, and generates the integrated anchor data by integrating a plurality of the anchor data using the second abstraction anchor data received for each of the terminals.
2. The distributed data processing system according to claim 1. The receiving unit executes, for each terminal, as a process of receiving the abstraction anchor data, a process of receiving first abstraction anchor data generated by applying an abstraction process to first anchor data held by the terminal, and a process of receiving second abstraction anchor data generated by applying an abstraction process to second anchor data held by the terminal and different from the first anchor data;
The generation unit generates the integrated data by integrating a plurality of the abstraction partial data using the first abstraction anchor data received for each of the terminals, and generates the integrated anchor data by integrating a plurality of the anchor data using the second abstraction anchor data received for each of the terminals.
2. The distributed data processing system according to claim 1. the receiving unit executes, for each of the terminals, a process of receiving the abstracted anchor data represented by a square matrix or a rectangular matrix, the abstracted anchor data being generated by performing an abstraction process on the anchor data represented by a square matrix or a rectangular matrix, and a process of receiving the abstracted partial data being generated by performing an abstraction process on the partial data represented by a square matrix or a rectangular matrix, the abstracted partial data being generated by performing an abstraction process on the partial data represented by a square matrix or a rectangular matrix,
4. A distributed data processing device according to claim 1. The receiving unit executes, for at least one of the terminals, a process of receiving the abstracted anchor data, which is generated by applying an abstraction process to the anchor data represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing, and which is represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing; and a process of receiving the abstracted partial data, which is generated by applying an abstraction process to the partial data represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing, and which is represented by a matrix in which some of the elements of a square matrix are missing or a matrix in which some of the elements of a rectangular matrix are missing.
4. A distributed data processing device according to claim 1. a transmitting unit that transmits abstracted anchor data generated by performing an abstraction process on the anchor data and abstracted partial data generated by performing an abstraction process on partial data including a part of data included in the entire data held in a divided manner on a plurality of terminals and a part of feature amounts included in the entire data to the distributed data processing device;
a receiving unit that inputs integrated anchor data generated by integrating a plurality of anchor data using the abstraction anchor data received by the distributed data processing device for each of the terminals into a first integrated analysis model that outputs, based on the integrated data, first label data corresponding to content indicated by an element included in integrated data generated by integrating a plurality of the abstracted partial data using the abstraction anchor data received by the distributed data processing device for each of the terminals, and receives second label data corresponding to content indicated by an element included in the integrated anchor data from the distributed data processing device;
a generation unit that generates a second integrated analysis model using the anchor data and the second label data;
A terminal comprising: On the computer,
a receiving function for each of the terminals, which executes the following processes: a process for receiving abstracted anchor data generated by applying an abstraction process to anchor data held by the terminal; and a process for receiving abstracted partial data generated by applying an abstraction process to partial data including a part of data included in overall data and a part of features included in the overall data, the partial data being divided and held by the multiple terminals;
a generation function of generating integrated data by integrating a plurality of the abstract partial data using the abstraction anchor data received for each of the terminals, generating a first integrated analysis model that outputs first label data corresponding to the content indicated by the elements included in the integrated data based on the integrated data, generating integrated anchor data by integrating a plurality of the anchor data using the abstraction anchor data received for each of the terminals, and inputting the integrated anchor data into the first integrated analysis model to generate second label data corresponding to the content indicated by the elements included in the integrated anchor data;
a transmission function for transmitting the second label data to the terminal;
A distributed data processing program that executes the above steps. On one device,
a transmission function for transmitting abstracted anchor data generated by performing an abstraction process on the anchor data and abstracted partial data generated by performing an abstraction process on partial data including a part of data included in the overall data held in a divided manner on a plurality of terminals and a part of feature amounts included in the overall data to a distributed data processing device;
a receiving function for inputting integrated anchor data generated by integrating a plurality of the anchor data using the abstraction anchor data received by the distributed data processing device for each of the terminals into a first integrated analysis model that outputs, based on the integrated data, first label data corresponding to content indicated by an element included in integrated data generated by integrating a plurality of the abstraction partial data using the abstraction anchor data received by the distributed data processing device for each of the terminals, and receiving second label data corresponding to content indicated by an element included in the integrated anchor data from the distributed data processing device;
a generation function for generating a second integrated analysis model using the anchor data and the second label data;
A terminal control program that executes the following: A method of information processing performed by a distributed data processing device, which is a computer device, under the control of software, comprising the steps of:
The distributed data processing device executes, for each of the terminals, a process of receiving abstracted anchor data generated by applying an abstraction process to anchor data held by the terminal, and a process of receiving abstracted partial data generated by applying an abstraction process to partial data including a part of data included in overall data held in a divided manner among the multiple terminals and a part of feature amounts included in the overall data;
the distributed data processing device generates integrated data by integrating a plurality of the abstracted partial data using the abstracted anchor data received for each of the terminals, generates a first integrated analysis model that outputs first label data corresponding to content indicated by elements included in the integrated data based on the integrated data, generates integrated anchor data by integrating a plurality of the anchor data using the abstracted anchor data received for each of the terminals, and inputs the integrated anchor data into the first integrated analysis model to generate second label data corresponding to content indicated by elements included in the integrated anchor data,
the distributed data processing device transmits the second label data to the terminal;
A method for distributed data processing. A method of information processing performed by a terminal that is a computer device under the control of software, comprising the steps of:
the terminal transmits to a distributed data processing device abstracted anchor data generated by performing an abstraction process on the anchor data, and abstracted partial data generated by performing an abstraction process on partial data including a part of data included in the overall data held in a divided manner on the multiple terminals and a part of the feature amount included in the overall data;
the terminal inputs integrated anchor data generated by integrating a plurality of the anchor data using the abstraction anchor data received by the distributed data processing device for each of the terminals into a first integrated analysis model that outputs, based on the integrated data, first label data corresponding to content indicated by elements included in integrated data generated by integrating a plurality of the abstracted partial data using the abstraction anchor data received by the distributed data processing device for each of the terminals, and receives second label data corresponding to content indicated by elements included in the integrated anchor data from the distributed data processing device;
The terminal generates a second integrated analytical model using the anchor data and the second label data.
Terminal control methods.

Images