JP7709662B2 - Specification-compliant data estimation device, machine learning method, specification-compliant data estimation method, and program - Google Patents

Description

The present disclosure relates to a technology that, when there are certain specifications that an input/output data pair for a given machine learning model must satisfy, enables the input/output data to satisfy the given specifications while reducing prediction errors.

Let X be the input domain and Y be the output domain. Machine learning is an algorithm that, when a sequence of training examples S = [( _x1 , _y1 ), ( _x2 , _y2 ), ..., ( _xn , _yn )], _which is a sequence of pairs of element xi of X and element yi of _Y , is given, selects from a set of hypotheses H that can predict _yi corresponding to _xi as accurately as possible among hypotheses h: X → Y, which is a mapping from the input domain X to the output domain Y. In other words, machine learning selects a predetermined hypothesis from the given set of hypotheses H, where there are many training examples ( _xi , _yi ) ∈ S that satisfy h( _xi ) = _yi .

For example, the input domain X may be a set of N-dimensional real vectors, a set of images, a set of news articles, etc. Also, the output domain Y for each input domain X may be a set of real numbers, a set of names of objects in images, a set of news categories, etc. In particular, when the output domain Y is a discrete finite set of natural numbers Y={1,2,…,K}, the hypothesis h∈H is called a multi-class classifier. Also, the problem of estimating the corresponding output domain y∈Y for a given input domain x∈X using a hypothesis is called a multi-class classification problem.

Methods for solving multi-class classification problems using such machine learning techniques are widely used. For example, a text classification problem is a problem of estimating the genre of a news document given as input, and this problem can be regarded as a multi-class classification problem and solved using machine learning techniques.

Machine learning technology is used in a wide range of application fields, but in some cases, the value that the hypothesis output domain y = h(x) should satisfy for a given input domain x is predetermined by specifications. For example, a mechanism for recommending products that are expected to interest users of an online shopping service can be formulated as a multi-class classification problem, with the set of users being X and the set of products to be recommended being Y. When solving such a problem using machine learning technology, the hypothesis h:X estimated from the training (learning) examples is
→Y may recommend any product to any user.

However, if the terms of service prohibit the recommendation of certain products to minors, the product y=h(x) output by the hypothesis to a minor user x may be
Recommending someone may be in violation of the terms of use.

Usually, it is difficult to guarantee that a hypothesis h∈H estimated from a set of hypotheses H satisfies a specification (rule). Therefore, in the past, when such a specification exists, it was necessary to use a specific model such as Markov Logic Networks as a machine learning model (set of hypotheses) capable of taking the specification into account (see Non-Patent Document 1).

Matthew Rechardson and Pedro Domingos, "Markov Logic Networks," Machine Learning Vol.62, pp.107-136, 2006.

However, there are various specifications, not limited to the above example of minors, and the machine learning models that perform well according to the various specifications are different. Therefore, in order to guarantee that the specifications are met, a specific machine learning model must be selected according to the content of the specifications, but in reality, it is difficult to select an appropriate specific model from among various machine learning models.

The present invention has been made in consideration of the above points, and has an object to perform appropriate machine learning in accordance with various specifications.

In order to solve the above problem, the invention of claim 1 is a specification-compliant data estimation device that trains a machine learning model in a learning phase in accordance with specified specifications, the specification-compliant data estimation device having an acquisition unit that acquires multiple training example data, which are pairs of input data and output data, and a specification function that outputs whether each of the multiple training examples satisfies the specified specifications, a specification verification unit that uses the specification function to verify specific training example data among the multiple training example data that satisfies the specified specifications, and a machine learning unit that trains the machine learning model by machine learning using the specific training example data.

As described above, the present invention has the effect of enabling appropriate machine learning to be performed in accordance with various specifications.

1 is an overall configuration diagram of a communication system according to an embodiment of the present invention; 1 is a diagram illustrating an electrical hardware configuration of a specification correspondence data estimation device according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an electrical hardware configuration of a communication terminal according to the present embodiment. FIG. 2 is a functional configuration diagram of the specification correspondence data estimation device in a learning phase. FIG. 2 is a functional configuration diagram of a specification correspondence data estimation device in an estimation phase. 10 is a flowchart showing the processing or operation executed by the specification-compliant data estimation device in a learning phase. 10 is a flowchart showing the process or operation executed by the specification correspondence data estimation device in the estimation phase.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[System configuration of the embodiment]
First, the overall configuration of the communication system according to the present embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing the overall configuration of the communication system according to the present embodiment.

1, a communication system 1 according to the present embodiment is constructed by a specification-compliant data estimation device 3 and a communication terminal 5. The communication terminal 5 is managed and used by a user. The user refers to the output result of the specification-compliant data estimation device and determines the subsequent action to be taken.

Furthermore, the specification correspondence data estimation device 3 and the communication terminal 5 can communicate with each other via a communication network 100 such as the Internet. The connection form of the communication network 100 may be either wireless or wired.

The specification-compliant data estimation device 3 is composed of one or more computers. When the specification-compliant data estimation device 3 is composed of multiple computers, it may be referred to as a "specification-compliant data estimation device" or a "specification-compliant data estimation system."

The specification-based data estimation device 3 performs appropriate machine learning and estimation in accordance with various specifications (rules), such as the prohibition of recommending a specific product group to minors.

The communication terminal 5 is a computer, and a notebook computer is shown as an example in Fig. 1. In Fig. 1, a user operates the communication terminal 5. Note that the specification correspondence data estimation device 3 may perform processing alone without using the communication terminal 5.

[Hardware configuration]
<Hardware configuration of the specification-compliant data estimation device>
Next, the electrical hardware configuration of the specification correspondence data estimation device 3 will be described with reference to Fig. 2. Fig. 2 is a diagram showing the electrical hardware configuration of the specification correspondence data estimation device.

The specification correspondence data estimation device 3 is a computer, and as shown in FIG. 2, includes a CPU (Central Processing Unit) 301, a ROM (Read Only Memory) 302, a RAM (Random Access Memory) 303, and a RAM (RAM).
Access Memory 303, SSD (Solid State Drive) 304, external device connection I/F (Interface) 305, network I/F 306, media I/F 309, and bus line 31
It has 0.

Of these, the CPU 301 controls the overall operation of the specification correspondence data estimation device 3. The ROM 302 stores programs used to drive the CPU 301, such as an IPL (Initial Program Loader). The RAM 303 is used as a work area for the CPU 301.

The SSD 304 reads and writes various data under the control of the CPU 301. Note that instead of the SSD 304, a hard disk drive (HDD) may be used.

The external device connection I/F 305 is an interface for connecting various external devices, such as a display, a speaker, a keyboard, a mouse, a USB (Universal Serial Bus) memory, and a printer.

The network I/F 306 is an interface for performing data communication via the communication network 100 .

The media I/F 309 controls reading and writing (storing) of data to a recording medium 309m such as a flash memory. The recording medium 309m includes a DVD (Digital Versatile Disc) and a Blu-ray Disc (registered trademark).

The bus line 310 is an address bus, a data bus, or the like for electrically connecting the various components such as the CPU 301 shown in FIG.

<Hardware configuration of communication terminal>
Next, the electrical hardware configuration of the communication terminal 5 will be described with reference to Fig. 3. Fig. 3 is a diagram showing the electrical hardware configuration of the communication terminal.

As shown in FIG. 3, the communication terminal 5 is a computer and includes a CPU 501, a ROM 502, a RAM 503, an SSD 504, an external device connection I/F (Interface) 505, a network I/F 506, a display 507, a pointing device 508, a media I/F 509, and a bus line 510.

Of these, the CPU 501 controls the overall operation of the communication terminal 5. The ROM 502 stores programs, such as IPL, used to drive the CPU 501. The RAM 503 is used as a work area for the CPU 501.

The SSD 504 reads and writes various data under the control of the CPU 501. Note that instead of the SSD 504, a hard disk drive (HDD) may be used.

The external device connection I/F 505 is an interface for connecting various external devices, such as a display, a speaker, a keyboard, a mouse, a USB memory, and a printer.

The network I/F 506 is an interface for performing data communication via the communication network 100 .

The display 507 is a type of display means, such as a liquid crystal display or an organic EL (Electro Luminescence) display, that displays various images.

The pointing device 508 is a type of input means for selecting and executing various instructions, selecting a processing target, moving a cursor, etc. When the user uses a keyboard, the function of the pointing device 508 may be turned off.

The media I/F 509 controls reading and writing (storing) of data from and to a recording medium 509m such as a flash memory. The recording medium 509m includes DVDs and Blu-ray Discs (registered trademark).

The bus line 510 is an address bus, a data bus, or the like for electrically connecting the various components such as the CPU 501 shown in FIG.

[Functional configuration of the specification-compliant data estimation device]
Next, the functional configuration of the specification correspondence data estimation device 3 according to this embodiment in the learning phase and the estimation (prediction) phase will be described.

Here, the input domain is X and the output label is Y. In the following, Y is a set of discrete values Y={1,2,…,K}. The hypothesized function h is a function that receives elements of X and outputs elements of Y, and h
This is expressed as follows.

<Functional configuration of the learning phase>
In the learning phase, the specification-compliant data estimation device 3 calculates a training loss based on a set S'=[( _x1 , _y1 ), ( _x2 , _y2 ), ..., ( _xn , _yn )] of training examples and a finite or infinite set H of hypotheses h.

The hypothesis that minimizes

Here, the set of hypotheses H is, for example, the parameters

For example, the set of all neural networks characterized by

represents a neural network when specific values are set for the parameters.

is a loss function that takes two arguments y, y'∈Y and outputs a non-negative real number.

A loss function is a function that outputs a large value depending on the degree of difference between two inputs. For example, 0-1 loss

If y = y', then

If y ≠ y',

This is a function that outputs

Next, each function of the specification correspondence data estimation device 3 in the learning phase will be described in detail. Fig. 4 is a functional configuration diagram of the specification correspondence data estimation device in the learning phase.

4, the specification correspondence data estimation device 3 includes an acquisition unit 31, a specification verification unit 32, and a machine learning unit 35. These units are functions realized by instructions from the CPU 301 in Fig. 2 based on a program. In addition, the RAM 303 or the SSD 304 stores a machine learning model 30a.

The acquisition unit 31 acquires training example data S (input data x, output data y) and a specification function c from the communication terminal 7. The acquisition unit 31 also serves as an input unit for inputting learning data for machine learning. Specifically, the acquisition unit 31 acquires a plurality of training example data S=[( _x1 , _y1 ), ..., ( _xn , _yn )] from an external device such as the communication terminal 5, and a specification function c that outputs whether each of the plurality of training examples satisfies a predetermined specification A. Here, the specification function c is a function c(x, y) that receives an input/output data pair (x, y) ∈ X×Y and outputs result information indicating whether it satisfies the predetermined specification. The specification function c(x, y) outputs c(x, y)=1 if the input/output data pair (x, y) satisfies the specification, and outputs c(x, y)=0 if it does not. In addition, the predetermined specification A can be, for example,
These include age restrictions on those to whom information is being offered, such as when it is prohibited to recommend certain goods or services to minors.

The specification verification unit 32 uses the specification function c acquired by the acquisition unit 31 to verify the predetermined training example data S' that satisfies the predetermined specification A among the multiple training example data S. In this case, the specification verification unit 32 manages (holds) the specification function c for use in the estimation phase. Specifically, the specification verification unit 32 uses the specification function c acquired from the acquisition unit 31 to verify whether or not the pair (x _i , y _i ) in the training example S satisfies the above-mentioned predetermined specification A.

The machine learning unit 35 trains the machine learning model 30g based on a hypothesis h (input data x, output data y) that minimizes the training loss for a set S' of training examples. Specifically, the machine learning unit 35 trains the machine learning model 30g based on a hypothesis h (input data x, output data y) that minimizes the training loss by receiving only the training examples S' that satisfy the specified specification A from the specification verification unit 32.

Perform the process to obtain the value.

<Functional configuration of the estimation phase>
In the inference phase, the specification-based data estimation device 3 estimates the hypothesis estimated in the learning phase.

and based on the input data x ∈ X,

In the following, the hypothesis h:X→Y is expressed by a function h(x,y) that takes (x,y)∈X×Y as an argument and outputs a real number.

Let us assume that the conditional probability distribution, such as neural networks,

A hypothesis that represents

Next, each function of the specification correspondence data estimation device 3 in the estimation phase will be described in detail. Fig. 5 is a functional configuration diagram of the specification correspondence data estimation device in the estimation phase.

As shown in Fig. 5, the specification correspondence data estimation device 3 has an acquisition unit 31, a specification verification unit 32, an estimation unit 37, and an output unit 39. These units are functions realized by instructions from the CPU 301 in Fig. 2 based on a program. In addition, the RAM 303 or the SSD 304 stores the trained machine learning model 30b. Note that the same reference numerals are used for the functional configurations similar to those in the learning phase, and the description thereof will be omitted.

The estimation unit 37 uses the trained machine learning model 30b and estimates and outputs output data y that satisfies a predetermined specification by a specification function c managed by the specification verification unit 32. Specifically, the estimation unit 37 uses the hypothesis obtained in the learning phase

Using the specification function c, we find an appropriate y∈Y that satisfies c(x,y)=1.

In addition, y is

It is required as follows.

The output unit 39 outputs information on the estimation result indicating the output data y satisfying the predetermined specifications estimated by the estimation unit 37 from the specification-compliant data estimation device 3. Examples of output include displaying the information on a display connected to the external device connection I/F 305 in Fig. 2, transmitting the information to an external device via the network I/F 306, etc.

[Processing or operation of the specification-compliant data estimation device]
Next, the processing or operation of the specification correspondence data estimation device 3 in the learning phase and the estimation phase will be described with reference to FIG. 6 and FIG.

<Processing or Actions in the Learning Phase>
FIG. 6 is a flowchart showing the processes or operations executed by the specification-compliant data estimation device in the learning phase.

S11: The acquisition unit 31 acquires training examples S (input data x, output data y) and a specification function c based on input from the communication terminal 7 or directly to its own device (specification-compliant data estimation device).

S12: The specification verification unit 32 acquires the training example S acquired by the acquisition unit 31 from the training example S.
In this case, the specification verification unit 32 stores the specification function c for use in the estimation phase.

For example, the specification verification unit 32 selects all samples that satisfy the specification function c(x _i , y _i )=1 from the training example data S input from the acquisition unit 31 and verifies the set data obtained by the selection.

Create a.

S13: The machine learning unit 35 trains the machine learning model 30a based on a hypothesis h (input data x, output data y) that minimizes the training loss for the set data S' of training examples. The machine learning model 30a uses an existing algorithm.

This makes it possible to theoretically demonstrate that the specification-compliant data estimation device 3 can select a hypothesis with fewer estimation errors while taking external specifications into account.

This completes the explanation of the processing or operation in the learning phase.

<Processing or Actions in Estimation Phase>
FIG. 7 is a flowchart showing the processes or operations executed by the specification correspondence data estimation device in the estimation phase.

S21: The acquisition unit 31 acquires input data xεX based on an input from the communication terminal 7 or directly to the own device (specification correspondence data estimation device 3).

S22: The estimation unit 37 uses the trained machine learning model 30b and outputs output data y that satisfies the predetermined specification A using the specification function c managed (stored) in the specification verification unit 32.

S23: The output unit 39 outputs information on the estimation result.

This completes the description of the processing or operation in the estimation phase.

[Effects of the embodiment]
As described above, according to the present embodiment, it is possible to provide an effect that appropriate machine learning can be performed in accordance with various specifications (rules).

〔supplement〕
The present invention is not limited to the above-described embodiment, and may have the following configurations or processes (operations).
(1) The specification correspondence data estimation device 3 can be realized by a computer and a program, but this program can also be recorded on a (non-transitory) recording medium or provided via the communication network 100.
(2) In the communication between the specification correspondence data estimation device 3 and the communication terminal 5, another device (such as a server or a router) may relay data. For example, for the sake of simplicity, this specification describes that the input unit 31 of the specification correspondence data estimation device 3 transmits data to the communication terminal 5, but this transmission process is intended to include cases where another device relays data.
(3) In the above embodiment, a notebook computer is shown as an example of the communication terminal 5, but this is not limited to this and may be, for example, a desktop computer, a tablet terminal, a smartphone, a smart watch, a car navigation device, a refrigerator, a microwave oven, etc.
(4) Each of the CPUs 301 and 501 may be multiple instead of single.

1 Communication system 3 Specification correspondence data estimation device 5 Communication terminal 30a Machine learning model 30b Trained machine learning model 31 Acquisition unit (input unit)
32 Specification verification unit 35 Machine learning unit 37 Estimation unit 39 Output unit

Claims (8)

A specification-compliant data estimation device that performs machine learning on a machine learning model in a learning phase in accordance with a predetermined specification,
an acquisition unit that acquires a plurality of training example data, which are pairs of input data and output data, and a specification function that outputs whether each of the plurality of training examples satisfies the predetermined specification;
a specification verification unit that verifies, using the specification function, a predetermined training example data that satisfies the predetermined specification among the plurality of training example data;
a machine learning unit that trains the machine learning model by machine learning using the predetermined training example data;
A specification-compliant data estimation device having the above configuration. The specification-compliant data estimation device according to claim 1 , wherein the predetermined specification indicates an age restriction for a person to whom the specification is to be provided. A machine learning method executed by a specification-compliant data estimation device that trains a machine learning model by machine learning in a learning phase in accordance with a predetermined specification, comprising:
The specification correspondence data estimation device includes:
An acquisition process for acquiring a plurality of training example data, which are pairs of input data and output data, and a specification function for outputting whether each of the plurality of training examples satisfies the predetermined specification;
a specification verification process for verifying, using the specification function, a predetermined training example data that satisfies the predetermined specification among the plurality of training example data;
A machine learning process for training the machine learning model by machine learning using the predetermined training example data;
A machine learning method to perform. A program causing a computer to execute the method according to claim 3. A specification-compliant data estimation device that performs estimation corresponding to a predetermined specification using a trained machine learning model in an estimation phase,
An acquisition unit that acquires predetermined input data;
An estimation unit that uses a trained machine learning model and estimates output data that satisfies the predetermined specifications using a specification function that outputs whether each of a plurality of training examples, which are pairs of input data and output data, satisfies the predetermined specifications;
an output unit that outputs information of an estimation result indicating the estimated output data;
A specification-compliant data estimation device having the above configuration. The specification-compliant data estimation device according to claim 5 , wherein the predetermined specification indicates an age restriction for a person to whom the data is to be provided. A specification-compliant data estimation method executed by a specification-compliant data estimation device that performs estimation corresponding to a predetermined specification using a trained machine learning model in an estimation phase, comprising:
The specification correspondence data estimation device includes:
An acquisition process for acquiring predetermined input data;
An estimation process that uses a trained machine learning model and estimates output data that satisfies the predetermined specifications using a specification function that outputs whether each of a plurality of training examples, which are pairs of input data and output data, satisfies the predetermined specifications;
an output process for outputting information of an estimation result indicating the estimated output data;
A specification-enabled data estimation method to perform. A program causing a computer to execute the method according to claim 7.