WO2023175977A1

WO2023175977A1 - Learning device

Info

Publication number: WO2023175977A1
Application number: PCT/JP2022/012882
Authority: WO
Inventors: バトニヤマエンケタイワン; 勇寺西; 邦大伊東; 諒古川
Original assignee: 日本電気株式会社
Priority date: 2022-03-18
Filing date: 2022-03-18
Publication date: 2023-09-21

Abstract

A learning device 400 comprises a reception unit 421 that receives a learner from another learning device, and a generation unit 422 that generates an adapted learner so as to execute appropriate prediction for data in the learning device 400 by using the learner received by the reception unit 421 and the data in the learning device 400.

Description

learning device

The present invention relates to a learning device, a learning method, a recording medium, and an inference device.

In order to obtain better performance, the learning devices learned by each participant may be combined.

For example, in Non-Patent Document 1, there is a technique called Gradient Boosting Forest (GBF) in which participants create a decision tree in each step and combine the created decision trees to create a model with better performance. The technology is described.

Additionally, as a related document, for example, there is Patent Document 1. Patent Document 1 describes GBDT (Gradient Boosting Decision Tree) and the like.

JP 2021-140296 Publication

In the case of the technique described in Non-Patent Document 1, learning devices learned by each participant are combined. Therefore, the learning device that is finally created will be, for example, in accordance with the learning data that each participant has. On the other hand, for example, in cases where there is a bias in the distribution of learning data held by each participant, it may be desirable to prepare a learning device that is more suitable for a specific individual while improving performance. However, it is difficult to deal with the above-mentioned cases simply by combining as described in Non-Patent Document 1.

Therefore, an object of the present invention is to provide a learning device, a learning method, and a recording medium that can solve the above-mentioned problems.

In order to achieve this purpose, a learning device that is one form of the present disclosure includes:
a receiving unit that receives the learning device from another learning device;
a generation unit that uses the learning device received by the receiving unit and the data possessed by the device itself to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the device itself;
It has the following structure.

In addition, a learning method that is another form of the present disclosure includes:
The information processing device
Receive learning devices from other learning devices,
Using the received learning device and data possessed by the device itself, an appropriate learning device is generated so as to make predictions suitable for the data possessed by the device itself.

Furthermore, a recording medium according to another embodiment of the present disclosure includes:
In the information processing device,
Receive learning devices from other learning devices,
Using the received learning device and the data held by the own device, a computer records a program for realizing the process of generating an appropriate learning device so as to make predictions suitable for the data held by the own device. It is a readable recording medium.

Further, an inference device according to another embodiment of the present disclosure includes:
An inference device that infers a label based on an input of a feature amount, the inference device comprising:
a storage device that stores a learning device received from another learning device and a coupling coefficient calculated to perform a prediction suitable for the data of the own device using the data of the own device;
an inference unit that performs inference using the learning device and the coupling coefficient stored in the storage device according to the input of the feature amount;
It has the following structure.

According to each of the configurations described above, it is possible to prepare a learning device that is more suitable for a specific individual while improving its performance.

FIG. 1 is a diagram for explaining an overview of the present disclosure. FIG. 1 is a diagram illustrating a configuration example of a learning system according to a first embodiment of the present disclosure. FIG. 2 is a block diagram showing a configuration example of a learning device. It is a figure showing an example of learning data. It is a figure showing an example of validation data. It is a figure showing an example of an algorithm. It is a figure which shows another example of an algorithm. It is a flow chart which shows an example of operation of a learning device. FIG. 3 is a diagram illustrating an example hardware configuration of a learning device according to a second embodiment of the present disclosure. It is a block diagram showing an example of the configuration of a learning device. FIG. 2 is a block diagram showing a configuration example of an inference device.

[First embodiment]
A first embodiment of the present disclosure will be described with reference to FIGS. 1 to 8. FIG. 1 is a diagram for explaining an overview of the present disclosure. FIG. 2 is a diagram showing a configuration example of the learning system 100. FIG. 3 is a block diagram showing a configuration example of the learning device 300. FIG. 4 is a diagram showing an example of the learning data information 341. FIG. 5 is a diagram showing an example of the validation data information 342. 6 and 7 are diagrams showing an example of an algorithm performed by the learning system 100. FIG. 8 is a flowchart showing an example of the operation of the learning device 300.

In the first embodiment of the present disclosure, as shown in FIG. 1, based on learning devices received from other participants in the learning system 100 and learning devices learned by the own device, A learning system 100 including a learning device 300 that generates a combined learning device to make more appropriate predictions for data such as learning data and validation data will be described. As will be described later, when the learning device 300 described in this embodiment receives a decision tree, which is a learning device, from another participant in the learning system 100, it inputs the learning data it owns into the received decision tree. By doing so, new feature quantities are calculated. Then, the learning device 300 generates a decision tree, which is a new learning device of the own device, based on the learning data and the calculated feature amount.

Furthermore, the learning device 300 uses the generated decision trees, decision trees received from other participants, and pre-stored validation data to calculate a coupling coefficient corresponding to each decision tree. For example, the learning device 300 calculates a coupling coefficient for each decision tree based on the decision tree and validation data so that the prediction performance for its own validation data is optimal. Thereafter, the learning device 300 creates a new combined decision tree by combining the decision tree received from another learning device 200 or the generated decision tree with the past combined decision tree using the calculated combination coefficient. generate. That is, the learning device 300 generates a combined decision tree, which is an appropriate learning device, by performing combination using the combination coefficient.

For example, as shown in FIG. 1, in the learning system 100 described in this embodiment, the above-described process can be repeated. In other words, in the learning system 100, the other learning device 200 generates a decision tree, the learning device 300 generates a decision tree using the decision tree generated by the other learning device 200, the learning device 300 calculates a coupling coefficient, and each decision tree. A series of steps of combining can be repeated multiple times, for example, until a predetermined condition is satisfied.

Note that the learning data includes, for example, a plurality of feature amounts such as gender, age, height, weight, etc., and a label indicating whether the person is sick or not. The feature amount may also be called an explanatory variable, an attribute, or the like. Further, a label can also be called an objective variable. Specific examples of feature amounts and labels may be other than those exemplified above. Furthermore, validation data refers to data for verification that can be used when evaluating a learning device. Similar to the learning data, the validation data includes a plurality of feature amounts and labels. Furthermore, a decision tree is a model that is trained by performing the task of sorting input data multiple times using a binary tree based on conditional branching of its feature values until the label explanatory performance becomes sufficiently good. Decision trees include regression trees that handle regression tasks, classification trees that handle classification tasks, and the like.

Additionally, in this embodiment, a case will be described in which a decision tree is used as a learning device. However, the learning device to which the present invention is applied is not limited to decision trees. For example, each participant in the learning system 100 may generate a shallow neural network, a support vector machine, or the like as a learning device. Even if each participant in the learning system 100 generates a neural network, support vector machine, or the like as a learning device, the present invention can be applied without problems.

FIG. 2 shows an example of the overall configuration of the learning system 100. Referring to FIG. 2, the learning system 100 includes one or more other learning devices 200 and a learning device 300. As shown in FIG. 2, the other learning device 200 and the learning device 300 are connected via a network or the like so that they can communicate with each other.

The other learning device 200 is an information processing device that generates a decision tree, which is a learning device, by performing learning based on training data that the other learning device 200 has. Further, the other learning device 200 can transmit the generated decision tree to other learning devices 200 and learning devices 300.

Further, the other learning device 200 may be configured to combine decision trees that are learning devices received from other learning devices 200, learning devices 300, etc. using predetermined coefficients or the like. Here, the predetermined coefficient can be predetermined, for example, based on the number of data held by each participant such as the other learning devices 200 and the learning device 300 in the learning system 100. For example, the predetermined coefficient may be calculated by dividing the number of training data that the own device has by the sum of the number of training data that

other learning devices

200 and 300 in the learning system 100 have.

For example, as described above, the other learning device 200 generates a decision tree, which is a learning device, based on the training data of the own device, and transmits the generated decision tree to another information processing device in the learning system 100. or send it. Further, the other learning device 200 can receive decision trees from other information processing devices in the learning system 100, and combine the received decision trees using a predetermined coefficient or the like. For example, the other learning device 200 may implement each of the above processes using the method described in Non-Patent Document 1.

The learning device 300 is an information processing device that generates a joint decision tree, which is a joint learning device, to make more appropriate predictions for data such as learning data and validation data that the learning device 300 has. FIG. 3 shows a configuration example of the learning device 300. Referring to FIG. 3, the learning device 300 includes, as main components, an operation input section 310, a screen display section 320, a communication I/F section 330, a storage section 340, an arithmetic processing section 350, have.

Note that FIG. 3 illustrates a case where the function of the learning device 300 is realized using one information processing device. However, the learning device 300 may be realized using a plurality of information processing devices, such as being realized on a cloud, for example. Furthermore, the learning device 300 may not include some of the configurations exemplified above, such as not having the operation input unit 310 or the screen display unit 320, or may have a configuration other than those exemplified above.

The operation input unit 310 consists of an operation input device such as a keyboard and a mouse. The operation input section 310 detects the operation of the operator who operates the learning device 300 and outputs the detected operation to the arithmetic processing section 350 .

The screen display unit 320 is composed of a screen display device such as an LCD (Liquid Crystal Display). The screen display unit 320 can display various information stored in the storage unit 340 on the screen in response to instructions from the arithmetic processing unit 350.

The communication I/F section 330 consists of a data communication circuit and the like. The communication I/F section 330 performs data communication with an external device connected via a communication line.

The storage unit 340 is a storage device such as a hard disk or memory. The storage unit 340 stores processing information and programs 345 necessary for various processes in the arithmetic processing unit 350. The program 345 implements various processing units by being read and executed by the arithmetic processing unit 350. The program 345 is read in advance from an external device or a recording medium via a data input/output function such as the communication I/F section 330, and is stored in the storage section 340. The main information stored in the storage unit 340 includes, for example, learning data information 341, validation data information 342, learning device information 343, coefficient information 344, and the like.

The learning data information 341 includes learning data used when learning a decision tree, which is a learning device. For example, the learning data information 341 is acquired in advance using a method such as acquiring from an external device via the communication I/F unit 330 or inputting using the operation input unit 310, and is stored in the storage unit 340. has been done.

FIG. 4 shows an example of the learning data information 341. Referring to FIG. 4, in the learning data information 341, a plurality of feature amounts and labels are associated with each other. For example, in the example shown in FIG. 4, the feature amounts (x ₁ , x ₂ , . . . , x _d ) are associated with the label y ₁ . As shown in FIG. 4, the learning data information 341 may include a plurality of learning data.

The validation data information 342 includes validation data that is data used when verifying the performance of a decision tree. For example, the validation data information 342 is obtained in advance using a method such as obtaining it from an external device via the communication I/F section 330 or inputting it using the operation input section 310, and is stored in the storage section 340. has been done.

FIG. 5 shows an example of the validation data information 342. Referring to FIG. 5, in the validation data information 342, like the learning data information 341, a plurality of feature amounts and labels are associated. For example, in the example shown in FIG. 5, the feature amounts (x ₁₁ , x ₁₂ , . . . , x _1d ) are associated with the label y ₁₀ . As shown in FIG. 5, the validation data information 342 may include a plurality of pieces of validation data.

The learning device information 343 includes information indicating a decision tree received from another learning device 200, a combined decision tree combined by a combining unit 355, which will be described later, and the like. For example, in the learning device information 343, the decision tree is associated with identification information indicating the source of the decision tree. The learning device information 343 may include a decision tree or a combined decision tree for each step. For example, the learning device information 343 is updated in response to a receiving unit 351 (described later) receiving a decision tree from another learning device 200, a combining unit 355 generating a combined decision tree, and the like.

The coefficient information 344 includes coupling coefficients corresponding to each decision tree, such as a decision tree received from another learning device 200 or a decision tree generated by a learning unit 353, which will be described later. For example, in the coefficient information 344, decision tree identification information and coupling coefficients are associated. The coefficient information 344 may include a coupling coefficient for each step and each decision tree. For example, the coefficient information 344 is updated in response to calculation of a coupling coefficient by a coefficient calculation unit 354, which will be described later.

The arithmetic processing unit 350 includes an arithmetic unit such as a CPU (Central Processing Unit) and its peripheral circuits. The arithmetic processing unit 350 reads the program 345 from the storage unit 340 and executes it, thereby causing the hardware and the program 345 to work together to implement various processing units. The main processing units realized by the arithmetic processing unit 350 include, for example, a receiving unit 351, a feature value addition calculation unit 352, a learning unit 353, a coefficient calculation unit 354, a combining unit 355, an inference unit 356, an output unit 357, etc. be.

The receiving unit 351 receives a decision tree, which is a learning device, from the other learning device 200. For example, the receiving unit 351 can receive a learning device from each other learning device 200 included in the learning system 100 in each step. Further, the receiving unit 351 stores the received decision tree in the storage unit 340 as learning device information 343.

Note that the receiving unit 351 may receive, for example, information indicating the difference between the decision tree in the previous step and the like from the other learning device 200. In this case, the receiving unit 351 may be configured to update the corresponding decision tree based on the received information indicating the difference.

The feature value addition calculation unit 352 calculates additional learning data based on the decision tree received by the receiving unit 351 and the learning data included in the learning data information 341. For example, the feature quantity addition calculation unit 352 obtains an output from the learning device by inputting each learning data included in the learning data information 341 to the decision tree received by the receiving unit 351. The feature quantity addition calculation unit 352 can acquire the above output as an additional feature quantity.

For example, it is assumed that the learning data information 341 includes learning data (x _i , y _i ) including feature quantity x _i and label y _i (i may be arbitrary). Further, assume that decision trees f ₁ (), f ₂ (), . . . are received from other learning devices 200. In this case, the feature amount additional calculation unit 352 calculates additional feature amounts f ₁ (x _i ), f ₂ (x _i ), . . . by inputting the feature amount xi into each decision tree. As a result, the learning data to be learned by the learning unit 353, which will be described later, becomes (x _i , f ₁ (x _i ), f ₂ (x _i ), . . . , y _i ).

For example, the feature value addition calculation unit 352 can perform the above-described processing for each decision tree and each learning data. For example, the feature amount addition calculation unit 352 may perform the above process for each decision tree extracted by an arbitrary method.

The learning unit 353 generates a decision tree, which is a learning device, by performing learning based on the feature quantities calculated by the feature quantity addition calculation unit 352 and the learning data indicated by the learning data information 341. Further, the learning unit 353 stores the generated decision tree in the storage unit 340 as learning device information 343.

For example, as described above, the additional feature amount is calculated by the feature amount addition calculation unit 352. Therefore, the learning unit 353 performs machine learning using learning data including additional feature quantities, such as (x _i , f ₁ (x _i ), f ₂ (x _i ), ..., y _i ). This will generate a decision tree.

Note that the learning unit 353 may perform machine learning by directly adding the additional feature calculated by the feature addition calculation unit 352 to the learning data as described above; The configuration may be such that machine learning is performed by adding the results of linearly combining the additional feature quantities calculated by the above to the learning data. The learning unit 353 may perform machine learning by adding the additional feature calculated by the feature addition calculation unit 352 and the result of linearly combining the additional feature to the learning data.

The coefficient calculating unit 354 calculates a coupling coefficient for each decision tree using the validation data indicated by the validation data information 342. For example, the coefficient calculating unit 354 calculates a coupling coefficient so that the prediction performance for the validation data indicated by the validation data information 342 is optimal. The coefficient calculating unit 354 can calculate a coupling coefficient for each decision tree received by the receiving unit 351 or for each decision tree generated by the learning unit 353. Further, the coefficient calculation unit 354 stores the calculated coupling coefficient in the storage unit 340 as coefficient information 344.

For example, it is assumed that the validation data information 342 includes validation data (x _1i ,y _1i ) including a feature amount x _{1i and} a label y _1i . Further, assume that the other learning device 200 or the learning unit 353 receives or generates decision trees f ₁₁ (), f ₁₂ (), . . . . In this case, first, the coefficient calculation unit 354 obtains an output by inputting validation data to each decision tree. For example, the coefficient calculation unit 354 obtains the output u _i by inputting the validation data (x _1i , y _1i ) into the decision tree f ₁₁ ( ). Further, the coefficient calculation unit 354 obtains the output v _i by inputting the validation data (x _1i , y _1i ) to the decision tree f ₁₂ (). Then, the coefficient calculating unit 354 calculates a coupling coefficient for each decision tree by using (u _i , v _i , y _1i ). For example, the coefficient calculation unit 354 may calculate the coupling coefficient by performing linear regression. For example, the coefficient calculating unit 354 may determine the coupling coefficient corresponding to the decision tree f11( ) by performing linear regression using the validation data (x _1i , y _1i ) and the output u _i .

Specifically, for example, the coefficient calculating unit 354 calculates the coefficient a corresponding to each decision tree by performing linear regression on Equation 1 using the validation data (x _1i , y _1i ) and u _i , v _i _i can be determined.

Note that the coefficient calculation unit 354 may calculate the coupling coefficient using the entire validation data, or may calculate the coupling coefficient using a part of the validation data. For example, by referring to the model information about the decision tree generated by the learning unit 353, such as the model structure and branching conditions, it is possible to specify the leaf node where each validation data falls. Therefore, the coefficient calculation unit 354 may calculate a coupling coefficient for each leaf node, for example, by performing linear regression using validation data for each leaf node. Even when calculating a coupling coefficient for each leaf node, by coupling each leaf node, a coupling decision tree can be generated in the same way as in the case described above. Note that when the coefficient calculating unit 354 calculates a coupling coefficient using the entire validation data, it can also be said that the coupling coefficient for the entire decision tree is calculated.

Furthermore, the coefficient calculation unit 354 may calculate the coupling coefficient using a method other than the one exemplified above. For example, the coefficient calculating unit 354 may calculate the coupling coefficient using the learning data indicated by the learning data information 341 instead of the validation data. However, from the viewpoint of suppressing excessive bias, it is preferable to calculate the coupling coefficient using validation data rather than using learning data. The coefficient calculation unit 354 may calculate the coupling coefficient using any other method.

The combination unit 355 uses the combination coefficients calculated by the coefficient calculation unit 354 to combine the decision tree received by the reception unit 351 or the decision tree generated by the learning unit 353 with the previous combination decision included in the learning device information 343. By combining the trees, a new combined decision tree is generated. Furthermore, the combining unit 355 stores the newly generated combined decision tree in the storage unit 340 as learning device information 343.

For example, the coupling unit 355 performs coupling using coupling coefficients by solving Equation 2 below.
Here, f ^(t-1) indicates the joint decision tree one step before, and f ^(t) indicates the newly generated joint decision tree. Further, a _k indicates a coupling coefficient, and f _k indicates a decision tree received by the receiving section 351 or a decision tree generated by the learning section 353. For example, k has a value corresponding to the number of other learning devices 200 and learning devices 300 included in the learning system 100.

The inference unit 356 performs inference using a joint decision tree. For example, the inference unit 356 can perform inference using the latest joint decision tree.

Note that, as shown in Equation 3, the latest joint decision tree f _final ( ) includes decision trees, joint coefficients, etc. that were received or generated in past steps. Therefore, the inference unit 356 can also perform inference using decision trees, coupling coefficients, etc. generated in past steps.
Note that, as shown in Equation 4, ft() is a term added to the model at the t-th step, and is a linear sum of the decision trees f _k () created by each participant.
Note that Equation 5 shows the coefficients of the decision tree created by the k-th participant at the t-th step.

The output unit 357 outputs a decision tree or a combined decision tree, or outputs an inference result by the inference unit 356.

For example, the output unit 357 transmits the decision tree generated by the learning unit 353 or the combined decision tree generated by the combining unit 355 to an external device such as another learning device 200 via the communication I/F unit 330. be able to. The output unit 357 may output at any timing, such as after processing by the learning unit 353 or the combining unit 355.

Further, the output unit 357 can display the result of the inference by the inference unit 356 on the screen display unit 320 or transmit it to an external device via the communication I/F unit 330.

The above is an example of the configuration of the learning device 300. Note that FIG. 6 shows an example of an algorithm of the learning system 100 when handling a regression task, which is described in Non-Patent Document 1. In the example shown in FIG. 6, a case is illustrated in which participants such as the other learning device 200 and the learning device 300 are included in the group K learning system 100. As illustrated in FIG. 6, in the learning system 100, decision trees f _k ( ) generated by each participant are combined using a connection coefficient a. Here, as described above, the coupling coefficient a is calculated so that the prediction performance for the validation data is optimized. Therefore, by performing a combination using the combination coefficient a, it is possible to improve performance and generate a combination decision tree in a form more suitable for the learning device 300 having validation data.

Note that the example shown in FIG. 6 illustrates an example of the algorithm of the learning system 100 when dealing with a regression task, but even when dealing with a classification task instead of a regression task, the same algorithm as in the case of the regression task can be used. , performs the coupling using the coupling coefficient a. Therefore, even when handling a classification task, by performing a combination using the combination coefficient a, as in the case of a regression task, it is possible to improve the performance and create a form more suitable for the learning device 300 that has validation data. can generate a joint decision tree. For example, Non-Patent Document 1 describes an example of an algorithm for handling a classification task as shown in FIG. Referring to FIG. 7, it can be seen that even when dealing with a classification task, the combination is performed using the combination coefficient a, as in the case of the regression task.

Next, an example of the operation of the learning device 300 will be described with reference to FIG. 8. FIG. 8 is a flowchart showing an example of the operation of the learning device 300. Referring to FIG. 8, the receiving unit 351 receives a decision tree, which is a learning device, from the other learning device 200 (step S101).

The feature value addition calculation unit 352 calculates additional learning data based on the decision tree received by the receiving unit 351 and the learning data included in the learning data information 341 (step S102). For example, the feature quantity addition calculation unit 352 obtains an output from the learning device by inputting each learning data included in the learning data information 341 to the decision tree received by the receiving unit 351. The feature quantity addition calculation unit 352 can acquire the above output as an additional feature quantity.

The learning unit 353 generates a decision tree, which is a learning device, by performing learning based on the feature quantity calculated by the feature quantity addition calculation unit 352 and the learning data indicated by the learning data information 341 (step S103). That is, the learning unit 353 generates a decision tree by performing learning by adding the feature amount calculated by the feature amount addition calculation unit 352 to the learning data indicated by the learning data information 341.

The coefficient calculating unit 354 calculates a coupling coefficient for each decision tree using the validation data indicated by the validation data information 342 (step S104). For example, the coefficient calculating unit 354 calculates a coupling coefficient so that the prediction performance for the validation data indicated by the validation data information 342 is optimal. The coefficient calculating unit 354 can calculate a coupling coefficient for each decision tree received by the receiving unit 351 or for each decision tree generated by the learning unit 353.

The combination unit 355 uses the combination coefficients calculated by the coefficient calculation unit 354 to combine the decision tree received by the reception unit 351 or the decision tree generated by the learning unit 353 with the previous combination decision included in the learning device information 343. By combining the trees, a new combined decision tree is generated (step S105).

The above is an example of the operation of the learning device 300.

In this way, the learning device 300 includes the feature quantity addition calculation section 352 and the learning section 353. According to such a configuration, the learning unit 353 can generate a decision tree, which is a learning device, by performing learning using the learning data to which the feature quantity calculated by the feature quantity addition calculation unit 352 is added. As a result, a decision tree can be generated that also incorporates the results of learning by other learning devices 200. Thereby, it is possible to generate a decision tree, which is a learning device more suitable for the data held by the device itself, while improving performance.

Further, the learning device 300 includes a coefficient calculating section 354 and a combining section 355. According to such a configuration, the combination unit 355 can calculate each decision tree using the combination coefficients calculated by the coefficient calculation unit 354. As described above, the coupling coefficient is calculated so that the prediction performance for the validation data is optimal. Therefore, by combining each decision tree using the above-mentioned connection coefficient, it is possible to improve performance and generate a decision tree that is a learning device more suitable for the learning device 300 having validation data.

Note that in this embodiment, the case where the learning device 300 includes both the feature value addition calculation unit 352 and the coefficient calculation unit 354 is illustrated. However, the learning device 300 may include only one of the feature amount addition calculation section 352 and the coefficient calculation section 354.

For example, if the learning device 300 does not include the feature amount addition calculation unit 352, the learning unit 353 performs learning based on learning data included in the learning data information 341 to generate a decision tree. Even in such a case, as mentioned above, in order to calculate the coupling coefficient so that the prediction performance for the validation data is optimal, by combining each decision tree using the coupling coefficient, it is possible to improve the performance while , a decision tree that is a learning device more suitable for the learning device 300 having validation data can be generated.

For example, if the learning device 300 does not have the coefficient calculating unit 354, the combining unit 355 combines a decision tree generated by incorporating the results of learning by other learning devices 200, and a combined decision tree one step before. Combine with. As a result, it is possible to generate a decision tree, which is a learning device more suitable for the data held by the device itself, while improving performance. Note that if the learning device 300 does not include the coefficient calculating unit 354, the combining unit 355 may combine the decision tree generated without using a combining coefficient and the combined decision tree one step before.

Furthermore, as described above, in this embodiment, a case has been described in which a decision tree is used as a learning device. However, the learning device to which the present invention is applied is not limited to decision trees. For example, the learning device may be a shallow neural network, a support vector machine, or the like. Of course, the finally generated learning device may also correspond to each of the learning devices described above. For example, when a decision tree is used as a learning device, the finally generated combined decision tree becomes a GBDT (Gradient Boosting Decision Tree) model. Furthermore, when a neural network is used as a learning device, the finally generated model will be a gradient boosting neural network.

[Second embodiment]
Next, a second embodiment of the present disclosure will be described with reference to FIGS. 9 to 11. FIG. 9 is a diagram showing an example of the hardware configuration of the learning device 400. FIG. 10 is a block diagram showing a configuration example of the learning device 400. FIG. 11 is a block diagram showing a configuration example of the inference device 500.

In the second embodiment of the present disclosure, a configuration example of a learning device 400, which is an information processing device that receives learning devices from other devices and combines the learning devices so as to perform optimal prediction for its own data, will be described. explain. FIG. 9 shows an example of the hardware configuration of the learning device 400. Referring to FIG. 9, the learning device 400 has the following hardware configuration, as an example.
・CPU (Central Processing Unit) 401 (arithmetic unit)
・ROM (Read Only Memory) 402 (storage device)
・RAM (Random Access Memory) 403 (storage device)
- Program group 404 loaded into RAM 403
- Storage device 405 that stores program group 404
- A drive device 406 that reads and writes from a recording medium 410 external to the information processing device
- A communication interface 407 that connects to a communication network 411 outside the information processing device
・I/O interface 408 that inputs and outputs data
・Bus 409 connecting each component

Further, the learning device 400 can realize the functions of the receiving section 421 and the generating section 422 shown in FIG. 10 by the CPU 401 acquiring the program group 404 and executing the program group 404. Note that the program group 404 is stored in the storage device 405 or ROM 402 in advance, for example, and is loaded into the RAM 403 or the like by the CPU 401 and executed as necessary. Further, the program group 404 may be supplied to the CPU 401 via the communication network 411, or may be stored in the recording medium 410 in advance, and the drive device 406 may read the program and supply it to the CPU 401.

Note that FIG. 9 shows an example of the hardware configuration of the learning device 400. The hardware configuration of learning device 400 is not limited to the above case. For example, the learning device 400 may be configured from part of the configuration described above, such as not having the drive device 406.

The receiving unit 421 receives learning devices from other learning devices.

The generation unit 422 uses the learning device received by the receiving unit 421 and the data possessed by the own device to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the own device. For example, the generation unit 422 combines the learning device received by the reception unit 421 using a coupling coefficient calculated using data possessed by the own device, so as to make an appropriate prediction suitable for the data possessed by the own device. Generate a learning machine. Alternatively, the generation unit 422 generates an appropriate learning device by generating a learning device using learning data to which additional feature amounts calculated using the learning device received by the receiving unit are added. For example, the generation unit 422 can generate the appropriate learning device by performing any of the methods exemplified above or a combination thereof.

In this way, the learning device 400 includes a receiving section 421 and a generating section 422. According to such a configuration, the generation unit 422 uses the learning device received by the receiving unit 421 and the data possessed by the own device to generate an appropriate learning device so as to perform a prediction suitable for the data possessed by the own device. can be generated. As a result, it is possible to generate a learning device that is more suitable for the data held by the device itself while improving performance.

Note that the learning device 400 described above can be realized by incorporating a predetermined program into an information processing device such as the learning device 400. Specifically, a program according to another embodiment of the present invention causes an information processing device such as the learning device 400 to receive a learning device from another learning device, and uses the received learning device and data possessed by the own device. This is a program for realizing processing that generates an appropriate learning machine so as to make predictions suitable for the data possessed by the own device.

Further, in a learning method executed by an information processing device such as the learning device 400 described above, the information processing device such as the learning device 400 receives a learning device from another learning device, and the received learning device and the own device are connected to each other. In this method, an appropriate learning machine is generated using the data that the device has, so as to make a prediction suitable for the data that the device has.

Even with the invention of a program, a computer-readable recording medium on which the program is recorded, or a learning method having the above-described configuration, in order to achieve the same operation and effect as the learning device 400 described above, The objectives of the present disclosure described above can be achieved.

Further, the purpose of the present disclosure can be achieved in the same way even with the inference device 500 or the like that performs inference using the appropriate learning device generated by the above-mentioned learning device 400 or the like. For example, as shown in FIG. 11, the inference device 500 uses a learning device received from another learning device and the data it owns to perform a combination calculated to make a prediction suitable for the data it owns. It has a storage device 521 that stores coefficients, and an inference unit 522 that performs inference using the learning device and the coupling coefficients stored in the storage device according to the input of the feature amount. Note that the hardware configuration of the inference device 500 may be the same as that of the learning device 400.

<Additional notes>
Part or all of the above embodiments may also be described as in the following additional notes. Hereinafter, the outline of the learning device etc. in the present invention will be explained. However, the present invention is not limited to the following configuration.

(Additional note 1)
a receiving unit that receives the learning device from another learning device;
a generation unit that uses the learning device received by the receiving unit and data possessed by the device itself to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the device;
A learning device.
(Additional note 2)
The learning device according to Supplementary Note 1,
It has a calculation unit that calculates a coupling coefficient to make a prediction suitable for the data using data possessed by the own device,
The generation unit generates the appropriate learning machine so as to perform prediction suitable for data possessed by the own device by combining the learning machines received by the reception unit using the coupling coefficient calculated by the calculation unit. A learning device.
(Additional note 3)
The learning device according to appendix 2,
A learning device, wherein the calculation unit calculates a coupling coefficient so as to perform a prediction suitable for the data by performing linear regression using a result of inputting data possessed by the learning device to the learning device.
(Additional note 4)
The learning device according to appendix 2 or appendix 3,
The learning device wherein the calculation unit calculates the coupling coefficient using validation data that is data for verification.
(Appendix 5)
The learning device according to any one of Supplementary Notes 2 to 4,
The calculation unit specifies data that falls on each leaf node in a decision tree that is a learning device, and calculates the coupling coefficient using data for each leaf node.
(Appendix 6)
The learning device according to any one of Supplementary Notes 1 to 5,
a feature amount calculation unit that calculates an additional feature amount using the learning device and the learning data received by the receiving unit;
a learning unit that generates a learning device by adding the feature calculated by the feature calculation unit to the learning data and learning;
has
The generation unit generates the appropriate learning device using the learning device generated by the learning unit. The learning device.
(Appendix 7)
The learning device according to appendix 6,
a calculation unit that calculates a coupling coefficient corresponding to the learning device received by the reception unit using data possessed by the own device;
The generation unit generates the appropriate learning device by combining the learning device received by the receiving unit and the learning device generated by the learning unit using the coupling coefficient calculated by the calculation unit.Learning device .
(Appendix 8)
The information processing device
Receive learning devices from other learning devices,
A learning method that uses a received learning device and data possessed by the device itself to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the device itself.
(Appendix 9)
In the information processing device,
Receive learning devices from other learning devices,
Using the received learning device and the data held by the own device, a computer records a program for realizing the process of generating an appropriate learning device so as to make predictions suitable for the data held by the own device. A readable recording medium.
(Appendix 10)
An inference device that infers a label based on an input of a feature amount, the inference device comprising:
a storage device that stores a learning device received from another learning device and a coupling coefficient calculated to perform a prediction suitable for the data of the own device using the data of the own device;
an inference unit that performs inference using a learning device and a coupling coefficient stored in the storage device according to input of a feature amount;
has a reasoning device.

Although the present invention has been described above with reference to each of the embodiments described above, the present invention is not limited to the embodiments described above. The configuration and details of the present invention can be modified in various ways within the scope of the present invention by those skilled in the art.

100 Learning system 200 Other learning devices 300 Learning device 310 Operation input section 320 Screen display section 330 Communication I/F section 340 Storage section 341 Learning data information 342 Validation data information 343 Learning device information 344 Coefficient information 345 Program 350 Arithmetic processing section 351 Reception Unit 352 Feature value addition calculation unit 353 Learning unit 354 Coefficient calculation unit 355 Combining unit 356 Inference unit 357 Output unit 400 Learning device 401 CPU
402 ROM
403 RAM
404 Program group 405 Storage device 406 Drive device 407 Communication interface 408 Input/output interface 409 Bus 410 Recording medium 411 Communication network 421 Receiving section 422 Generation section 500 Inference device 521 Storage device 522 Inference section

Claims

a receiving unit that receives the learning device from another learning device;
a generation unit that uses the learning device received by the receiving unit and data possessed by the device itself to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the device;
A learning device.
The learning device according to claim 1,
It has a calculation unit that calculates a coupling coefficient to make a prediction suitable for the data using data possessed by the own device,
The generation unit generates the appropriate learning device so as to perform prediction suitable for data possessed by the own device by combining the learning devices received by the reception unit using the coupling coefficient calculated by the calculation unit. A learning device.
The learning device according to claim 2,
A learning device, wherein the calculation unit calculates a coupling coefficient so as to perform a prediction suitable for the data by performing linear regression using a result of inputting data possessed by the learning device to the learning device.
The learning device according to claim 2 or 3,
The learning device wherein the calculation unit calculates the coupling coefficient using validation data that is data for verification.
The learning device according to any one of claims 2 to 4,
The calculation unit specifies data that falls on each leaf node in a decision tree that is a learning device, and calculates the coupling coefficient using data for each leaf node.
The learning device according to any one of claims 1 to 5,
a feature amount calculation unit that calculates an additional feature amount using the learning device and the learning data received by the receiving unit;
a learning unit that generates a learning device by adding the feature calculated by the feature calculation unit to the learning data and learning;
has
The generation unit generates the appropriate learning device using the learning device generated by the learning unit. The learning device.
The learning device according to claim 6,
a calculation unit that calculates a coupling coefficient corresponding to the learning device received by the reception unit using data possessed by the own device;
The generation unit generates the appropriate learning device by combining the learning device received by the receiving unit and the learning device generated by the learning unit using the coupling coefficient calculated by the calculation unit.Learning device .
The information processing device
Receive learning devices from other learning devices,
A learning method that uses a received learning device and data possessed by the device itself to generate an appropriate learning device so as to make a prediction suitable for the data possessed by the device itself.
In the information processing device,
Receive learning devices from other learning devices,
Using the received learning device and the data held by the own device, a computer records a program for realizing the process of generating an appropriate learning device so as to make predictions suitable for the data held by the own device. A readable recording medium.
An inference device that infers a label based on an input of a feature amount, the inference device comprising:
a storage device that stores a learning device received from another learning device and a coupling coefficient calculated to perform a prediction suitable for the data of the own device using the data of the own device;
an inference unit that performs inference using a learning device and a coupling coefficient stored in the storage device according to input of a feature amount;
has a reasoning device.