JP2015230570A - Learning model creation device, determination system and learning model creation method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To create a learning model having high determination accuracy.SOLUTION: A learning model creation device comprises: a reception part for receiving learning data, in which input information and a label corresponding to the input information are combined, about each of plural users; a calculation part for, calculating an index value indicating difference between a feature of the user and a feature of other user, for every user, based on the feature of the user indicated by the learning data for each user; and a learning model creation part for, when creating the learning model for determining the label which corresponds to unknown input data, based on the learning data for each of plural users, reducing a degree that the learning data of the user having large difference between the feature of the user and feature of other user, contributes to the learning model, based on the index value calculated for every user, for creating the learning model.

Description

  The present invention relates to a learning model creation device, a determination system, and a learning model creation method.

  Conventionally, in machine learning, there is a technique called supervised learning as a technique for receiving input of data to be discriminated and determining and outputting the type of data content (this is called a label). In this supervised learning, a learning model is created based on learning data given from a plurality of users. Next, the input of data whose label is unknown is received, and the label of the data is determined based on the created learning model. Note that the learning data is a collection of a set of data such as text, image, and sound and a label that is a type of content of the data from a plurality of users. The learning model is a parameter for determining a label corresponding to data.

  The determination accuracy of data whose label is unknown is greatly influenced by the original learning model. For this reason, in creating a learning model, a learning model (user model) unique to each user is created based on learning data from each user, and a learning model used for determination is determined by a simple average of the user models of each user. There is something to be created (see Non-Patent Document 1). Also, based on learning data (expert data) where the reliability of the label satisfies a predetermined standard, the reliability is calculated for each non-expert data, and a learning model is created by adding the calculated reliability to the learning data. (See Patent Document 1).

JP 2009-282686 A

Hiroshi Kajino, Yuta Tsuboi, Hisashi Kashima, "A Convex Formulation for Learning from Crowds", Proceedings of the Twenty-Sixth AAAI Conference on Artificial Intelligence

  However, in the above-described conventional technology, even when there is a user who gives data with an erroneous label (hereinafter referred to as erroneous learning data), the user model of the user is also obtained by a simple average, Since it affects the creation of the learning model used for the determination, the determination accuracy may be reduced. Moreover, when there is little expert data, it becomes difficult to calculate the reliability of non-expert data. For this reason, it may be difficult to create a learning model capable of highly accurate determination.

  The present invention has been made in view of the above, and an object of the present invention is to provide a learning model creation device, a determination system, and a learning model creation method that can create a learning model with high determination accuracy.

  In order to solve the above-described problems and achieve the object, the learning model creation device according to the embodiment accepts learning data including a combination of input information and a label corresponding to the input information for a plurality of users. A calculation unit that calculates, for each user, an index value indicating a difference between the feature of the user and the feature of another user, based on the feature of the user indicated by the reception unit and the learning data for each user; When creating a learning model for determining a label corresponding to input data whose label is unknown based on the learning data of the plurality of users, other users are based on the index value calculated for each user. And a learning model creation unit that creates the learning model by reducing the degree to which the learning data of the user whose feature difference is large contributes to the learning model. And butterflies.

  According to the present invention, it is possible to create a learning model with high determination accuracy.

FIG. 1 is a block diagram illustrating an example of a configuration of a determination system according to the first embodiment. FIG. 2 is an explanatory diagram for explaining learning data. FIG. 3 is an explanatory diagram illustrating the user model creation unit. FIG. 4 is an explanatory diagram for explaining the calculation of the degree of abnormality. FIG. 5 is an explanatory diagram for explaining the calculation of the degree of abnormality. FIG. 6 is an explanatory diagram for explaining the calculation of the degree of abnormality. FIG. 7 is an explanatory diagram illustrating a processing flow of the determination system according to the first embodiment. FIG. 8 is a block diagram illustrating an example of a configuration of a determination system according to the second embodiment. FIG. 9 is a flowchart illustrating a learning model creation process according to the second embodiment. FIG. 10 is a diagram illustrating that the learning model creating apparatus and the determination apparatus of the determination system according to the embodiment are specifically implemented using a computer.

  Hereinafter, a learning model creation device, a determination system, and a learning model creation method according to embodiments will be described in detail with reference to the accompanying drawings. In the following description, common constituent elements are given common reference numerals, and redundant description is omitted.

(First embodiment)
FIG. 1 is a block diagram illustrating an example of a configuration of a determination system according to the first embodiment. As shown in FIG. 1, the determination system 1 receives learning data D1 from a plurality of users and creates a learning model M1, and receives an input of data D2 whose label is unknown, and creates a learning model. And a determination device 20 that determines the label of the data D2 based on the learning model M1 created by the device 10.

  The learning model creation device 10 includes a learning data conversion unit 101, a learning data storage unit 102, a user model creation unit 103, an abnormality degree calculation unit 104, and a learning model creation unit 105.

  The learning data conversion unit 101 receives input of learning data D1. The learning data D1 is data that is given by a plurality of users and corresponds to teacher data in supervised learning. Specifically, the learning data D1 is data including at least a set of data such as text, image, and voice that are user input information and a label that is a type of content of the data.

  FIG. 2 is an explanatory diagram for explaining the learning data D1. As shown in FIG. 2, when the user input information is text, the learning data D1 identifies text such as “I want to play baseball”, a label such as “sports” indicating the type of the text content, and the user. And a user ID for The learning data D1 may be, for example, collected from user input to a bulletin board on the Internet, a blog, an SNS (social networking service), etc., in addition to input from a console.

  The learning data conversion unit 101 converts the input learning data D1 into a set of a feature vector in which the feature of the data is expressed by an n-dimensional vector and a numerical label in which the contents of the label are digitized. For the conversion to the feature vector, a known method generally used in machine learning is used. For example, when the data is text, there is a method of converting to a feature vector by morphological analysis, n-gram or delimiter. In general, morphological analysis is often applied to Japanese text and white space is applied to English text, and the method may be changed depending on the type of text (Japanese / English).

  In the case of morphological analysis, the data is divided into parts of speech based on preset dictionary data, and converted into feature vectors based on the divided elements. For example, the text “Today plays soccer” is divided as “Today / ha /, // soccer ////”, and is converted into a feature vector based on the divided elements.

  In the case of n-gram, a set of n characters is created while shifting the data one character at a time from the beginning, and these are converted into feature vectors. For example, if the text “Today plays soccer” is n-gram (n = 6), an element “Today is a day / day is a soccer /.../ car” is created. Based on these elements, it is converted into a feature vector.

  In the case of a delimiter character, the data is divided by a predetermined delimiter character (space, punctuation mark, comma, etc.), and converted into a feature vector based on the divided elements. For example, the text “Today we play soccer” is divided into “Today / to play soccer” in the case of the delimiter (“,”), and converted to a feature vector based on the divided elements. To do.

  Conversion to a feature vector includes a method in which the number of appearances of each element is used as it is as a feature vector, and a method in which a certain element appears (1) / not appear (0) regardless of the number of appearances. Labels are also digitized according to predetermined rules. For example, when there is a label of “sports” / “non-sports”, “sports” is digitized as “1” and “non-sports” is expressed as “0”.

  The learning data storage unit 102 stores a set of learning data D1 converted by the learning data conversion unit 101 and a user ID. The learning data storage unit 102 may use a general database (MySQL, PostgreSQL, etc.), or may be stored in a table format or a text format. The type of storage method in the learning data storage unit 102 is not limited. Further, the learning data storage unit 102 may be substituted with a work area secured on a RAM (Random Access Memory).

  The user model creation unit 103 retrieves a set of the converted learning data D1 and user ID from the learning data storage unit 102, and creates a user model that is a unique learning model for each user ID (user).

  FIG. 3 is an explanatory diagram for explaining the user model creation unit 103. As shown in FIG. 3, the user model creation unit 103 creates user models U1 and U2 corresponding to learning data D11 and D12 for each user ID. The user models U1 and U2 are learning models for each user that are created based on a set of user feature vectors and numerical labels with user IDs “1” and “2”. The user models U1 and U2 may be created by using general machine learning techniques such as logistic regression, support vector machines, and perceptrons. The created user models U1 and U2 are represented by n-dimensional vectors.

  The degree-of-abnormality calculation unit 104 is an index value indicating how much the user's characteristics and other users' characteristics are different for each user based on the characteristics of each user indicated by the learning data D1 for each user. A certain degree of abnormality is calculated. Specifically, the degree of abnormality indicating how much a certain user is different from other users is calculated by comparing the learning data D1 of each user and the user models U1 and U2. The calculation method of the degree of abnormality is roughly divided into a first calculation method that is calculated based on a difference in learning data D1 of each user, a second calculation method that is calculated based on a difference between user models U1 and U2, and a first calculation. There is a third calculation method that combines the method and the second calculation method.

  In the first calculation method, the degree of abnormality of each user is calculated based on the difference in the elements of the learning data D1 of each user. For example, the degree of abnormality of a user having extremely few or many learning data elements compared to other users is calculated to be high.

  FIG. 4 is an explanatory diagram for explaining the calculation of the degree of abnormality, and more specifically, a diagram for explaining the calculation of the degree of abnormality by the first calculation method. As shown in FIG. 4, as a specific example of the first calculation method, an average of the number of learning data of all users (“40” in the illustrated example) is calculated, and the calculated average and the number of learning data of each user are calculated. There is a method of defining the difference between each user as the degree of abnormality of each user. In addition, there is a method of defining the difference between the median value of the number of learning data of all users (“30” in the illustrated example) and the number of learning data of each user as the degree of abnormality of each user. Note that the first calculation method is not limited to the example of FIG. 4, and any other method may be used as long as the elements of the learning data D1 of each user are compared with each other.

  In the second calculation method, the degree of abnormality of each user is calculated based on the difference in the size and direction of the vectors representing the user models U1 and U2. For example, the degree of abnormality of a user whose magnitude and direction of vectors representing the user models U1 and U2 are different from others is calculated high.

5 and 6 are explanatory diagrams for explaining the calculation of the degree of abnormality, and more specifically, for explaining the calculation of the degree of abnormality by the second calculation method. Here, V ₁ , V _2, and V ₃ in FIG. 5 and FIG. 6 represent user models having user IDs “1”, “2”, and “3” as two-dimensional vectors. Further, V _a in FIG. 5 is an average vector of V ₁ , V ₂ and V ₃ . Further, V _1n , V _2n and V _3n in FIG. 6 are vectors obtained by normalizing V ₁ , V ₂ and V ₃ . Further, V _an in FIG. 6 is _an average vector of V _1n , V _2n, and V _3n .

As shown in FIG. 5, as the second calculation method of taking the average of _V 1, _{V 2} and _{V 3} is a user model for each user, the distance between the average and the _V 1, _{V 2} and _{V 3 D 1} , D ₂ and D ₃ are defined as the degree of abnormality for each user. In this case, the degree of abnormality increases as distances D ₁ , D ₂ , D ₃ between V _a and V ₁ , V _2, and V ₃ increase.

Further, as shown in FIG. 6, the second calculation method is to normalize V ₁ , V ₂ and V ₃ which are user models of each user, and calculate the average of normalized V _1n , V _2n and V _3n . Take. Then, there is a method of defining the angle formed by the averaged V _an and V ₁ , V _2, and V ₃ as the degree of abnormality for each user. In this case, the degree of abnormality increases as the angle between V _an and V ₁ , V _2, and V ₃ increases. Note that the second calculation method is not limited to the examples of FIGS. 5 and 6, and other methods may be used as long as the comparison is performed based on the user models U1 and U2 of each user.

  As a third calculation method, for example, there is a method of calculating (abnormality calculated by the first calculation method) + (abnormality calculated by the second calculation method). Note that the combination of the first calculation method and the second calculation method may use multiplication in addition to the above-described simple addition, and is not particularly limited.

  The learning model creation unit 105 creates a learning model M1 for the determination device 20 to determine a label corresponding to the data D2 based on the input learning data D1. Specifically, the learning model creation unit 105 creates a learning model M1 that integrates the user models of each user created by the user model creation unit 103, and outputs the learning model M1 to the determination device 20. In creating the learning model M1, the learning model creating unit 105 has a large feature difference from other users based on the degree of abnormality of each user calculated by the degree of abnormality calculating unit 104, that is, a user of a user having a large degree of abnormality. The learning model M1 is created by reducing the degree of contribution of the model to the learning model M1. As a method of changing the contribution of the learning data D1 of the user according to the degree of abnormality, the weight of each user model is defined from the degree of abnormality, a method of creating a learning model by the weighted average, and a weight less than a predetermined threshold There is a method of creating a learning model after excluding a corresponding user model.

  In the method of creating a learning model by weighted average, as a method of determining the weight of the user model, for example, there is a method in which the reciprocal of the degree of abnormality is used as the weight. When the reciprocal of the degree of abnormality is used as a weight, when the degree of abnormality is 0, the reciprocal cannot be calculated. Therefore, exception processing such as using a preset numerical value as a weight is performed. The method for determining the weight from the importance is not limited to the above.

For example, when w _i is the user model of user ID = i, λ _i is the weight of the user model w _i , and K is the total number of users, the learning model w ₀ may be calculated by the following equation (1). .

In the method using Expression (1), the user model w _i having a high degree of abnormality has a small contribution to the learning model w ₀ .

As a method of creating a learning model after eliminating a user model corresponding to a weight less than a predetermined threshold, the learning model w ₀ is used without using the user model w _i whose weight λ _i is lower than a predetermined threshold. There is a way to create. The threshold value setting method includes a method in which the user predetermines from the console or the like, and a method in which the average of the weights λ _i is used as the threshold value, but is not limited thereto.

For example, the learning model w ₀ when the predetermined threshold value is a may be calculated by the following equation (2).

Further, all the user models w _i whose weight is equal to or greater than the threshold value a may be handled equally to calculate the learning model w ₀ . In this case, the learning model w ₀ is given by the following equation (3).

If the weights λ _i of all user models are below the threshold value a, the learning model w ₀ may be a 0 vector.

  The determination device 20 includes a data conversion unit 201, a determination unit 202, a learning model storage unit 203, and a determination result output unit 204.

  The data conversion unit 201 receives data D2 whose label is unknown as an input. For example, when the input is text, the data D2 may be collected from arbitrary writing on a bulletin board, blog, SNS, etc. on the Internet. Similar to the learning data conversion unit 101, the data conversion unit 201 converts the received data D2 into a feature vector. The data D2 converted to the feature vector is output to the determination unit 202.

  When the determination unit 202 receives the data D2 converted into the feature vector, the determination unit 202 refers to the learning model M1 stored in the learning model storage unit 203 and determines a label corresponding to the data D2. Specifically, if the inner product of the feature vector and the learning model M1 is greater than 0, the quantification label = 1, if it is less than 0, the quantification label = 0, and if it is 0, the quantification label = 1 (“−1”). The determination unit 202 performs label determination based on the inner product value of the feature vector and the learning model M1.

  The learning model storage unit 203 stores the learning model M1 created by the learning model creation device 10. The learning model storage unit 203 may use a general database (MySQL, PostgreSQL, etc.), or may be stored in a table format or a text format. The type of storage method in the learning model storage unit 203 does not matter.

  The determination result output unit 204 converts the determination result (numerical label) obtained by the determination unit 202 into a label before digitization with reference to a correspondence table between the numerical label and label notation. Then, the determination result output unit 204 outputs the converted label as a determination result R1 that can be identified by the user. However, when the determination result obtained by the determination unit 202 is “−1”, a character string such as “determination impossible” is output as the determination result R1. Here, the output of the determination result R1 performed by the determination result output unit 204 is displayed on a display, written to a storage medium such as an HDD (Hard Disk Drive), and other information processing apparatuses on the network (for example, a PC (Personal Computer) ) Etc.).

  FIG. 7 is an explanatory diagram illustrating a processing flow of the determination system 1 according to the first embodiment. As illustrated in FIG. 7, the learning model creation device 10 receives learning data D11, D12, D13... That includes at least a set of input information and a label corresponding to the input information for a plurality of users. Then, the learning model creation apparatus 10 performs learning data conversion processing such as conversion to the feature vector indicated by the learning data D11, D12, and D13 for each user in the learning data conversion unit 101 (S1).

  Next, the learning model creation apparatus 10 creates a user model based on the converted data in the user model creation unit 103, and the abnormality degree calculation unit 104 determines the user characteristics and other users for each user. The degree of abnormality indicating the difference from the feature is calculated. Then, the learning model creation device 10 performs processing (S2) in the learning model creation unit 105 that creates the learning model M1 based on the calculated degree of abnormality and the user model. In S <b> 2, based on the degree of abnormality calculated for each user, the degree to which the user model of the user who has a large feature difference from other users contributes to the learning model M <b> 1 is reduced.

  The determination device 20 receives input of data D2 whose label is unknown. Then, in the data conversion unit 201, the determination device 20 performs data conversion processing such as conversion of the received data D2 into a feature vector (S3). Next, in the determination unit 202, the determination unit 202 receives the data D2 after data conversion, refers to the learning model M1 created in S2, and determines a label corresponding to the data D2 (S4). Then, the determination device 20 outputs the determined label (“sports” in the illustrated example) as the determination result R1.

  As described above, the learning model creation device 10 accepts learning data D1 including at least a set of input information and a label corresponding to the input information for a plurality of users. Then, the learning model creation apparatus 10 calculates, for each user, a degree of abnormality indicating a difference between the user's feature and another user's feature based on the feature indicated by the learning data D1 for each user. The learning model creation device 10 then creates a learning model M1 for determining a label corresponding to the data D2 based on the learning data D1 when the determination device 20 determines a label corresponding to the data D2. The degree of contribution of the learning data D1 of the user who has a large feature difference from the other users to the learning model M1 is reduced.

  For this reason, even when there is a user who gives erroneous learning data D1, the learning model creation apparatus 10 can reduce the degree to which the learning data D1 of the user contributes to the learning model M1, and the determination accuracy Can be created. For example, when there are a small number of users who give erroneous learning data D1, the learning data D1 and user models given by those users are significantly different from those of many users. Therefore, a user who gives erroneous learning data D1 has a high degree of abnormality. Since the degree of contribution of the learning data D1 of the user having a high degree of abnormality to the learning model M1 is small, it is possible to create the learning model M1 with high determination accuracy without taking the reliability into account with the expanded data or the like. In addition, since the determination device 20 performs label determination of the data D2 with reference to the learning model M1 created by the learning model creation device 10, the influence of the user who gives erroneous learning data D1 is suppressed and the accuracy is high. A determination can be made.

(Second Embodiment)
Next, a second embodiment will be described. In the first embodiment described above, after creating a user model for each user, the degree of abnormality for each user is obtained, and the obtained degree of abnormality is converted into a weight. Then, the learning model M1 is created using the user model and the weight. That is, in the first embodiment, a sequential method in which the above-described processing is sequentially performed to create the learning model M1 is exemplified. In the second embodiment, a learning model (a user model for each user and a learning model that integrates them) and a weight update for each user are repeated, and an iterative method that outputs a learning model when a predetermined convergence condition is satisfied is provided. Illustrate.

  FIG. 8 is a block diagram illustrating an example of the configuration of the determination system 1a according to the second embodiment. As shown in FIG. 8, the learning model creation device 10a of the determination system 1a is different from the determination system 1 in that a learning model creation unit 110 having an initialization unit 111, a learning model update unit 112, and a parameter update unit 113 is used. .

When the initialization unit 111 receives the learning data D1 from the learning data storage unit 102, the user model w _k , the learning model w ₀ , the weight λ _k , and the regularity of each user (arbitrary user ID = k is a user k) The initialization parameter η (real number) is initialized with a predetermined value. Next, the initialization unit 111 outputs the user model w _k , the learning model w ₀ and the learning data D 1 to the learning model update unit 112, and outputs the weight λ _k and the regularization parameter η to the parameter update unit 113.

The learning model update unit 112 updates the user model w _k and the learning model w ₀ of each user. Here, the details of the update process in the learning model update unit 112 will be described. First, the i-th learning data D1 given by the user k with the user ID = k is represented by (x _ki , y _ki ). Here, x _ki represents an n-dimensional feature vector, y _ki represents a label, and this label takes a value of 0 or 1. The feature vector x _ki is a column vector.

Also, assuming that the total number of elements of the learning data D1 given by the user k is I _k , an I _k × n matrix formed by vertically transposing I _k feature vectors is expressed as X _{k in the} following equation (4). Represented by

Here, x ^T _ki represents transposition of x _ki . The inverse matrix of the matrix A is denoted as A- ¹ .

The learning model update unit 112 updates the user model w _k and the learning model w ₀ of each user according to the following equation (5). In Equation (5), w _k represents a user model before update, and w ^new _k represents a user model after update. In addition, _{w 0} is updated before the learning ^{model, w new} ₀ indicates the learning model after the update.

  Here, β is a positive real number that needs to be determined in advance, and H and g are given by the following equation (6).

B _k is an I _{k -order} diagonal matrix whose (i, i) component is (1−σ (w ^T _k x _ki )) σ (w ^T _k x _ki ). Further, D _k is an I _k -th order column vector whose i component is y _ki −σ (w ^T _k x _ki ). I is an n-th unit matrix. Here, σ is a function (sigmoid function) given by the following equation (7).

  The update in the learning model update unit 112 is performed until a predetermined end condition (convergence condition) is satisfied. Here, as the end condition, there are a case where the number of updates exceeds a predetermined number, a case where the change of the learning model before and after the update is a predetermined value or less, and the like.

The parameter updating unit 113 updates the weight λ _k and the regularization parameter η in the user model of each user according to the following equation (8). The weight λ _k is an index value indicating the difference from other users.

  Here, Tr [A] indicates a trace of the matrix A. Note that, as a method of determining the degree of abnormality, for example, a method in which the reciprocal of the weight is used as the degree of abnormality can be considered, but other methods may be used and there is no particular limitation.

The learning model w ₀ obtained by repeating the updating of the learning model (the user model for each user and the learning model integrating them) and the updating of the weight for each user by the parameter updating unit 113 described above by the learning model updating unit 112. Becomes the following equation (9) from equation (5).

From the above equation (9), it can be seen that the learning model w ₀ is given by a constant multiple of the weighted average of each user model w _k . As in the case of truncating less than a predetermined threshold in the first embodiment, a certain threshold a is defined, and a user model whose weight is lower than the threshold a is not used in the above formula, but a learning model w ₀ is created. May be.

FIG. 9 is a flowchart illustrating a learning model creation process according to the second embodiment. As shown in FIG. 9, when the process in the learning model creation unit 110 is started, the initialization unit 111 initializes the user model w _k , the learning model w ₀ , the weight λ _k , and the regularization parameter η (real number). (S11). Next, the learning model update unit 112 updates the initialized user model w _k and learning model w ₀ (S12).

Next, the parameter updating unit 113 updates the weight λ _k and the regularization parameter η in the user model (S13). Next, the learning model update unit 112 updates the user model w _k and the learning model w ₀ based on the weight λ _k updated in S13 and the regularization parameter η (S14). Next, the learning model creation unit 110 satisfies a predetermined convergence condition such as whether or not the number of updates exceeds a predetermined number, whether or not the change in the learning model before and after the update is equal to or less than a predetermined value, and so on. Is determined (S15).

When the predetermined convergence condition is not satisfied (S15: NO), the learning model creation unit 110 returns the process to S13. Thus, while the predetermined convergence condition is not satisfied, the updating of the weight λ _k and the regularization parameter η in the user model and the updating of the user model w _k and the learning model w ₀ are alternately and repeatedly performed. .

If predetermined convergence condition is satisfied (S15: YES), the learning model creation unit 110 ends the processing, and outputs a learning model w ₀ computed until convergence condition is satisfied to the decision unit 20.

  It is also possible to create a program in which the processing executed by the learning model creation device 10 and the determination device 20 described in the above-described embodiment is described in a language that can be executed by a computer. In this case, when the computer executes the program, the same effect as the processing in the learning model creation device 10 and the determination device 20 according to the above-described embodiment can be obtained. Further, such a program may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read by the computer and executed to realize the same processing as that of the above-described embodiment. An example of a computer that executes the above-described program will be described below.

  FIG. 10 is a diagram illustrating that the processing in the learning model creation device 10 and the determination device 20 of the determination system 1 according to the embodiment is specifically realized using a computer. As illustrated in FIG. 10, the computer 1000 includes, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These units are connected by a bus 1080.

  The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012. The ROM 1011 stores a boot program such as BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1031. The disk drive interface 1040 is connected to the disk drive 1041. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive. The serial port interface 1050 is connected to a mouse 1051 and a keyboard 1052, for example. The video adapter 1060 is connected to the display 1061, for example.

  The hard disk drive 1031 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the above program is stored in, for example, the hard disk drive 1031 as a program module 1093 in which a command to be executed by the computer 1000 is described. For example, the functional configuration of the learning model creation device 10 illustrated in FIG. 1 (the learning data conversion unit 101, the learning data storage unit 102, the user model creation unit 103, the abnormality degree calculation unit 104, the learning model creation unit 105) and the determination device 20 A program module 1093 for executing information processing similar to the functional configuration (data conversion unit 201, determination unit 202, determination result output unit 204, learning model storage unit 203) is stored in the hard disk drive 1031.

  In addition, a work area for executing the above-described program is secured in the RAM 1012. Setting data necessary for processing is stored as program data 1094 in, for example, the memory 1010 or the hard disk drive 1031. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1031 to the RAM 1012 as necessary, and executes them.

  The program module 1093 and the program data 1094 are not limited to being stored in the hard disk drive 1031, but may be stored in a removable storage medium, for example, and read by the CPU 1020 via the disk drive or the like. Alternatively, the program module 1093 and the program data 1094 are stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.) and read by the CPU 1020 via the network interface 1070. May be issued. Needless to say, the determination system 1 including the learning model creation device 10 and the determination device 20 may be realized by the above-described one computer.

DESCRIPTION OF SYMBOLS 1, 1a ... Determination system 10, 10a ... Learning model creation apparatus 20 ... Determination apparatus 101 ... Learning data conversion part 102 ... Learning data storage part 103 ... User model creation part 104 ... Abnormality calculation part 105, 110 ... Learning model creation part 111 ... Initialization unit 112 ... Learning model update unit 113 ... Parameter update unit 201 ... Data conversion unit 202 ... Determination unit 203 ... Learning model storage unit 204 ... Determination result output units D1, D11, D12, D13 ... Learning data D2 ... Data M1 ... Learning model U1, U2 ... User model R1 ... Determination result 1000 ... Computer

Claims (8)

For a plurality of users, a reception unit that receives learning data including at least a set of input information and a label corresponding to the input information;
Based on the characteristics of the user indicated by the learning data for each user, for each user, a calculation unit that calculates an index value indicating the difference between the characteristics of the user and the characteristics of other users;
When creating a learning model for determining a label corresponding to input data whose label is unknown based on the learning data of the plurality of users, other users are based on the index value calculated for each user. A learning model creation device comprising: a learning model creation unit that creates a learning model by reducing a degree of contribution of learning data of a user having a large feature difference to the learning model. A user model creation unit that creates a user model in which the features of the user are represented by vectors for each user based on the learning data of the plurality of users,
The calculation unit calculates the index value by comparing the size or direction indicated by the user model vector created for each user, or both the size and direction for each user with other users. The learning model creation apparatus according to claim 1. The said calculation part calculates the said index value by comparing the average user model of the said user model produced for every said user with the user model for every said user. Learning model creation device. 2. The learning model creation according to claim 1, wherein the calculation unit calculates the index value by comparing the number of learning data of the user with the number of learning data of another user for each user. apparatus. The learning model creation unit changes a degree of contribution of the learning data of the user to the learning model based on whether or not an index value calculated for each user exceeds a predetermined threshold value. The learning model creation device according to any one of claims 1 to 4. The learning model creation unit, for each of the users, a process of updating the degree that the learning data of the user contributes to the learning model, and a process of updating the learning model based on the updated degree. The learning model creation apparatus according to claim 1, wherein the learning model is created repeatedly until a predetermined convergence condition is satisfied. A determination system having a learning model generation device that generates a learning model and a determination device that determines a label corresponding to input data whose label is unknown based on the learning model,
The learning model creation device includes:
For a plurality of users, a reception unit that receives learning data including at least a set of input information and a label corresponding to the input information;
Based on the characteristics of the user indicated by the learning data for each user, for each user, a calculation unit that calculates an index value indicating the difference between the characteristics of the user and the characteristics of other users;
When creating the learning model for determining the label corresponding to the input data whose label is unknown based on the learning data of the plurality of users, based on the index value calculated for each user, A learning model creating unit that creates the learning model by reducing the degree to which the learning data of the user having a large feature difference from the user contributes to the learning model;
The determination device includes:
A storage unit for storing the learning model created by the learning model creation unit;
And a determination unit configured to determine a label corresponding to input data whose label is unknown based on the stored learning model. A learning model creation method executed by a learning model creation device,
For a plurality of users, receiving learning data including at least a pair of input information and a label corresponding to the input information;
Calculating an index value indicating a difference between the feature of the user and the feature of another user for each user based on the feature of the user indicated by the learning data for each user;
When creating a learning model for determining a label corresponding to input data whose label is unknown based on the learning data of the plurality of users, other users are based on the index value calculated for each user. And a step of creating the learning model by reducing the degree to which learning data of a user having a large feature difference contributes to the learning model.

Images