Estimation device

Abstract

An estimation device 400 has: an acquisition unit 421 that acquires a plurality of inference results that are respectively inferred as a result of inputting, to a learning model, multiple pieces of candidate data created on the basis of information indicating an unknown attribute candidate; a calculation unit 422 that for each inference result calculates the distance between the inference result acquired by the acquisition unit 421 and a label corresponding to the candidate data; and an estimation unit 423 that estimates an unknown attribute value in accordance with the result calculated by the calculation unit 422.

Description

[Technical Field]

[0001] The present invention relates to an estimation device, an estimation method, and a recording medium. [Background technology]

[0002] BACKGROUND ART For the purpose of risk assessment of a learning model trained using machine learning or the like, a technique is known that estimates data used during learning based on output from the learning model.

[0003] For example, Non-Patent Document 1 describes a method for outputting a plausible attribute value by inputting known attributes and true labels of target data and executing a predetermined process. For example, according to Non-Patent Document 1, unknown attributes to be estimated are fixed at a certain value, and an estimated label to be output by a decision tree is calculated. Then, a hypothesized error function is used to calculate the deviation between the true label and the estimated label, and the calculated deviation is used as a weight to evaluate the marginal probability. According to Non-Patent Document 1, for example, a plausible attribute value is identified as a result of the process described above. Note that technologies related to Non-Patent Document 1 include, for example, Non-Patent Document 2.

[0004] Furthermore, examples of documents describing machine learning include Patent Document 1. For example, Patent Document 1 describes providing acquired data to a trained machine learning model, causing the trained machine learning model to execute a predetermined inference, and as a result, obtaining an inference result for the data. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2021 / 014878 [Non-patent literature]

[0006] [Non-Patent Document 1] Matthew Fredrikson et al., Privacy in pharmacogenetics: An end-to-end case study of personalized warfarin dosing,Aug 2014 [Non-patent document 2] Matthew Fredrikson et al., Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures, October 2015 Summary of the Invention [Problem to be solved by the invention]

[0007] In the case of the techniques described in Non-Patent Document 1 and Non-Patent Document 2, various prior knowledge and assumptions are required to make an estimation, such as the need to assume the shape of the error function and the need to know the marginal probabilities, etc. Therefore, there is a problem that if such knowledge is not available or no assumptions are made, it may not be possible to accurately estimate data.

[0008] SUMMARY OF THE INVENTION It is therefore an object of the present invention to provide an estimation device, an estimation method, and a recording medium that can solve the above-mentioned problems. [Means for solving the problem]

[0009] In order to achieve this object, an estimation device according to one embodiment of the present disclosure includes: an acquisition unit that acquires a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; a calculation unit that calculates, for each of the inference results acquired by the acquisition unit, a distance between the inference result and a label corresponding to the candidate data; an estimation unit that estimates a value of an unknown attribute according to the result calculated by the calculation unit; have The structure is as follows.

[0010] Furthermore, an estimation method according to another aspect of the present disclosure includes: The information processing device acquiring a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; Calculating a distance between the acquired inference result and a label corresponding to the candidate data for each of the inference results; Estimate the value of the unknown attribute based on the calculated result The structure is as follows.

[0011] Furthermore, a recording medium according to another aspect of the present disclosure includes: In the information processing device, acquiring a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; Calculating a distance between the acquired inference result and a label corresponding to the candidate data for each of the inference results; Estimate the value of the unknown attribute based on the calculated result It is a computer-readable recording medium that records a program for implementing the processing. [Effects of the Invention]

[0012] According to the above-described configurations, it is possible to provide an estimation device, an estimation method, and a recording medium that are capable of accurately estimating data. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a diagram illustrating an example of the configuration of a risk assessment system according to a first embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram illustrating a configuration example of a model storage device. [Figure 3] FIG. 1 is a block diagram showing an example of the configuration of a risk assessment device. [Figure 4] FIG. 10 is a diagram illustrating an example of advance information. [Figure 5] FIG. 10 is a diagram illustrating another example of advance information. [Figure 6] 10 is a flowchart illustrating an example of the operation of the risk assessment device when estimating attributes. [Figure 7] 10 is a flowchart showing an example of the operation of the risk assessment device during risk assessment. [Figure 8] FIG. 10 is a diagram illustrating another example of advance information. [Figure 9] FIG. 10 is a diagram illustrating an example of a hardware configuration of an estimation device according to a second embodiment of the present disclosure. [Figure 10] FIG. 1 is a block diagram illustrating an example of the configuration of an estimation device. DETAILED DESCRIPTION OF THE INVENTION

[0014] [First embodiment] A first embodiment of the present disclosure will be described with reference to FIGS. 1 to 8. FIG. 1 is a diagram illustrating an example of the configuration of a risk assessment system 100. FIG. 2 is a block diagram illustrating an example of the configuration of a model storage device 200. FIG. 3 is a block diagram illustrating an example of the configuration of a risk assessment device 300. FIG. 4 is a diagram illustrating an example of prior information 341. FIG. 5 is a diagram illustrating another example of the prior information 341. FIG. 6 is a flowchart illustrating an example of the operation of the risk assessment device 300 during attribute estimation. FIG. 7 is a flowchart illustrating an example of the operation of the risk assessment device 300 during risk assessment. FIG. 8 is a diagram illustrating another example of the prior information 341.

[0015] In the first embodiment of the present disclosure, a risk assessment system 100 will be described that, when some of the attributes constituting the training data used in training the learning model 241 are missing due to reasons such as concealment, estimates the value of the missing attribute using known attributes. For example, the risk assessment system 100 estimates the value of the attributes (x1, x2, x3) constituting the training data. 2、 …, x d ) some attribute values ​​(x 2、 …, x d ) and the unknown attribute x1 has k values ​​(v 11, …, v 1k ) in the risk assessment system 100. In this case, the risk assessment system 100 determines whether the unknown attribute X1 can take any of the following values: (v 11 , ……, v 1k ), and creates candidate data corresponding to each value. The risk assessment system 100 inputs each created candidate data into the learning model 241 to obtain an inference result corresponding to each candidate data. The risk assessment system 100 then calculates the distance (e.g., residual) between each of the obtained inference results and the known label, and determines whether the unknown attribute is (v 11 , ……, v 1k ) that the unknown attribute value will take. In this way, the risk assessment system 100 described in this embodiment estimates the unknown attribute value by calculating the distance between the inference result for the candidate data created based on known knowledge and the known label. Furthermore, the risk assessment system 100 can perform risk assessment according to the risk of training data leakage, etc., based on the attribute value estimation result.

[0016] In this embodiment, the learning model 241 is generated by supervised learning using a plurality of training data. For example, the learning model 241 is trained using a plurality of training data including a plurality of attributes and labels so as to output a label indicating whether or not a subject is sick in response to input of a plurality of attributes such as gender, age, height, weight, etc. Specific examples of attributes and labels are not limited to the above examples and may be set arbitrarily. The model trained using the training data may be any model such as a decision tree or a neural network. The attributes may also be referred to as explanatory variables or features. The labels may also be referred to as objective variables.

[0017] Furthermore, the risk assessment system 100 described in this embodiment estimates unknown attributes, for example, when the learning model 241 is in a black box setting. For example, a model generated by machine learning may be in a black box setting in which only the output for an input is made public to the user, or in a white box setting in which model information such as the model structure and branching conditions is also made public. As will be described later, the risk assessment system 100 in this embodiment can estimate unknown attributes without using information made public by the white box setting.

[0018] Fig. 1 shows an example of the configuration of a risk assessment system 100 in this embodiment. Referring to Fig. 1, the risk assessment system 100 includes, for example, a risk assessment device 300 and a model storage device 200. As shown in Fig. 1, the risk assessment device 300 and the model storage device 200 are connected to each other so as to be able to communicate with each other, for example, via a network or the like.

[0019] The model storage device 200 is an information processing device in which a learning model 241 learned using training data is stored. FIG. 2 shows an example of the configuration of the model storage device 200. For example, referring to FIG. 2, the model storage device 200 has a memory unit 240 in which the learning model 241 is stored, as well as a receiving unit 210, an inference unit 220, and an output unit 230. For example, the model storage device 200 has a calculation unit such as a CPU (Central Processing Unit) and a storage device. The model storage device 200 can realize each of the above processing units by having the calculation unit execute a program stored in the storage device.

[0020] 2, the learning model 241 stored in the storage unit 240 is trained in advance using a plurality of training data including a plurality of attributes and labels. The learning model 241 may be trained within the model storage device 200 or outside the model storage device 200.

[0021] The receiving unit 210 receives candidate data, which will be described later, from the risk assessment device 300. For example, the receiving unit 210 receives 11 , x 2、 …, x d " and "v 12 , x 2、 …, x d " and receives training data including values ​​of attributes known to the risk assessment device 300, such as "." and also including candidates for unknown attributes. As an example, the receiving unit 210 receives from the risk assessment device 300 a number of candidate data corresponding to the number of attribute candidates unknown to the risk assessment device 300. The receiving unit 210 may receive information other than the above examples, such as identification information, along with the candidate data.

[0022] The inference unit 220 inputs each piece of candidate data received by the receiving unit 210 into the learning model 241. As a result of the input, the inference unit 220 also acquires an inference label, which is an inference result corresponding to each piece of candidate data.

[0023] The output unit 230 transmits the inference label acquired by the inference unit 220 to the risk assessment device 300. For example, the output unit 230 may transmit the inference label to the risk assessment device 300 together with identification information of the candidate data so that it is possible to determine which candidate data the inference label is based on.

[0024] For example, as described above, the model storage device 200 has a learning model 241 that has been learned using training data. Furthermore, when the model storage device 200 receives candidate data from the risk assessment device 300, it performs inference using the learning model 241 based on the received candidate data, thereby acquiring an inference label corresponding to the candidate data. Then, the model storage device 200 transmits the acquired inference label to the risk assessment device 300.

[0025] The risk assessment device 300 is an information processing device that estimates the value of a concealed attribute using known knowledge such as information about known attributes. The risk assessment device 300 can also perform risk assessment based on the estimation result.

[0026] Fig. 3 shows an example of the configuration of the risk assessment device 300. Referring to Fig. 3, the risk assessment device 300 has, as main components, for example, an operation input unit 310, a screen display unit 320, a communication I / F unit 330, a storage unit 340, and an arithmetic processing unit 350.

[0027] 3 illustrates an example in which the functions of the risk assessment device 300 are realized using one information processing device. However, the risk assessment device 300 may be realized using multiple information processing devices, for example, on the cloud. For example, the functions of the risk assessment device 300 may be realized by two information processing devices: an estimation device having the functions of a candidate data creation unit 351, a candidate data transmission unit 352, an inference result acquisition unit 353, a distance calculation unit 354, and an estimation unit 355, and an evaluation device having the functions of an evaluation unit 356 and an output unit 357. Furthermore, the risk assessment device 300 may not include some of the components exemplified above, such as not having an operation input unit or a screen display unit, or may have components other than those exemplified above.

[0028] The operation input unit 310 is made up of operation input devices such as a keyboard, a mouse, etc. The operation input unit 310 detects operations of the operator operating the risk assessment device 300 and outputs the operations to the arithmetic processing unit 350.

[0029] 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.

[0030] The communication I / F unit 330 is composed of a data communication circuit etc. The communication I / F unit 330 performs data communication with an external device such as the model storage device 200 connected via a communication line.

[0031] The storage unit 340 is a storage device such as a hard disk or memory. The storage unit 340 stores processing information and a program 345 required for various processes in the arithmetic processing unit 350. The program 345 is read into the arithmetic processing unit 350 and executed to realize various processing units. The program 345 is read in advance from an external device or recording medium via a data input / output function such as the communication I / F unit 330, and is stored in the storage unit 340. Main information stored in the storage unit 340 includes, for example, prior information 341, inference result information 342, distance information 343, and estimated information 344.

[0032] The prior information 341 includes information known in advance about the training data used when training the learning model 241 stored in the model storage device 200. For example, the prior information 341 is acquired in advance using a method such as acquiring it from an external device via the communication I / F unit 330 or inputting it using the operation input unit 310, and is stored in the storage unit 340.

[0033] Fig. 4 shows an example of the prior information 341. Referring to Fig. 4, the prior information 341 includes partial training data information and missing attribute information. For example, as shown in Fig. 4, the prior information 341 can include a plurality of pieces of information that associate partial training data information with missing attribute information.

[0034] Here, the partial training data information indicates the values ​​of known attributes and corresponding labels in a state where some attributes are concealed (missing) from the training data used to train the learning model 241. For example, in FIG. 4, the attributes (x2, ..., x d ) and label y are known, and attribute x1 is missing. The missing attribute information indicates information about the value of the missing attribute. For example, in Figure 4, the missing attribute x1 is given by k values ​​(v 11 , …, v 1k ) In this embodiment, the missing attribute is, for example, a categorical variable (discrete variable).

[0035] Furthermore, the prior information 341 can include information other than the information exemplified in FIG. 4. For example, FIG. 5 shows another example of the prior information 341. For example, referring to FIG. 5, the prior information 341 can include, in addition to the information exemplified above, information indicating the marginal probability that the missing attribute takes on each candidate value. For example, as shown in FIG. 5, the prior information can include information indicating the marginal probability that the missing attribute takes on each candidate value for the unknown attribute x1. 11 , ……, v 1k ) can include information indicating the marginal probabilities corresponding to each of the above. The prior information 341 may include information other than the above examples.

[0036] The inference result information 342 includes information indicating inference labels obtained by inputting candidate data, which is created by a candidate data creation unit 351 (described later) based on prior information 341, into the learning model 241. For example, the inference result information 342 may include information indicating inference labels corresponding to the number of candidates for a missing attribute. For example, the inference result information 342 is generated and updated in response to an inference result acquisition unit 353 (described later) acquiring inference labels from the model storage device 200.

[0037] The distance information 343 includes information indicating the result of calculation of the distance between the inference label included in the inference result information 342 and the label used as training data by the distance calculation unit 354 described below. For example, the distance information 343 may include information indicating the distance according to the number of candidates for the missing attribute. For example, the distance information 343 is generated and updated as the distance calculation unit 354 calculates the distance between the inference label and the label.

[0038] The estimated information 344 includes information indicating the result of estimation made by the estimation unit 355 (described later) based on the distance information 343. For example, the estimated information 344 may include information indicating the attribute value estimated by the estimation unit 355 among the unknown attribute candidates. For example, the estimated information 344 is generated and updated in response to the estimation unit 355 estimating a likely value as the value of the unknown attribute among the candidates based on the distance between the inferred label and the label.

[0039] The arithmetic processing unit 350 has an arithmetic device such as a CPU and its peripheral circuits. The arithmetic processing unit 350 reads and executes a program 345 from the storage unit 340, thereby causing the above hardware and the program 345 to work together to realize various processing units. Major processing units realized by the arithmetic processing unit 350 include, for example, a candidate data creation unit 351, a candidate data transmission unit 352, an inference result acquisition unit 353, a distance calculation unit 354, an estimation unit 355, an evaluation unit 356, and an output unit 357.

[0040] The candidate data creating unit 351 creates candidate data based on the prior information 341. For example, the candidate data creating unit 351 creates candidate data according to the number of candidates indicated by the missing attribute information. The candidate data creating unit 351 may create candidate data at any timing.

[0041] Specifically, for example, partial training data information (x2, ..., x d , y) is stored, and the unknown attribute x1 is stored as missing attribute information (v 11 , …, v 1k In this case, the candidate data creation unit 351 stores the value of the unknown attribute x1 as (v 11 , …, v 1k ), where (v 11 , …, v 1k ) corresponding to each of the candidate data. 11 , x2, …, x d ), …, (v 1k , x2, …, x d ) candidate data is created.

[0042] As described above, the prior information 341 may include a plurality of pieces of information in which partial training data information and missing attribute information are associated with each other. The candidate data creating unit 351 may create candidate data for each piece of associated information using the method described above.

[0043] The candidate data transmission unit 352 transmits the candidate data created by the candidate data creation unit 351 to the model storage device 200. The candidate data transmission unit 352 may transmit, together with the candidate data, identification information of the candidate data according to partial training data information used when creating the candidate data.

[0044] The inference result acquisition unit 353 receives and acquires an inference label as a result of inference based on candidate data from the model storage device 200. For example, the inference result acquisition unit 353 acquires the inference label from the model storage device 200 together with identification information etc. so that the candidate data that is the subject of inference can be identified. Furthermore, the inference result acquisition unit 353 stores the received inference label in the storage unit 340 as inference result information 342. The inference result acquisition unit 353 may store the inference label in the storage unit 340 together with identification information etc. of the corresponding candidate data.

[0045] The distance calculation unit 354 calculates the distance between the inference label and the label included in the partial training data information from which the candidate data to be inferred was created, based on the prior information 341 and the inference result information 342. That is, the distance calculation unit 354 calculates the distance between the inference label and the label corresponding to the candidate data to be inferred. The distance calculation unit 354 also stores the calculated distance as distance information 343 in the storage unit 340. The distance calculation unit 354 may store the calculation result in the storage unit 340 together with identification information of the corresponding candidate data, etc.

[0046] Specifically, for example, the distance calculation unit 354 calculates the residual between the inferred label and the label as the distance between the inferred label and the label. For example, let us say that the label is represented as y and the inferred label is represented as Equation 1. In this case, the distance calculation unit 354 calculates Equation 2, which will be described later, to calculate the residual between the inferred label and the label.

number

number

[0047] For example, as described above, the distance calculation unit 354 calculates the residual between the inferred label and the label as the distance between the inferred label and the label. The distance calculation unit 354 may be configured to calculate the distance between the inferred label and the label using a known method other than the example given above, such as calculating twice the value of Equation 2 as the distance.

[0048] The estimation unit 355 estimates the value of an attribute that is likely to be an unknown attribute from among the candidates, based on the distance information 343. The estimation unit 355 also stores the estimation result as estimated information 344 in the storage unit 340.

[0049] For example, the estimation unit 355 identifies a candidate with the smallest distance based on the distance information 343, and estimates a value according to the identified result. Specifically, for example, the estimation unit 355 identifies i' by solving the following equation 3. Then, v corresponding to the identified i' is used as the most likely attribute value. 1i’ where i' takes any value from 1 to k.

number

[0050] In addition, there may be multiple i's when the residual is 0. In this case, for example, the estimation unit 355 randomly selects one of the multiple i's and calculates v 1i’ As described above, the prior information 341 may contain information indicating marginal probabilities. In this case, the estimation unit 355 may select one of the multiple i's based on the marginal probability. For example, the estimation unit 355 may select the i' with the largest marginal probability from the multiple i's. The estimation unit 355 may also be configured to select i' with a probability according to the marginal probability. In this way, when there are multiple i's, the estimation unit 355 selects one of the multiple i's by any method and outputs v according to the selection result. 1i’When there are multiple i's, the estimation unit 355 may be configured to output multiple v 1i’ may be configured to output the

[0051] The evaluation unit 356 performs an evaluation based on the estimation information 344. In other words, the evaluation unit 356 performs a risk evaluation based on the result of estimation by the estimation unit 355.

[0052] For example, the evaluation unit 356 has correct answer information, which is information indicating what value the unknown attribute indicated by the prior information 341 actually had. For example, in the example shown in FIG. 4, the evaluation unit 356 evaluates whether x1 is (v 11 , …, v 1k ) is the value. The evaluation unit 356 can compare the result of estimation by the estimation unit 355 with the actual value indicated by the correct answer information, and evaluate the risk based on the comparison result. For example, when the result of estimation by the estimation unit 355 matches the actual value indicated by the correct answer information, the evaluation unit 356 can evaluate that the risk is high. On the other hand, when the result of estimation by the estimation unit 355 does not match the actual value indicated by the correct answer information, the evaluation unit 356 can evaluate that the risk is low.

[0053] As described above, the prior information 341 includes multiple pieces of information associating partial training data information with missing attribute information. Therefore, the estimation unit 355 can estimate candidates for each piece of associated information. Therefore, for example, the evaluation unit 356 may evaluate the risk based on a comparison result between multiple estimations by the estimation unit 355 and correct answer information corresponding to each estimation. Specifically, for example, the evaluation unit 356 calculates a correct answer rate indicating the rate at which the estimation result matches the correct answer information based on the results of the multiple comparisons. The evaluation unit 356 can then output, for example, the calculated correct answer rate as information indicating the risk. The evaluation unit 356 may be configured to evaluate the risk based on whether the calculated correct answer rate exceeds a predetermined threshold, and output the evaluation result.

[0054] The output unit 357 outputs information indicating the candidates estimated by the estimation unit 355, information indicating the evaluation results by the evaluation unit 356, etc. For example, the output unit 357 displays the above information on the screen display unit 320, or transmits it to an external device via the communication I / F unit 330.

[0055] The above is an example of the configuration of the risk assessment device 300. Next, an example of the operation of the risk assessment device 300 will be described with reference to FIGS.

[0056] First, an example of the operation of the risk assessment device 300 when estimating an unknown attribute will be described with reference to Fig. 6. Fig. 6 is a flowchart showing an example of the operation of the risk assessment device 300 when estimating an unknown attribute. Referring to Fig. 6, the candidate data creation unit 351 creates candidate data based on the prior information 341 (step S101). For example, the candidate data creation unit 351 creates candidate data according to the number of candidates indicated by the missing attribute information.

[0057] The candidate data transmission unit 352 transmits each piece of candidate data created by the candidate data creation unit 351 to the model storage device 200 (step S102).

[0058] The inference result acquisition unit 353 acquires an inference label for each candidate data item from the model storage device 200 as a result of inference based on the candidate data item (step S103).

[0059] The distance calculation unit 354 calculates the distance between the inference label and the training label indicated by the corresponding partial training data information based on the inference label acquired by the inference result acquisition unit 353 (step S104). For example, the distance calculation unit 354 calculates the residual between each received inference label and the label as the distance.

[0060] The estimation unit 355 estimates a likely value as an unknown attribute from among the candidates based on the result calculated by the distance calculation unit 354 (step S105). For example, the estimation unit 355 identifies a candidate with the smallest distance based on the distance information 343, and estimates a value according to the identified result.

[0061] The above is an example of the configuration of the risk assessment device 300 when estimating attributes. For example, the risk assessment device 300 can perform the processes from step S101 to step S105 for each target to be estimated.

[0062] Next, an example of the operation of the risk assessment device 300 during risk assessment will be described with reference to Fig. 7. Fig. 7 is a flowchart showing an example of the operation of the risk assessment device 300 during risk assessment. Referring to Fig. 7, the risk assessment device 300 performs a process of estimating unknown attributes described with reference to Fig. 6 (step S201).

[0063] If the estimation target remains in the prior information 341 (step S202, No), the risk assessment device 300 returns to the processing of step S201 and performs the estimation process. On the other hand, if the estimation target no longer exists in the prior information 341 (step S202, Yes), the risk assessment device 300 performs risk assessment according to the results of each estimation (step S203). For example, the risk assessment device 300 can calculate a correct answer rate based on the comparison result between the result of each estimation and the correct answer information corresponding to each estimation, and perform output according to the calculated correct answer rate.

[0064] The above is an example of the operation of the risk assessment device 300 during risk assessment. Note that the processing of step S203 does not necessarily have to be performed immediately after the processing of steps S201 and S202. For example, the processing of step S203 may be performed at any timing after the processing of steps S201 and S202.

[0065] As described above, the risk assessment device 300 includes a distance calculation unit 354 and an estimation unit 355. With this configuration, the estimation unit 355 can estimate the value of a candidate attribute that is most likely to be an unknown attribute, based on the distance between the estimated label and the label calculated by the distance calculation unit 354. In other words, with the above configuration, it is possible to estimate an unknown attribute value without assuming an error function or knowledge of marginal probabilities. As a result, even when such knowledge is not available or no assumptions are made, i.e., even when no prior knowledge is assumed, data can be estimated more accurately.

[0066] In this embodiment, the case where there are x11 unknown attributes is exemplified, but the present invention can be applied without any problems even when there are multiple unknown attributes.

[0067] For example, Figure 8 shows the relationship between the unknown attribute x1 and x n For example, in FIG. 8, the attribute (x n+1 , …, x d ) and label y are known, and the attributes (x1, ..., x n ) is missing. In this case, the missing attribute information indicates information about the value of each missing attribute. As shown in FIG. 5, the prior information 341 may also include information indicating the marginal probability of each candidate even when there are multiple unknown attributes.

[0068] As shown in FIG. 8, when there are multiple unknown attributes, the candidate data creation unit 351 creates a number of candidate data corresponding to the combination of unknown attribute candidates, assuming that each unknown attribute takes one of the candidates. The candidate data transmission unit 352 and subsequent units can be processed in the same manner as when there is one unknown attribute. For example, when there are multiple i's, the estimation unit 355 may select one of the multiple i's, for example, randomly or based on marginal probability, as in the case when there is one unknown attribute. For example, as described above, even when there are multiple unknown attributes, the unknown attribute value can be estimated by the same processing as when there is one unknown attribute, except that the number of candidate data created by the candidate data creation unit 351 increases.

[0069] In this embodiment, the risk assessment system 100 has been exemplified as having the model storage device 200 and the risk assessment device 300. However, the risk assessment system 100 may be configured, for example, from a single information processing device that has the functions of the model storage device 200 and the risk assessment device 300 described in this embodiment. The risk assessment system 100 may also employ other known modified examples.

[0070] [Second embodiment] Next, a second embodiment of the present disclosure will be described with reference to Fig. 9 and Fig. 10. Fig. 9 is a diagram showing an example of the hardware configuration of an estimation device 400. Fig. 10 is a block diagram showing an example of the configuration of the estimation device 400.

[0071] In the second embodiment of the present disclosure, a configuration example of an estimating device 400 that is an information processing device that estimates an unknown attribute value based on information about known attributes, etc. will be described. Fig. 9 shows an example of the hardware configuration of the estimating device 400. Referring to Fig. 9, the estimating device 400 has, as an example, the following hardware configuration. ·CPU (Central Processing Unit) 401 (computing unit) ROM (Read Only Memory) 402 (storage device) RAM (Random Access Memory) 403 (storage device) Programs 404 loaded into RAM 403 A storage device 405 for storing the program group 404 A drive device 406 that reads and writes data from a recording medium 410 outside the information processing device A communication interface 407 for connecting to a communication network 411 outside the information processing device Input / output interface 408 for inputting and outputting data Bus 409 connecting each component

[0072] 10 by the CPU 401 acquiring the program group 404 and executing it. The program group 404 is stored in advance in the storage device 405 or the ROM 402, for example, and is loaded into the RAM 403 or the like by the CPU 401 for execution as needed. The program group 404 may be supplied to the CPU 401 via the communication network 411, or may be stored in advance in the recording medium 410, and the drive device 406 may read out the program and supply it to the CPU 401.

[0073] 9 shows an example of the hardware configuration of the estimating device 400. The hardware configuration of the estimating device 400 is not limited to the above-described case. For example, the estimating device 400 may be configured with a part of the above-described configuration, such as excluding the drive device 406.

[0074] The acquisition unit 421 acquires a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into the learning model.

[0075] The calculation unit 422 calculates the distance between the inference result acquired by the acquisition unit 421 and the label corresponding to the candidate data for each inference result. For example, the calculation unit 422 calculates a residual as the distance.

[0076] The estimation unit 423 estimates the value of the unknown attribute according to the result calculated by the calculation unit 422.

[0077] As described above, the estimation device 400 includes a calculation unit 422 and an estimation unit 423. With this configuration, the estimation unit 423 can estimate the value of an unknown attribute based on the distance calculated by the calculation unit 422. In other words, with the above configuration, it is possible to estimate the unknown attribute value without assuming an error function or knowledge of marginal probabilities. As a result, even when such knowledge is not available or no assumptions are made, i.e., even when no prior knowledge is assumed, data can be estimated more accurately.

[0078] The above-described estimation device 400 can be realized by incorporating a predetermined program into an information processing device such as the estimation device 400. Specifically, a program according to another aspect of the present invention is a program for implementing a process in an information processing device such as the estimation device 400, in which a plurality of candidate data created based on information indicating unknown attribute candidates is input to a learning model, a plurality of inference results are acquired, each of which is inferred as a result of the input, a calculation is made for each inference result, of the distance between the acquired inference result and the label corresponding to the candidate data, and a calculation is made to estimate a value of the unknown attribute based on the calculated result.

[0079] Furthermore, the estimation method executed by an information processing device such as the above-described estimation device 400 is a method in which the information processing device such as the estimation device 400 acquires multiple inference results inferred as a result of inputting multiple pieces of candidate data created based on information indicating unknown attribute candidates into a learning model, calculates the distance between the acquired inference result and the label corresponding to the candidate data for each inference result, and estimates the value of the unknown attribute based on the calculated result.

[0080] Even if the invention is a program having the above-described configuration, or a computer-readable recording medium having the program recorded thereon, or an estimation method, it has the same functions and effects as the above-described estimation device 400, and therefore can achieve the above-described object of the present invention.

[0081] <Additional Notes> A part or all of the above-described embodiments can be described as follows: An outline of the estimation device and the like according to the present invention will be described below. However, the present invention is not limited to the following configuration.

[0082] (Appendix 1) an acquisition unit that acquires a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; a calculation unit that calculates, for each of the inference results acquired by the acquisition unit, a distance between the inference result and a label corresponding to the candidate data; an estimation unit that estimates a value of an unknown attribute according to the result calculated by the calculation unit; have Estimation device. (Appendix 2) 10. The estimation device of claim 1, the calculation unit calculates a residual between the inference result and the label as a distance between the inference result and the label; The estimation unit estimates a value of an unknown attribute according to the residual calculated by the calculation unit. Estimation device. (Appendix 3) 10. The estimation device according to claim 1 or 2, The estimation unit estimates a value according to the specified result by identifying a candidate that has the smallest distance according to the result calculated by the calculation unit. Estimation device. (Appendix 4) 1. The estimation device according to claim 1, further comprising: The estimation unit estimates the value of the unknown attribute using information indicating the marginal probability of the unknown attribute candidate. Estimation device. (Appendix 5) 1. The estimation device according to claim 1, further comprising: a generating unit that generates candidate data corresponding to each unknown attribute candidate based on information about known attributes and information indicating unknown attribute candidates; The acquisition unit acquires an inference result inferred as a result of inputting the plurality of candidate data created by the creation unit into a learning model. Estimation device. (Appendix 6) 6. The estimation device according to claim 5, When there are a plurality of unknown attributes, the creation unit creates candidate data according to a combination of candidates for the plurality of unknown attributes. Estimation device. (Appendix 7) 10. The estimation device according to claim 1, further comprising: an evaluation unit that performs a predetermined evaluation based on the result of the estimation by the estimation unit; Estimation device. (Appendix 8) The information processing device acquiring a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; Calculating a distance between the acquired inference result and a label corresponding to the candidate data for each of the inference results; Estimate the value of the unknown attribute based on the calculated result Estimation method. (Appendix 9) 10. The estimation method according to claim 8, comprising: Calculating a residual between the inference result and the label as a distance between the inference result and the label; Estimate the value of the unknown attribute based on the calculated residual Estimation method. (Appendix 10) In the information processing device, acquiring a plurality of inference results that are inferred as a result of inputting a plurality of candidate data created based on information indicating unknown attribute candidates into a learning model; Calculating a distance between the acquired inference result and a label corresponding to the candidate data for each of the inference results; Estimate the value of the unknown attribute based on the calculated result A computer-readable recording medium that records a program for implementing processing.

[0083] Although the present invention has been described above with reference to the above-mentioned embodiments, the present invention is not limited to the above-mentioned embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention. [Explanation of symbols]

[0084] 100 Risk Assessment System 200 Model Enclosure 210 Receiving unit 220 Reasoning Department 230 Output section 240 Storage section 241 Learning Model 300 Risk Assessment Device 310 Operation input section 320 Screen display section 330 Communication I / F section 340 Storage section 341 Advance Information 342 Inference result information 343 Distance Information 344 Estimated Information 350 Processing Unit 351 Candidate Data Creation Department 352 Candidate Data Transmission Unit 353 Inference result acquisition part 354 Distance calculation unit 355 Estimation Department 356 Evaluation Department 357 Output Section 400 Estimator 401 CPU 402 ROM 403 RAM 404 Programs 405 Storage device 406 Drive Unit 407 Communication Interface 408 Input / Output Interface 409 Bus 410 Recording Media 411 Communication Network 421 Acquisition Department 422 Calculation Unit 423 Estimation Department

Claims

1. A creation unit that uses partial training data information that indicates values ​​and labels of known attributes when some attributes of training data are missing, and missing attribute information that indicates candidate values ​​that the missing attributes can take, to create candidate data corresponding to each candidate value indicated by the missing attribute information; an acquisition unit that acquires a plurality of inference results that are inferred as a result of inputting the plurality of candidate data created by the creation unit into a learning model; a calculation unit that calculates, for each of the inference results, a distance between the inference results acquired by the acquisition unit and a label indicated by the partial training data information; an estimation unit that estimates a likely unknown attribute value from among the candidate values ​​according to the result calculated by the calculation unit; have Estimation device.

2. The estimation device according to claim 1, the calculation unit calculates a residual between the inference result and the label as a distance between the inference result and the label; The estimation unit estimates a value of an unknown attribute according to the residual calculated by the calculation unit. Estimation device.

3. 3. The estimation device according to claim 1 or 2, The estimation unit estimates a value according to the specified result by identifying a candidate that has the smallest distance according to the result calculated by the calculation unit. Estimation device.

4. The estimation device according to claim 3, The estimation unit estimates the value of the unknown attribute using information indicating the marginal probability of the unknown attribute candidate when there are multiple candidates with the smallest distance. Estimation device.

5. The estimation device according to claim 1, When there are a plurality of unknown attributes, the creation unit creates candidate data according to a combination of candidates for the plurality of unknown attributes. Estimation device.

6. The estimation device according to any one of claims 1 to 5, An evaluation unit that performs risk evaluation by comparing the result of the estimation by the estimation unit with correct answer information stored in advance. Estimation device.

7. The information processing device Using partial training data information indicating the values ​​and labels of known attributes when some attributes of the training data are missing, and missing attribute information indicating candidate values ​​that the missing attributes can take, candidate data corresponding to each candidate value indicated by the missing attribute information is created; Obtaining multiple inference results that are inferred as a result of inputting each of the multiple candidate data that have been created into the learning model; Calculating a distance between the acquired inference result and a label indicated by the partial training data information for each of the inference results; Based on the calculated results, estimate the most likely unknown attribute value from among the candidate values. Estimation method.

8. 8. The estimation method according to claim 7, Calculating a residual between the inference result and the label as a distance between the inference result and the label; Estimate the value of the unknown attribute based on the calculated residual Estimation method.

9. In the information processing device, Using partial training data information indicating the values ​​and labels of known attributes when some attributes of the training data are missing, and missing attribute information indicating candidate values ​​that the missing attributes can take, candidate data corresponding to each candidate value indicated by the missing attribute information is created; Obtaining multiple inference results that are inferred as a result of inputting each of the multiple candidate data that have been created into the learning model; Calculating a distance between the acquired inference result and a label indicated by the partial training data information for each of the inference results; Based on the calculated results, estimate the most likely unknown attribute value from among the candidate values. A program to realize the processing.

Images