JP7792688B2 - Terminal, data processing device, terminal control program, data processing program, terminal control method, and data processing method - Google Patents

Description

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

In recent years, social networking has progressed rapidly, creating an environment for collecting data in various sectors of society. For example, various institutions, such as medical institutions and financial institutions, independently collect and store large amounts of data. Against this background, there are hopes 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 distributed data integration device disclosed in Patent Document 1.

This distributed data integration device comprises an acquisition unit, an anchor data conversion unit, a calculation unit, and an analysis target data conversion unit. The acquisition unit acquires, for each piece of analysis target data, an anchor data intermediate representation, which is an intermediate representation obtained by converting anchor data, which is data commonly used in integrating multiple pieces of distributed analysis target data, using a first function, and an analysis target intermediate representation, which is an intermediate representation obtained by converting the analysis target data using the first function. The anchor data conversion unit converts the multiple anchor data intermediate representations acquired by the acquisition unit for each piece of analysis target data using a second function. The calculation unit calculates a second function for each piece of analysis target data that minimizes the difference between the anchor data intermediate representations converted by the anchor data conversion. The analysis target data conversion unit converts the analysis target intermediate representation acquired by the acquisition unit for each piece of analysis target data using the second function calculated by the calculation unit.

International Publication No. 2020/137728

However, because the distributed data integration device described above is capable of comparing the original data with the intermediate representation to be analyzed, it cannot completely eliminate the possibility that the content represented by the original data may become known to persons other than the institution that holds the original data.

The present invention was made in light of these circumstances, and aims to provide a terminal, data processing device, terminal control program, data processing program, terminal control method, and data processing method that make it impossible to compare data held by an institution or the like with data that has undergone abstraction processing.

One aspect of the present invention includes an abstraction function generation unit that generates an abstraction function used to perform abstraction processing on partial data, first anchor data, and second anchor data; an abstraction execution unit that applies the abstraction function to the partial data to generate abstracted partial data, applies the abstraction function to the first anchor data to generate first abstracted anchor data, and applies the abstraction function to the second anchor data to generate second abstracted anchor data; an abstraction function discarding unit that discards the abstraction function after the abstracted partial data, the first abstracted anchor data, and the second abstracted anchor data have been generated; and The terminal includes a data sending unit that sends first abstraction anchor data, the second abstraction anchor data, and label data to a data processing device; a data receiving unit that receives from the data processing device second label data that indicates labels corresponding to content indicated by elements included in integrated second anchor data generated by integrating multiple second abstraction anchor data using multiple first abstraction anchor data; and a model generating unit that uses the second anchor data and the second label data to generate an integrated analysis model that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data.

Furthermore, in the above-mentioned terminal, the abstraction function discarding unit discards the abstraction function before the abstraction partial data, the first abstraction anchor data, and the second abstraction anchor data are transmitted to the data processing device.

Furthermore, in the above-mentioned terminal, the abstraction function generation unit generates the abstraction function by at least one of a process of randomly extracting a portion of the partial data and a process of adding a random element to the abstraction function.

Furthermore, in the above-mentioned terminal, the abstraction execution unit randomly interchanges the rows of the abstraction partial data and the label data equally.

One aspect of the present invention is a data processing device comprising: a data receiving unit that receives abstraction partial data, first abstraction anchor data, second abstraction anchor data, and label data for each terminal; a label data generating unit that generates integrated data by integrating multiple pieces of abstraction partial data using the first abstraction anchor data received for each terminal; generates integrated second anchor data by integrating multiple pieces of second abstraction anchor data using the first abstraction anchor data received for each terminal; generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data; and generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model; and a data transmitting unit that transmits the second label data to the terminal.

One aspect of the present invention includes an abstraction function generation function that generates an abstraction function used to perform abstraction processing on partial data, first anchor data, and second anchor data; an abstraction execution function that applies the abstraction function to the partial data to generate abstracted partial data, applies the abstraction function to the first anchor data to generate first abstracted anchor data, and applies the abstraction function to the second anchor data to generate second abstracted anchor data; an abstraction function discarding function that discards the abstraction function after the abstracted partial data, the first abstracted anchor data, and the second abstracted anchor data have been generated; and an abstraction function discarding function that discards the abstraction function after the abstracted partial data, the first abstracted anchor data, and the second abstracted anchor data have been generated. The terminal control program causes a terminal to execute the following functions: a data transmission function that transmits car data, the second abstraction anchor data, and label data to a data processing device; a data reception function that receives from the data processing device second label data that indicates labels corresponding to content indicated by elements included in integrated second anchor data generated by integrating multiple second abstraction anchor data using multiple first abstraction anchor data; and a model generation function that uses the second anchor data and the second label data to generate an integrated analysis model that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data.

One aspect of the present invention is a data processing program that causes a data processing device to execute: a data receiving function that receives abstraction partial data, first abstraction anchor data, second abstraction anchor data, and label data for each terminal; a label data generation function that generates integrated data by integrating multiple pieces of abstraction partial data using the first abstraction anchor data received for each terminal; generates integrated second anchor data by integrating multiple pieces of second abstraction anchor data using the first abstraction anchor data received for each terminal; generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data; and generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model; and a data transmission function that transmits the second label data to the terminal.

One aspect of the present invention is a terminal control method executed by a computer, wherein the computer generates, using an abstraction function generation unit or abstraction function generation function, an abstraction function used to perform abstraction processing on partial data, first anchor data, and second anchor data; the computer uses an abstraction execution unit or abstraction execution function to apply the abstraction function to the partial data to generate abstracted partial data, apply the abstraction function to the first anchor data to generate first abstract anchor data, and apply the abstraction function to the second anchor data to generate second abstract anchor data; the computer uses an abstraction function discarding unit or abstraction function to discard the abstraction function after the abstracted partial data, the first abstract anchor data, and the second abstract anchor data have been generated; This is a terminal control method in which the computer uses a data transmitting unit or data transmitting function to transmit the abstraction partial data, the first abstraction anchor data, the second abstraction anchor data, and label data to a data processing device; the computer uses a data receiving unit or data receiver to receive from the data processing device second label data indicating labels corresponding to content indicated by elements included in integrated second anchor data generated by integrating a plurality of the second abstraction anchor data using a plurality of the first abstraction anchor data; and the computer uses a model generation unit or model generation function to generate an integrated analysis model indicating the relationship between the content indicated by the second anchor data and the content indicated by the second label data.

One aspect of the present invention is a data processing method executed by a computer, in which the computer receives abstraction partial data, first abstraction anchor data, second abstraction anchor data, and label data for each terminal using a data receiving unit or data receiving function; the computer generates integrated data by integrating multiple pieces of abstraction partial data using the first abstraction anchor data received for each terminal using a label data generating unit or label data generating function; generates integrated second anchor data by integrating multiple pieces of second abstraction anchor data using the first abstraction anchor data received for each terminal; generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data; generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model; and the computer transmits the second label data to the terminal using a data transmitting unit or data transmitting function.

The present invention provides a terminal, data processing device, terminal control program, data processing program, terminal control method, and data processing method that make it impossible to compare data held by an institution or the like with data that has undergone abstraction processing.

FIG. 1 is a diagram illustrating an example of an integrated analysis model generation system according to an embodiment of the present invention. FIG. 10 is a diagram showing an example of overall data according to an embodiment of the present invention. FIG. 10 is a diagram illustrating an example of partial data according to the embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a functional configuration of a terminal according to an embodiment of the present invention. FIG. 1 is a diagram illustrating an example of a functional configuration of a data processing device according to an embodiment of the present invention. FIG. 4 is a sequence diagram showing an example of processing executed by a terminal and a data processing device according to an embodiment of the present invention. FIG. 10 is a diagram showing an example of the relationship between the number of institutions and the accuracy rate when abstracted partial data indicating corporate ratings is integrated and analyzed using an integrated analysis model generated by a terminal according to an embodiment of the present invention. FIG. 10 is a diagram showing an example of the relationship between the number of positions and the accuracy rate when abstracted partial data representing handwritten numbers is integrated and analyzed using an integrated analysis model generated by a terminal according to an embodiment of the present invention.

[Embodiment]
First, an overview of an integrated analytical model generation system according to an embodiment will be described with reference to Figures 1 to 3. Figure 1 is a diagram showing an example of an integrated analytical model generation system according to an embodiment of the present invention. As shown in Figure 1, the integrated analytical model generation system 1 includes a terminal 10-1, ..., a terminal 10-c (c: natural number equal to or greater than 2), and a data processing device 20.

Terminals 10-1, ..., 10-c and data processing device 20 are all connected to the network NW shown in Figure 1 and are capable of communicating with each other. The network NW may be, for example, the Internet, a WAN (Wide Area Network), an intranet, or a LAN (Local Area Network).

Terminals 10-1, ... and 10-c are computers, such as servers, managed by different organizations or institutions. The institutions referred to here are, for example, a corporate rating agency or a medical institution. Data processing device 20 is a computer, such as a server, that collects and analyzes partial data from terminals 10-1, ... and 10-c.

Terminal 10-1, ... and terminal 10-c each possess partial data X ₁ , ... and partial data X _c, which are parts of the overall data. Partial data X ₁ , ... and partial data X _c each represent information that needs to be kept secret by the institution that manages terminal 10-1, ... or terminal 10-c.

Figure 2 is a diagram showing an example of overall data according to an embodiment of the present invention. For example, as shown in Figure 2, the overall data has a structure in which multiple feature quantities are defined for each sample, and is expressed as a matrix. For example, if the overall data is data related to a company's financial aspects, these feature quantities are indices that represent features related to the company's rating. Alternatively, if the overall data is data related to handwritten numeric imagery depicting handwritten numeric characters, these feature quantities are indices that represent parameters related to the colors assigned to each pixel included in the handwritten numeric imagery.

Fig. 3 is a diagram showing an example of partial data according to an embodiment of the present invention. The partial data is data indicating all feature quantities defined for a portion of samples included in the entire data, and is expressed as a matrix, as shown in Fig. 3. The entire data shown in Fig. 2 is divided into partial data _X1 , ... and partial data _Xc shown in Fig. 3, for example.

In addition, terminal 10-1, ... and terminal 10-c all possess first anchor data. The first anchor data is data that is distributed commonly to terminal 10-1, ... and terminal 10-c, and does not indicate any information that needs to be kept secret by the institutions that manage terminal 10-1, ... or terminal 10-c.

Furthermore, it is preferable that the first anchor data is data close to the partial data _X1 , ... and partial data _Xc . The first anchor data being data close to the partial data _X1 , ... and partial data _Xc means that the range of values that can be included in the first anchor data is close to the range of values that can be included in the partial data _X1 , ... and partial data _Xc . Alternatively, the first anchor data being data close to the partial data _X1 , ... and partial data _Xc means that the attribute included in the first anchor data is close to the attribute included in the partial data _X1 , ... and partial data _Xc .

However, the first anchor data may be randomly generated data. For example, the first anchor data may be data indicating random numbers in the range from the minimum value to the maximum value of the feature amount in at least one of the partial data X ₁ , ... and the partial data X _c . Alternatively, the first anchor data may be data generated by applying low-rank approximation to the partial data X ₁ , ... or the partial data X _c and then adding perturbation. Furthermore, the first anchor data may be data generated by each of the terminals 10-1, ... and 10-c and mutually exchanged between the terminals 10-1, ... and 10-c.

In addition, terminal 10-1, ... and terminal 10-c all possess second anchor data. The second anchor data is data that is distributed commonly to terminal 10-1, ... and terminal 10-c, and does not indicate any information that needs to be kept secret by the institutions that manage terminal 10-1, ... or terminal 10-c.

Furthermore, it is preferable that the second anchor data is data close to the partial data _X1 , ... and partial data _Xc . The second anchor data being close to the partial data _X1 , ... and partial data _Xc means that the range of values that can be included in the second anchor data is close to the range of values that can be included in the partial data _X1 , ... and partial data _Xc . Alternatively, the second anchor data being close to the partial data _X1 , ... and partial data _Xc means that the attribute included in the second anchor data is close to the attribute included in the partial data _X1 , ... and partial data _Xc .

However, the second anchor data may be randomly generated data. For example, the second anchor data may be data indicating random numbers in the range from the minimum value to the maximum value of the feature amount in at least one of the partial data X ₁ , ... and the partial data X _c . Alternatively, the second anchor data may be data generated by applying low-rank approximation to the partial data X ₁ , ... or the partial data X _c and then adding perturbation. Furthermore, the second anchor data may be data generated by each of the terminals 10-1, ... and 10-c and mutually exchanged between the terminals 10-1, ... and 10-c.

Note that the first anchor data and the second anchor data may be the exact same data or different data. Furthermore, terminal 10-1, for example, stores partial data X ₁ , the first anchor data, and the second anchor data in a storage medium that it can access. This storage medium only needs to be accessible from terminal 10-1, and may or may not be included in terminal 10-1. Terminals 10-2, ..., 10-c similarly store partial data X ₂ , the first anchor data and the second anchor data, ..., partial data X _c , the first anchor data, and the second anchor data in their storage media, respectively.

Next, an example of the functional configuration of an integrated analysis model generation system according to an embodiment will be described with reference to Figures 4 and 5.

Figure 4 is a diagram showing an example of the functional configuration of a terminal according to an embodiment of the present invention. As shown in Figure 4, terminal 10-1 includes an abstraction function generation unit 11-1, an abstraction execution unit 12-1, an abstraction function discarding unit 13-1, a data transmission unit 14-1, a data reception unit 15-1, and a model generation unit 16-1.

Similarly, terminals 10-2, ..., 10-c each include an abstraction function generation unit 11-2, an abstraction execution unit 12-2, an abstraction function discarding unit 13-2, a data transmission unit 14-2, a data reception unit 15-2, and a model generation unit 16-2, and ..., an abstraction function generation unit 11-c, an abstraction execution unit 12-c, an abstraction function discarding unit 13-c, a data transmission unit 14-c, a data reception unit 15-c, and a model generation unit 16-c.

Figure 5 is a diagram showing an example of the functional configuration of a data processing device according to an embodiment of the present invention. As shown in Figure 5, the data processing device 20 includes a data receiving unit 21, a label data generating unit 22, and a data transmitting unit 23.

The abstraction function generation unit 11-1 generates an abstraction function _f1 used to perform abstraction processing on the partial data _X1 , the first anchor data, and the second anchor data. Similarly, the abstraction function generation units 11-2, ..., and 11-c generate abstraction functions _f2 , ..., and _fc, respectively.

Furthermore, at least one of the abstraction function generation units 11-1, ... and 11-c may generate an abstraction function by at least one of a process of randomly extracting a portion of the partial data and a process of imparting a random element to the abstraction function. The process of randomly extracting a portion of the partial data is, for example, downsampling. The process of imparting a random element to the abstraction function is, for example, a process of imparting a random perturbation to the abstraction function.

Furthermore, at least one of the abstraction function generation units 11-1, ... and 11-c may randomly extract a portion of the partial data according to rules determined by a predetermined method based on, for example, the time when the process of generating the abstraction function started. Furthermore, at least one of the abstraction function generation units 11-1, ... and 11-c may impart a random perturbation to the abstraction function determined by a predetermined method based on, for example, the time when the process of generating the abstraction function started.

It is preferable that these times be measured in fine units, such as microseconds. This is because if times measured in fine units are used, it will be difficult for the person managing terminal 10-1, ... or terminal 10-c, to understand the relationship between time and randomness, making it difficult to generate the same abstraction function and difficult to derive the information indicated by the partial data from the abstract partial data, which will be described later.

The abstraction execution unit 12-1 applies an abstraction function _f1 to partial data _X1 to generate abstracted partial data _W1 . Similarly, the abstraction execution units 12-2, ... and 12-c apply abstraction functions _f2 , ... and _fc to partial data _X2 , ... and partial data _Xc , respectively, to generate abstracted partial data _W2 , ... and abstracted partial data _Wc .

The abstracted partial data _W1 is numerical data obtained by abstracting the feature quantities included in the partial data _X1 using a unique abstraction process represented by an abstraction function _f1 . Similarly, the abstracted partial data _W2 is numerical data obtained by abstracting the feature quantities included in the partial data _X2 using a unique abstraction process represented by an abstraction function _f2 . Furthermore, the abstracted partial data _Wc is numerical data obtained by abstracting the feature quantities included in the partial data _Xc using a unique abstraction process represented by an abstraction function _fc .

The abstraction execution unit 12-1 may equally and randomly interchange rows of the abstraction partial data _W1 and the label data _Y1 . Similarly, the abstraction execution units 12-2, ... and 12-c may equally interchange rows of the abstraction partial data _W1 and the label data _Y2 , ... and the abstraction partial data _Wc and the label data _YC , respectively.

Furthermore, the abstraction execution unit 12-1 applies an abstraction function f ₁ to the first anchor data to generate abstracted first anchor data. Similarly, the abstraction execution units 12-2, ... and 12-c apply abstraction functions f ₂ , ... and abstraction function f _c to the first anchor data, ... and first anchor data, respectively, to generate abstracted first anchor data, ... and abstracted first anchor data.

The first abstract anchor data W ^anc1 ₁ generated by terminal 10-1 is data indicating a linear combination of feature quantities included in the first anchor data by a unique abstraction process represented by abstraction function f _1. Similarly, the first abstract anchor data W ^anc1 ₂ generated by terminal 10-2 is data indicating a linear combination of feature quantities included in the first anchor data by a unique abstraction process represented by abstraction function f _2. Furthermore, the first abstract anchor data W ^anc1 _c generated by terminal 10-c is data indicating a linear combination of feature quantities included in the first anchor data by a unique abstraction process represented by abstraction function f _c .

Furthermore, the abstraction execution unit 12-1 applies an abstraction function f ₁ to the second anchor data to generate abstracted second anchor data. Similarly, the abstraction execution units 12-2, ... and 12-c apply abstraction functions f ₂ , ... and abstraction function f _c to the second anchor data, ... and second anchor data, respectively, to generate abstracted second anchor data, ... and abstracted second anchor data.

The second abstract anchor data W ^anc2 ₁ generated by terminal 10-1 is data indicating a linear combination of feature quantities included in the second anchor data by a unique abstraction process represented by abstraction function f _1. Similarly, the second abstract anchor data W ^anc2 ₂ generated by terminal 10-2 is data indicating a linear combination of feature quantities included in the second anchor data by a unique abstraction process represented by abstraction function f _2. Furthermore, the second abstract anchor data W ^anc2 _c generated by terminal 10-c is data indicating a linear combination of feature quantities included in the second anchor data by a unique abstraction process represented by abstraction function f _c .

The above-mentioned abstraction process uses a linear or nonlinear dimensionality reduction method that uses an objective function. Dimensionality reduction methods are broadly divided into supervised 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 t-distributed Stochastic Neighbor Embedding (t-SNE), Principal Component Analysis, Local Linear Embedding, Local Tangent Space Alignment, and Locality Preserving Projections. Alternatively, the above-mentioned abstraction process uses a substructure of deep learning.

The abstraction function discarding unit 13-1 discards the abstraction function f ₁ after the abstraction partial data W ₁ , the first abstraction anchor data W ^anc1 ₁ , and the second abstraction anchor data W ^anc2 ₁ have been generated. Similarly, the abstraction function discarding units 13-2, ... and 13-c discard the abstraction functions f 2, ... and f c after the abstraction partial data W ₂ , the first abstraction anchor data W ^anc1 ₂ , the second abstraction anchor data W ^anc2 ₂ , ..., the abstraction partial data W _c , the _first abstraction anchor data W ^anc1 _c , and the second abstraction anchor data _W ^anc2 _c have been generated, respectively.

Furthermore, it is preferable that the abstraction function discarding unit 13-1 discards the abstraction function f 1 before the abstraction partial data W ₁ , the first abstraction anchor data W ^anc1 ₁ , and the second abstraction anchor data W ^anc2 ₁ are transmitted to the data processing device 20. Similarly, it is preferable that the abstraction function discarding units 13-2, ... and 13-c discard the abstraction functions f ₂ , ... and the abstraction function f _c before the abstraction partial data W ₂ , the first abstraction anchor data W ^anc1 ₂ , and the second abstraction anchor data W ^anc2 ₂ , ..., the abstraction partial data W _c , the first abstraction anchor data W ^anc1 _c , and the second abstraction anchor data _W ^anc2 _c are transmitted to the data processing device 20, respectively.

The data transmitting unit 14-1 transmits the abstraction partial data W ₁ , the first abstraction anchor data W ^anc1 ₁ , the second abstraction anchor data W ^anc2 ₁ and the label data Y ₁ to the data processing device 20. Similarly, the data transmitting units 14-2, ... and 14-c transmit the abstraction partial data W ₂ , the first abstraction anchor data W ^anc1 ₂ , the second abstraction anchor data W ^anc2 ₂ and the label data Y ₂ , ..., the abstraction partial data W _c , the first abstraction anchor data W ^anc1 _c , the second abstraction anchor data W ^anc2 _c and the label data Y _c to the data processing device 20, respectively.

The data receiving unit 21 receives abstraction partial data W ₁ , first abstraction anchor data W ^anc1 ₁ , second abstraction anchor data W ^anc2 ₁ and label data Y ₁ from terminal 10-1. Similarly, the data receiving unit 21 receives abstraction partial data W ₂ , first abstraction anchor data W ^anc1 ₂ , second abstraction anchor data W ^anc2 _c and label data Y ₂ ..., abstraction partial data W _c , first abstraction anchor data W ^anc1 _c , second abstraction anchor data W ^anc2 _c and label data Y _C from terminal 10-2 ... and terminal 10-c, respectively.

The label data generation unit 22 generates integrated data Z by integrating the abstraction partial data W ₁ , ... and the abstraction partial data W c using the first abstraction anchor data W ^anc1 ₁ , ... received from terminal 10-1 and the first abstraction anchor data W ^anc1 _c received from terminal 10- _c . The label data generation unit 22 also generates integrated second anchor data Z anc2 by integrating the second abstraction anchor data W ^anc2 ₁ , ... and the second abstraction anchor data W ^anc2 _c using the first abstraction anchor data ^{W anc1 1} _... and the first abstraction anchor data W ^anc1 _c .

Next, the label data generation unit 22 generates an integrated analysis model h that outputs label data Y ₁ , ... and label data Y _c that indicate labels corresponding to the contents indicated by each of elements Z ₁ , ... and element Z _c included in the integrated data Z, based on the integrated data Z.

Next, the label data generation unit 22 generates second label data Y ^{anc2 1} , ... and second label data Y ^anc2 _c indicating labels corresponding to the contents indicated by each of the elements Z ^anc2 ₁ , ... and Z ^anc2 _c included in the integrated second anchor data Z ^anc2 from the integrated second anchor data Z ^anc2 and the integrated analytical model h. The second label data Y ^anc2 _{1 ,} ... and second label data Y ^anc2 _c all indicate labels corresponding to the contents indicated by the elements included in the integrated second anchor data Z ^anc2 .

Specifically, the label data generation unit 22 converts the first abstraction anchor data W ^anc1 ₁ received from terminal 10-1 into provisional integration data using a provisional integration function e _1. Similarly, the label data generation unit 22 converts the first abstraction anchor data W ^anc1 ₂ received from terminal 10-2 into provisional integration data using a provisional integration function e _2. Similarly, the label data generation unit 22 converts the first abstraction anchor data W ^anc1 _c received from terminal 10-c into provisional integration data using a provisional integration function e _c . The provisional integration functions e ₁ , ..., provisional integration function e _c are linear or nonlinear functions.

Next, the label data generation unit 22 calculates provisional integration functions _e1 , ..., ec that minimize the differences between these _c pieces of provisional integration data as integration functions g1, ..., gc. _{It is also preferable that the generation unit 220 calculates provisional integration functions e1 ,} ..., _ec that minimize the differences between these c pieces of provisional integration data as integration functions _g1 , ..., _gc .

The problem of calculating the integration functions g ₁ , ..., and g _c reduces to, for example, a minimization problem. In particular, when the provisional integration functions are linear functions, this problem reduces to a generalized total least squares (TLS) problem. The label data generation unit 22 generates an integrated analysis model h based on the integration functions g ₁ , ..., and g _c . The data Y, which is a combination of the label data Y ₁ , ..., and the label data Y _c , the integrated data Z, and the integrated analysis model h have a relationship expressed as Y≒h(Z). Then, the label data generation unit 22 generates second label data Y ^anc2 ₁ , ..., and second label data Y ^anc2 _c using the integrated second anchor data Z ^anc2 and the integrated analysis model h.

The data transmitting unit 23 transmits the second label data Y ^anc2 ₁ to the terminal 10-1. Similarly, the data transmitting unit 23 transmits the second label data Y ^anc2 ₂ , ... and the second label data Y ^anc2 _c to the terminals 10-2, ... and 10-c, respectively.

The data receiving unit 15-1 receives the second label data Y ^anc2 ₁ from the data processing device 20. Similarly, the data receiving units 15-2, ... and 15-c receive the second label data Y ^anc2 ₂ , ... and second label data Y ^anc2 _c from the data processing device 20, respectively.

The model generation unit 16-1 uses the second anchor data and the second label data Y ^anc2 ₁ to generate an integrated analytical model t ₁ that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data Y ^anc2 _1. Similarly, the model generation units 16-2, ... and 16-c generate integrated analytical models t ₂ , ... and integrated analytical model t _c , respectively. The integrated analytical model t ₁ is a model used for partial data newly received by terminal 10-1. Similarly, the integrated analytical models t ₂ , ... and integrated analytical model t _c are models used for partial data newly received by terminal 10-2, ... and partial data newly received by terminal 10-c, respectively.

Next, an example of processing executed by terminals 10-1, ..., terminal 10-c, and data processing device 20 will be described with reference to FIG. 6. FIG. 6 is a sequence diagram showing an example of processing executed by terminals and data processing device according to an embodiment of the present invention. In the description using FIG. 6, terminal 10-1 will be used as an example, but the same applies to terminals 10-2, ..., terminal 10-c.

In step S10, the abstraction function generation unit 11-1 generates an abstraction function.

In step S20, the abstraction execution unit 12-1 generates abstraction partial data, first abstraction anchor data, and second abstraction anchor data.

In step S30, the abstraction function discarding unit 13-1 discards the abstraction function.

In step S40, the data transmission unit 14-1 transmits the abstraction portion data, the first abstraction anchor data, the second abstraction anchor data, and the label data.

In step S50, the data receiving unit 21 receives the abstraction portion data, first abstraction anchor data, second abstraction anchor data, and label data.

In step S60, the label data generation unit 22 generates integrated data and generates an integrated analysis model.

In step S70, the label data generation unit 22 generates integrated second anchor data and generates second label data using the integrated second anchor data and the integrated analysis model.

In step S80, the data transmission unit 23 transmits the second label data.

In step S90, the data receiving unit 15-1 receives the second label data.

In step S100, the model generation unit 16-1 generates an integrated analysis model that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data.

The above describes the integrated analytical model generation system 1 according to an embodiment. The integrated analytical model generation system 1 includes terminals 10-1 and the like and a data processing device 20. The terminals 10-1 and the like generate an abstraction function, generate abstraction partial data, first abstraction anchor data, and second abstraction anchor data using the abstraction function, discard the abstraction function, and transmit these three pieces of data and label data to the data processing device 20. The data processing device 20 uses the first abstraction anchor data to integrate multiple abstraction partial data to generate integrated data Z, generates an integrated analytical model h that outputs label data indicating labels corresponding to the content indicated by the elements included in the integrated data Z based on the integrated data Z, integrates multiple second abstraction anchor data to generate integrated second anchor data, and generates second label data using the integrated second anchor data and the integrated analytical model h, which it then transmits to the terminals 10-1 and the like. The terminals 10-1 and the like use the second anchor data and the second label data to generate an integrated analytical model that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data.

As a result, the integrated analysis model generation system 1 discards the abstraction function after using it, making it impossible to match partial data with abstracted partial data.

Furthermore, terminal 10-1, etc. discards the abstraction function before the abstraction partial data, first abstraction anchor data, and second abstraction anchor data are sent to the data processing device. This allows the integrated analysis model generation system 1 to more reliably prevent the abstraction function from being leaked to anyone other than the person managing terminal 10-1, etc. after it has been used, resulting in partial data being compared with abstraction partial data.

In addition, terminals 10-1 and the like generate abstraction functions by at least one of a process of randomly extracting a portion of the partial data and a process of adding random elements to the abstraction function. This makes it impossible for the integrated analysis model generation system 1 to generate an abstraction function identical to the abstraction function in question, further reliably preventing partial data from being matched with abstracted partial data.

In addition, terminals 10-1 and the like randomly swap rows of abstracted partial data and label data equally. This makes it difficult for the integrated analysis model generation system 1 to match partial data with abstracted partial data, and more reliably prevents partial data from being matched with abstracted partial data.

Next, with reference to Figures 7 and 8, we will explain the results of applying the integrated analysis model generation system according to the embodiment to a specific example.

Figure 7 shows an example of the relationship between the number of institutions and the accuracy rate when abstracted partial data indicating corporate ratings is integrated and analyzed using an integrated analysis model generated by a terminal according to an embodiment of the present invention. The overall data in the case shown in Figure 7 consists of 3,932 pieces of data with feature values of working capital/total assets, retained earnings/total assets, earnings before interest and taxes/total assets, market capitalization/book value of all debt, sales/total assets, and industry, and each piece of data is assigned one of seven labels ranging from AAA to CCC.

In the case shown in Figure 7, the first anchor data is data representing uniform random numbers ranging from the minimum to maximum value of each feature. The second anchor data in this case is data generated by performing low-rank approximation using singular value decomposition on partial data held by each institution, adding noise, and linearly combining the data.

The abstraction function shown in Figure 7 is an abstraction function generated by applying principal component analysis to data with random perturbations added for each institution. Note that principal component analysis is a dimensionality reduction method that depends on the partial data, so the abstraction method will differ for each institution. Also, in the case shown in Figure 7, the rows of the abstracted partial data are randomly permuted.

In the case shown in Figure 7, the integration function is generated by solving a generalized total least squares problem. Also, in the case shown in Figure 7, the integrated data is analyzed using kernelized ridge regression with a Gaussian kernel. Furthermore, in the case shown in Figure 7, the integrated analysis model generated by the terminal is generated using kernelized ridge regression with a Gaussian kernel.

The horizontal axis of Figure 7 represents the number of institutions. The vertical axis of Figure 7 represents the accuracy rate of the analysis results. The solid line in Figure 7 represents the relationship between the number of institutions and the accuracy rate when abstracted partial data indicating corporate ratings is integrated and analyzed using an integrated analysis model generated by a terminal according to the embodiment. The dashed line in Figure 7 represents the relationship between the number of institutions and the accuracy rate when analysis is performed without performing processing to prevent matching of the partial data indicating corporate ratings with the abstracted partial data. The dashed-dotted line in Figure 7 represents the relationship between the number of institutions and the accuracy rate when partial data indicating corporate ratings is integrated and analyzed, rather than the abstracted partial data. The dashed-dotted line in Figure 7 represents the relationship between the number of institutions and the accuracy rate when one piece of partial data is analyzed only by the institution that holds that partial data.

Comparing the solid line shown in Figure 7 with the dashed line shown in Figure 7, it can be seen that the accuracy rate when the integrated analysis model generated by the terminal according to the embodiment is used to integrate and analyze the abstracted partial data indicating the ratings of companies is lower than the accuracy rate when the analysis is performed without performing the process to make it impossible to match the abstracted partial data, but still shows a relatively high accuracy rate.

Furthermore, when comparing the solid line shown in Figure 7 with the dashed-dotted line shown in Figure 7, it can be seen that the accuracy rate when the abstracted partial data indicating a company's rating is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment is higher than the accuracy rate when the partial data is integrated and analyzed.

Furthermore, when comparing the solid line shown in Figure 7 with the two-dot chain line shown in Figure 7, it can be seen that the accuracy rate when the abstracted partial data indicating a company's rating is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment is higher than the accuracy rate when one piece of partial data is analyzed only by the institution that holds that partial data.

Figure 8 is a diagram showing an example of the relationship between the number of units and the accuracy rate when abstracted partial data representing handwritten numbers is integrated and analyzed using an integrated analysis model generated by a terminal according to an embodiment of the present invention. The horizontal axis in Figure 8 represents the number of units.

In the case shown in Figure 8, the overall data represents handwritten numbers "0" to "9" displayed as a grayscale image with 28 vertical and horizontal pixels. Other conditions, procedures, etc. in the case shown in Figure 8 are the same as those in the case shown in Figure 7.

The vertical axis in Figure 8 represents the accuracy rate of the analysis results. The solid line in Figure 8 represents the relationship between the number of institutions and the accuracy rate when abstracted partial data representing handwritten digits is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment. The dashed line in Figure 8 represents the relationship between the number of institutions and the accuracy rate when analysis is performed without performing processing to prevent matching of partial data representing handwritten digits with abstracted partial data. The dotted-dash line in Figure 8 represents the relationship between the number of institutions and the accuracy rate when partial data representing handwritten digits is integrated and analyzed, rather than abstracted partial data. The two-dot-dash line in Figure 8 represents the relationship between the number of institutions and the accuracy rate when one piece of partial data is analyzed only by the institution that holds that partial data.

Comparing the solid line shown in Figure 8 with the dashed line shown in Figure 8, it can be seen that the accuracy rate when the abstract partial data representing handwritten numbers is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment is higher than the accuracy rate when the analysis is performed without performing processing to make matching with the abstract partial data impossible.

Furthermore, when comparing the solid line shown in Figure 8 with the dashed-dotted line shown in Figure 8, it can be seen that the accuracy rate when the abstract partial data representing handwritten numbers is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment is slightly lower than the accuracy rate when the partial data is integrated and analyzed.

Furthermore, when comparing the solid line shown in Figure 8 with the two-dot chain line shown in Figure 8, it can be seen that the accuracy rate when abstracted partial data representing handwritten numbers is integrated and analyzed using the integrated analysis model generated by the terminal according to the embodiment is higher than the accuracy rate when one piece of partial data is analyzed only by the institution that holds that partial data.

Note that at least some of the functions of terminals 10-1, ..., terminal 10-c, and data processing device 20 may be realized by hardware including circuitry executing a program. Examples of hardware include a CPU (Central Processing Unit), an LSI (Large Scale Integration), an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), and a GPU (Graphics Processing Unit). The above-mentioned programs are stored in a storage device that includes a storage medium. Examples of storage media include a HDD (Hard Disk Drive), flash memory, ROM (Read Only Memory), and DVD (Digital Versatile Disc). Furthermore, the above-mentioned programs may be differential programs that realize some of the functions of terminals 10-1, ..., terminal 10-c, and data processing device 20.

The above-described integrated analysis model generation system 1 can also be applied to data other than corporate financial data and data related to handwritten number images. For example, the above-described integrated analysis model generation system 1 can also be applied to data related to agriculture, data related to smart cities, and development data held by manufacturing companies.

Embodiments of the present invention have been described above with reference to the drawings. However, the integrated analysis model generation system 1 is not limited to the above-described embodiments, and various modifications, substitutions, combinations, and design changes may be made without departing from the spirit of the present invention.

1...Integrated analysis model generation system, 10-1, ..., 10-c...Terminal, 11-1, ..., 11-c...Abstraction function generation unit, 12-1, ..., 12-c...Abstraction execution unit, 13-1, ..., 13-c...Abstraction function discarding unit, 14-1, ..., 14-c...Data transmission unit, 15-1, ..., 15-c...Data reception unit, 16-1, ..., 16-c...Model generation unit, 20...Data processing device, 21...Data reception unit, 22...Label data generation unit, 23...Data transmission unit

Claims (9)

an abstraction function generation unit that generates an abstraction function used to perform abstraction processing on the partial data, the first anchor data, and the second anchor data;
an abstraction execution unit that applies the abstraction function to the partial data to generate abstracted partial data, applies the abstraction function to the first anchor data to generate first abstracted anchor data, and applies the abstraction function to the second anchor data to generate second abstracted anchor data;
an abstraction function discarding unit that discards the abstraction function after the abstraction partial data, the first abstraction anchor data, and the second abstraction anchor data are generated;
a data transmitting unit that transmits the abstraction partial data, the first abstraction anchor data, the second abstraction anchor data, and label data to a data processing device;
a data receiving unit that receives, from the data processing device, second label data indicating labels corresponding to contents indicated by elements included in integrated second anchor data generated by integrating a plurality of the second abstract anchor data using a plurality of the first abstract anchor data;
a model generation unit that generates an integrated analysis model that indicates a relationship between content indicated by the second anchor data and content indicated by the second label data, using the second anchor data and the second label data;
A terminal comprising: the abstraction function discarding unit discards the abstraction function before the abstraction portion data, the first abstraction anchor data, and the second abstraction anchor data are transmitted to the data processing device.
The terminal according to claim 1 . the abstraction function generation unit generates the abstraction function by at least one of a process of randomly extracting a part of the partial data and a process of adding a random element to the abstraction function.
The terminal according to claim 1 or claim 2. the abstraction execution unit equally and randomly permutes rows of the abstract partial data and the label data;
A terminal according to any one of claims 1 to 3. a data receiving unit that receives abstraction partial data, first abstraction anchor data, second abstraction anchor data, and label data for each terminal;
a label data generation unit that generates integrated data by integrating a plurality of the abstract partial data using the first abstraction anchor data received for each of the terminals, generates integrated second anchor data by integrating a plurality of the second abstraction anchor data using the first abstraction anchor data received for each of the terminals, generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data, and generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model;
a data transmission unit that transmits the second label data to the terminal;
A data processing device comprising: an abstraction function generation function that generates an abstraction function used to perform abstraction processing on the partial data, the first anchor data, and the second anchor data;
an abstraction execution function that applies the abstraction function to the partial data to generate abstract partial data, applies the abstraction function to the first anchor data to generate first abstract anchor data, and applies the abstraction function to the second anchor data to generate second abstract anchor data;
an abstraction function discarding function that discards the abstraction function after the abstraction partial data, the first abstraction anchor data, and the second abstraction anchor data are generated;
a data transmission function for transmitting the abstraction partial data, the first abstraction anchor data, the second abstraction anchor data, and label data to a data processing device;
a data receiving function that receives, from the data processing device, second label data that indicates labels corresponding to contents indicated by elements included in integrated second anchor data that is generated by integrating a plurality of the second abstract anchor data using a plurality of the first abstract anchor data;
a model generation function that uses the second anchor data and the second label data to generate an integrated analysis model that indicates the relationship between the content indicated by the second anchor data and the content indicated by the second label data;
A terminal control program that causes the terminal to execute the above. a data receiving function for receiving the abstraction part data, the first abstraction anchor data, the second abstraction anchor data, and the label data for each terminal;
a label data generation function that generates integrated data by integrating a plurality of the abstract partial data using the first abstraction anchor data received for each of the terminals, generates integrated second anchor data by integrating a plurality of the second abstraction anchor data using the first abstraction anchor data received for each of the terminals, generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data, and generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model;
a data transmission function for transmitting the second label data to the terminal;
A data processing program that causes a data processing device to execute the above. A terminal control method executed by a computer, comprising:
the computer generates, by an abstraction function generation unit or an abstraction function generation function, an abstraction function used to perform abstraction processing on the partial data, the first anchor data, and the second anchor data;
the computer, by an abstraction execution unit or an abstraction execution function, applies the abstraction function to the partial data to generate abstract partial data, applies the abstraction function to the first anchor data to generate first abstract anchor data, and applies the abstraction function to the second anchor data to generate second abstract anchor data;
the computer discards the abstraction function after the abstraction partial data, the first abstraction anchor data, and the second abstraction anchor data are generated by an abstraction function discarding unit or an abstraction function discarding function;
the computer transmits the abstraction partial data, the first abstraction anchor data, the second abstraction anchor data, and label data to a data processing device by a data transmission unit or a data transmission function;
the computer receives, via a data receiving unit or a data receiver, second label data from the data processing device, which indicates labels corresponding to content indicated by elements included in integrated second anchor data generated by integrating a plurality of the second abstract anchor data using a plurality of the first abstract anchor data;
the computer generates, by a model generation unit or a model generation function, an integrated analysis model that indicates a relationship between the content indicated by the second anchor data and the content indicated by the second label data, using the second anchor data and the second label data;
Terminal control methods. A computer-implemented data processing method, comprising:
the computer receives abstraction partial data, first abstraction anchor data, second abstraction anchor data, and label data for each terminal by a data receiving unit or a data receiving function;
the computer, by a label data generation unit or a label data generation function, generates integrated data by integrating a plurality of the abstract partial data using the first abstraction anchor data received for each of the terminals, generates integrated second anchor data by integrating a plurality of the second abstraction anchor data using the first abstraction anchor data received for each of the terminals, generates an integrated analysis model that outputs label data indicating labels corresponding to content indicated by elements included in the integrated data based on the integrated data, and generates second label data indicating labels corresponding to content indicated by elements included in the integrated second anchor data using the integrated second anchor data and the integrated analysis model;
the computer transmits the second label data to the terminal by a data transmission unit or a data transmission function;
Data processing methods.