Set desirability expression generation model, set target evaluation model, and grouping program and method

Abstract

To provide a model for generating expression of a set in which interaction of characteristics of objects belonging to the set is reflected.SOLUTION: A set preference level expression generation model generates set preference level expression of a set for an interest object belonging to the set. Specifically, the model causes a computer to function as: a first embedded expression generation unit for generating a plurality of first embedded data by performing dimensional compression processing relative to each of the interaction data generated for each combination of N elements that can be taken for the interest object and other objects and including interaction terms relative to characteristics of the interest object and other objects; and a second embedded expression generation unit for generating second embedded data by performing pooling processing relative to the plurality of first embedded data. Here, the model is connected to a data estimator and is trained using learning data including evaluation data of a correct answer, and as a result, the second embedded data is set preference level expression for the interest object.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a technique for making inferences regarding the characteristics and properties of a group including elements (for example, people) having predetermined characteristics, such as an organization or group related to human activities. [Background technology]

[0002] There are many situations where people interact with each other to form groups, such as English conversation classes, in-house projects, human activity organizations such as sports teams, and temporary groups such as those formed by call center operators and customers.

[0003] In this case, it is well known from experience that the characteristics of each member greatly affect the performance of the group, such as the results and satisfaction of the group. Therefore, if matching or grouping is carried out taking such characteristics into account, it is highly expected that the performance of the group will improve.

[0004] As a matching technique that takes such characteristics into consideration, for example, Patent Document 1 discloses a talent selection technique that improves the suitability of selected talent to the request of a service requester. Specifically, this technique extracts talent that matches the information of the requested service by referring to a job skill accumulation unit that stores information indicating the correspondence between the service and talent. Next, the information of the extracted talent is output to a requester device, and information on selected talent from the extracted talent is obtained from the requester device. After that, the requested service is requested to the selected talent, information indicating the evaluation content in natural language for the performed requested service is obtained from the requester device, and an evaluation corresponding to the evaluation content is determined by analyzing the character string indicated by the evaluation content, and the information in the job skill accumulation unit is updated based on the determined evaluation.

[0005] Also, for example, Patent Document 2 discloses a technology that supports optimal employee allocation while considering the compatibility between employees. In this technology, a database that stores employee information and organizational information sent from a personnel management terminal and an employee management server are provided, and the employee management server receives evaluation point information on the evaluation of compatibility between each employee and other employees from a specific external terminal, and stores employee relationship information consisting of the evaluating employee, the evaluated employee, and the evaluation points in the above database. Furthermore, a compatibility list between employees is created from the employee relationship information, and when an allocation request for a specific organization is received from the personnel management terminal, a process is performed to create an organizational compatibility list between the members of the specific organization and the members of other organizations using the compatibility list, employee information, and organizational information.

[0006] In this way, the technology disclosed in Patent Document 2 focuses on the compatibility between employees and allocates human resources, i.e., groups, based on this compatibility. As a technology for predicting the compatibility between two people that can be used for grouping, for example, Patent Document 3 discloses a speaker compatibility determination device that predicts the compatibility between two people based on the distance between their personalities.

[0007] This speaker compatibility determination device is a device capable of classifying the personality of a speaker and determining the compatibility between speakers, and specifically includes a voice analysis unit that analyzes voice data uttered by a speaker, a matrix storage unit that stores a matrix in which a plurality of segments are arranged and each of the plurality of segments is divided into a plurality of cells, each of the plurality of cells corresponding to a different personality type, a cell selection unit that selects a cell on the matrix stored in the matrix storage unit based on the analysis result by the voice analysis unit, and a compatibility determination unit that determines the compatibility between the plurality of speakers based on the relative distance on the matrix between the cells respectively selected by the cell selection unit for the plurality of speakers. Here, the compatibility determination unit calculates the above-mentioned relative distance from the vertical and horizontal path lengths of the matrix between the cells respectively selected by the cell selection unit.

[0008] Furthermore, Patent Document 4 discloses a technology for determining compatibility between objects by considering the personalities of each object. In this technology, a trained model is used in which a result value of an interaction between a first object and a second object is used as a response variable, and a compatibility score between one object and another object is determined based on a value obtained by applying the personality score of one object to be determined compatibility and the personality score of the other object to an interaction part in the trained model having an interaction part for a first explanatory variable, which is the personality score of the first object, and a second explanatory variable, which is the personality score of the second object. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] JP 2019-191975 A [Patent Document 2] JP 2009-271601 A [Patent Document 3] Patent No. 6763103 [Patent Document 4] JP 2021-117583 A Summary of the Invention [Problem to be solved by the invention]

[0010] As described above, in the past, there have been widespread attempts to evaluate organizations and groups of people with certain characteristics such as personality and job skills, i.e., groups, based on those characteristics and determine more effective human resource allocations, i.e., group divisions.

[0011] It is generally known that the performance of a group is greatly influenced not only by the characteristics of each person in the group, but also by the compatibility between the members. For example, it is empirically recognized that the compatibility between members in the workplace can have a large impact on the results of the workplace, the degree of stress, and even the tendency to quit. However, in the conventional technology disclosed in Patent Document 1, it is not at all expected that such compatibility will be reflected in the evaluation of the group.

[0012] On the other hand, the technology disclosed in Patent Document 2 certainly assigns human resources, i.e., divides employees into groups, based on the compatibility between employees. However, this technology requires each employee to evaluate their compatibility with other employees, and the group cannot be evaluated unless definitive information on the compatibility between employees is obtained. For example, it is impossible to evaluate a group that includes employees whose compatibility is unknown.

[0013] In addition, it is true that by using the technology disclosed in Patent Document 3, it is possible to classify the personality of the speaker and determine the compatibility between the speakers, and there is no need to obtain definite information regarding the compatibility between the speakers. However, Patent Document 3 does not mention using such a technology to evaluate a group, and the problem of how to automatically incorporate the compatibility determined by the technology disclosed in Patent Document 3 into the evaluation process of the group is not at all easily solved even by a person skilled in the art.

[0014] Under the circumstances described above, the inventors of the present application thought that if it were possible to generate a set embedded representation as a feature that represents the characteristics of a group (set) containing multiple elements such as members with predetermined characteristics, such as personality, and that also reflects the "interactions" between characteristics of the elements, such as compatibility between members (and their personalities), then it might be possible to use such a set embedded representation to evaluate a group, for example, taking into account the compatibility between members.

[0015] Incidentally, the above "interaction" refers to a synergistic effect within a group that only appears when two or more characteristics are combined. For example, an easy-to-understand example of "interaction" is when a teacher with a high level of extroversion is generally expected to improve students' grades in a school class, but such a teacher may be perceived as annoying by students with a high level of conscientiousness. In this case, apart from the teacher's "extroversion" itself, it is possible that the interaction between "extroversion" and "conscientiousness" may affect the results of the class, such as test scores.

[0016] In this regard, the technology disclosed in Patent Document 4, which includes the inventors of the present application, estimates compatibility by incorporating such "interactions." However, the technology does not generate the set embedding representation described above.

[0017] Therefore, the present invention aims to provide a set suitability representation generation model capable of generating a representation of a set including a plurality of objects having a predetermined characteristic, the representation of the set reflecting the interaction of the characteristic. Another aim is to provide an intra-set object evaluation model capable of performing an evaluation of a set taking into account the interaction of the characteristic using the representation of the set thus generated, and further to provide a grouping program and a grouping method capable of performing grouping taking into account the interaction of the characteristic using the intra-set object evaluation model. [Means for solving the problem]

[0018] According to the present invention, a set suitability expression generation model is provided that causes a computer to function to generate a set suitability expression, which is an embedded expression of a degree to which a set including a plurality of objects having a predetermined characteristic represented by at least one factor score is suitable for a target object belonging to the set, comprising: a first embedding representation generation unit that generates a plurality of first embedding data by performing a dimensionality reduction process on each of interaction data generated for each combination of N (N is an integer equal to or greater than 1) elements that include the target of interest as an element and are possible for the target of interest and other targets belonging to the set, the interaction data including interaction terms for the factor scores related to the target of interest and the other targets included in the combination; a second embedding representation generation unit that performs a pooling process on the plurality of first embedding data to generate second embedding data; A set desirability expression generation model that causes a computer to function as a set desirability expression generation model is provided. A first embedded expression generating unit and a second embedded expression generating unit which are functional means realized by a computer teeth, This second embedding representation generation unit: A data estimator is connected to the second embedded data and outputs evaluation data related to a predetermined output of the target object. On top of that , This first embedding generation part but As training data Interaction data The evaluation data output from this data estimator when but , as training data The correct evaluation data and The shape being compared The second embedding data has been trained and is a representation of the set preference for the object of interest.

[0019] As an embodiment of the set suitability expression generation model according to the present invention, The set desirability expression generation model causes a computer to function as a plurality of sets of a first embedding expression generation unit and a second embedding expression generation unit, the sets having different values ​​of N, and further causes the computer to function as a set desirability expression generation unit that combines the second embedding data generated in each set to generate new second embedding data, It is also preferable that the new second embedding data is, through the training, a representation of the set suitability for the object of interest.

[0020] Furthermore, as another embodiment of the set desirability expression generation model according to the present invention, the set desirability expression generation model is: a third embedding representation generating unit that performs a dimensionality reduction process on each of the interaction data generated for each possible combination of N elements for the other objects belonging to the set, the interaction data including an interaction term for the factor score for the other objects included in the combination, to generate a plurality of third embedding data; a fourth embedding representation generation unit that performs a pooling process on the plurality of third embedding data to generate fourth embedding data; a set desirability expression generating unit that combines the second embedding data and the fourth embedding data to generate new second embedding data; to make the computer function better, It is also preferable that the new second embedding data is, through the training, a representation of the set suitability for the object of interest.

[0021] In addition, in the above embodiment in which the fourth embedding data is generated, it is also preferable that the set desirability expression generation unit uses an attention mechanism to generate the new second embedding data by weighting and mixing the second embedding data and the fourth embedding data.

[0022] Further, in the above embodiment for generating the fourth embedding data, The set desirability expression generation model causes a computer to function as a plurality of sets of a third embedding expression generation unit and a fourth embedding expression generation unit, the plurality of sets having different values ​​of N between the sets, and further causes the computer to function as a set desirability expression generation unit that combines the second embedding data with the fourth embedding data generated in each set to generate new second embedding data; It is also preferable that the new second embedding data is, through the training, a representation of the set suitability for the object of interest.

[0023] In addition, as yet another embodiment of the set suitability expression generation model according to the present invention, it is also preferable that the interaction data is data that does not include an interaction term that has been previously determined to have a low significance of the interaction between the factor scores.

[0024] Furthermore, as yet another embodiment of the set suitability representation generation model according to the present invention, it is also preferable that the first embedded representation generation unit has been pre-trained as an encoding unit of an autoencoder, or is connected to an estimator that receives the first embedded data as input and outputs data related to a predetermined output, and has been pre-trained using training data including the interaction data and ground truth data.

[0025] According to the present invention, the above-mentioned set desirability expression generation model A second embedding representation generating unit, which is a functional means realized by a computer The in-set object evaluation model is provided, which causes a computer to function as a trained evaluation estimator that takes as input the set desirability representation of the object of interest output from and outputs evaluation information related to a predetermined output of the object of interest.

[0026] In the in-set target evaluation model according to the present invention, it is also preferable that the evaluation estimator receives as input at least one of factor score data including the factor score of the target of interest, factor statistical data including statistical values ​​of the set of factor scores related to the target of interest and the other targets, and / or other target factor statistical data including statistical values ​​of the set of factor scores related to the other targets, and outputs evaluation information related to a specified output of the target of interest.

[0027] According to the present invention, further comprising: a group candidate setting means for setting, from a predetermined object group, a plurality of group candidates whose members are objects included in the object group; The set suitability expression generation model described above The second embedding data output from the second embedding expression generating unit, which is a functional means realized by a computer, For each group candidate, a set suitability expression of each member of the group candidate is generated by using the above, and the set suitability expression is expressed as 、The above-mentioned in-group object evaluation model A functional means implemented in a computer by the input unit for expressing a set suitability Enter in Using the evaluation information output from an evaluation information output unit, which is a functional means realized by a computer using the in-set object evaluation model, a group candidate evaluation means for determining an estimate of evaluation information for each member of the group candidate; a grouping implementation means for selecting a plurality of the set group candidates based on a grouping condition including a condition that an estimated value of evaluation information for each member in each group candidate or a statistical value of the estimated value satisfies a predetermined criterion, and setting the selected plurality of group candidates as a grouping result for the predetermined target group; A grouping program is provided that causes a computer to function as a

[0028] According to the present invention, there is also provided a method for producing a medicament for a medicament comprising the steps of: A step of setting a plurality of group candidates from a predetermined target group, the group candidates having targets included in the target group as members; The set suitability expression generation model described above The second embedding data output from the second embedding expression generating unit, which is a functional means realized by a computer, For each group candidate, a set suitability expression of each member of the group candidate is generated by using the above, and the set suitability expression is expressed as 、 The above-mentioned in-group object evaluation model A functional means implemented in a computer by the input unit for expressing a set suitability Enter in Using the evaluation information output from an evaluation information output unit, which is a functional means realized by a computer using the in-set object evaluation model, determining an estimate of reputation information for each member of the candidate group; a step of selecting a plurality of the set group candidates based on a grouping condition including a condition that an estimated value of evaluation information for each member in each group candidate or a statistical value of the estimated value satisfies a predetermined criterion, and setting the selected plurality of group candidates as a grouping result for the predetermined target group; A computer-implemented grouping method is provided, comprising: Effect of the Invention

[0029] According to the set suitability representation generation model of the present invention, it is possible to generate a representation of a set including a plurality of objects having a predetermined characteristic, the representation reflecting the interaction of the characteristic. Also, according to the in-set object evaluation model of the present invention, it is possible to perform evaluation of a set taking into account the interaction of the characteristic. Furthermore, according to the grouping program and grouping method of the present invention, it is possible to perform grouping taking into account the interaction of the characteristic. [Brief description of the drawings]

[0030] [Figure 1] FIG. 2 is a schematic diagram for explaining an embodiment of a set desirability expression generation model according to the present invention. [Diagram 2] FIG. 11 is a schematic diagram for explaining another embodiment of a set desirability expression generation model according to the present invention. [Diagram 3] FIG. 11 is a schematic diagram for explaining still another embodiment of the set desirability expression generation model according to the present invention. [Figure 4] 13 is a schematic diagram for explaining a modified mode of the set desirability expression generation process in the set desirability expression generation model according to the present invention. FIG. [Diagram 5] 1 is a schematic diagram for explaining an embodiment of a set-in-set target evaluation model that estimates an evaluation result of a target of interest using a set desirability expression output from a set desirability expression generation model according to the present invention. FIG. [Figure 6] 1 is a functional block diagram showing a functional configuration of an embodiment of a set desirability expression generating device according to the present invention; [Figure 7] FIG. 2 is a schematic diagram showing a specific example of a grouping method including a set desirability expression generating method according to the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0031] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

[0032] [Set desirability expression generation model] FIG. 1 is a schematic diagram for explaining an embodiment of a set desirability expression generation model according to the present invention.

[0033] The set suitability representation generation model 1 of this embodiment shown in Figure 1 is a model that causes a computer to function to generate a set embedding representation, which is an embedding representation of a "set" to which multiple "objects" having a specified characteristic represented by at least one factor score belong.

[0034] Here, word embedding, which is a representation of each "word" in a sentence, is well-known as a so-called embedding representation, and this word embedding is an essential technology in natural language processing using artificial intelligence. The set suitability representation generation model 1 is a model that can generate the "set" version of this.

[0035] In addition, in FIG. 1, as a specific example, group A, which includes M (M is an integer of 2 or more) members m1, m2, ..., is the target for generating a set embedding representation. Here, each member has a personality represented by a score of the Big Five factors (Big Five score). The Big Five factors are personality information proposed in the Five Factor Model (FFM) proposed by Goldberg et al. and are widely accepted, and specifically, they consist of five factors: openness (O), conscientiousness (C), extraversion (E), agreeableness (A), and emotional stability (N). For example, the personality of member m1 can be represented by the Big Five score, which is a set of five score variables (O1, C1, E1, A1, N1), all of which take values ​​from 0 to 1.

[0036] Incidentally, the Big Five scores of each member may be collected by conducting a questionnaire survey using a known questionnaire, or may be estimated from various behavioral logs of each member. For example, it is preferable to estimate the Big Five scores of each member from the browsing history of web pages of the member using the technology disclosed in JP 2019-197338 A.

[0037] In this embodiment, the set suitability expression generation model 1 takes member m1 belonging to group A as the "target of interest" and generates a set suitability expression that embeds (expresses) the "degree to which group A (set) is suitable for member m1 (target of interest)," such as the compatibility between member m1 (target of interest) and group A (set) and the good or bad position of member m1 (target of interest) within group A (set).

[0038] Specifically, the set desirability expression generation model 1 generates (M-1) pieces of “interaction data” generated from the Big Five score of member m1 (target) and the Big Five scores of other members of group A, i.e. x1 = (O1×O2, O1×C2, O1×E2, ...): Interaction data between member m1 and member m2 x2 = (O1×O3, O1×C3, O1×E3, ...): Interaction data between member m1 and member m3 x3 = (O1×O4, O1×C4, O1×E4, ...): Interaction data between member m1 and member m4 ... ... is input, and a set desirability expression of group A is output. Here, in this embodiment, the output set desirability expression is a multi-dimensional set embedding expression vector, for example, an enumeration of multiple numerical values.

[0039] Furthermore, "interaction" refers to a synergistic effect within a group that only appears when two or more characteristics are combined. For example, an easy-to-understand example of "interaction" is when a teacher with high extraversion (E) is expected to generally improve students' grades in a school class, but such a teacher may be perceived as annoying by students with high conscientiousness (C). In this case, apart from the teacher's extraversion (E) itself, it is possible that the interaction between extraversion (E) and conscientiousness (C) may affect the results of the class, such as test scores.

[0040] Furthermore, the above-mentioned "interaction data" contains member m1 (target) as an element and is a total of (M-1) pieces of vector data, with the elements being "interaction terms" for the Big Five scores (factor scores) for the two members included in each combination (a total of (M-1) combinations) of two (N=2) elements possible for member m1 and other (M-1) members (member m2, member m3, ...).

[0041] Here, for example, in the combination of member m1 and member m2, the "interaction terms" are O1×O2, O1×C2, O1×E2, etc., which are the products of one Big Five score of member m1 and one Big Five score of member m1, resulting in a total of 25 (=5×5) terms. Therefore, one "interaction data", for example, the above x1=(O1×O2, O1×C2, O1×E2, . . . ), is vector data of 25 dimensions (25 channels). Incidentally, the "interaction terms" are not limited to those which take the product of all Big Five scores (factor scores) in this way. For example, as will be explained later, it is also preferable to exclude interaction terms which have a so-called p-value of a predetermined value or more and are judged to have a small effect on the characteristics of group A (set).

[0042] As shown in FIG. 1, in order to generate a set desirability expression using the “interaction data” as described above, the set desirability expression generation model 1 has the following features: (A) a first embedded representation generation unit 11 that generates a plurality of "first embedded data" by performing a dimensionality reduction process on each of "interaction data" generated for each combination of N (N is an integer equal to or greater than 2, in the above example, N=2) elements that include a target object (member m1) as an element and are possible for the target object (member m1) and other objects (members m2, m3, . . . ) and that include interaction terms for factor scores (Big Five scores) related to the target object (member m1) and other objects (members m2, m3, . . . ) included in the combination; (B) a second embedding representation generation unit 12 that performs a pooling process on a plurality of “first embedding data” to generate “second embedding data”; to make the computer function. (C1) The set desirability representation generation model 1 is connected to a “data estimator” (e.g., an evaluation estimator 81 in FIG. 5 described later) that receives the “second embedded data” as input and outputs evaluation data related to a predetermined output of the target of interest (member m1), and is trained using learning data including the “interaction data” and the correct evaluation data; (C2) As a result, the "second embedded data" represents a set desirability for the target object.

[0043] Here, the "second embedded data" which is a set suitability expression is generated by the pooling process (B) described in detail below, and as a result, it is a permutation invariant quantity that does not depend on rearrangement of data related to elements (members) belonging to the set (group A). ​​This permutation invariance is naturally required for the expression (feature) of a set, and the "second embedded data" is appropriate to be called the expression (feature) of a set.

[0044] As shown in (A) above, the set suitability expression generation model 1 receives as input the "interaction data" generated for each combination that includes the target of interest (member m1) as an element. As a result, it is possible to generate an embedded expression that reflects the interaction of a predetermined characteristic (personality in this specific example) between the target of interest (member m1) and other targets (members m2, m3, ...) in the set (group A).

[0045] Furthermore, the embedded representation reflecting such interactions can be regarded as an embedded representation reflecting the degree to which the set (group A) is suitable for the target (member m1) (in terms of its output (results)) as a result of the higher-level (C) training.

[0046] Therefore, according to the set suitability expression generation model 1, for example, it is possible to generate an embedded expression of the degree to which a company organization is suited (suitable) for a member of interest, which reflects the compatibility (degree of getting along, ease of working together, etc.) between a member of interest and other members belonging to the company organization. This also makes it possible to grasp how the compatibility between the member of interest and other members is related to the output (results) of the company organization, as will be described later, and to carry out suitable personnel allocation taking into account such compatibility between individual members.

[0047] In addition, even if the number of elements belonging to the set for which the expression is to be generated (the number of members M in the specific example of FIG. 1) is changed, the set suitability expression generation model 1 can handle it with the same weight coefficients and parameters without changing the basic configuration. In the schematic diagram of FIG. 1, the number of rectangles in the "gray rectangles" representing the "interaction data" group and the data group after a predetermined process can be appropriately set to handle the number of elements.

[0048] Furthermore, like known word distributed representations, the set suitability representation (vector) in which the suitability of the set for the target object is embedded can be expected to be used to, for example, use the distance between set suitability representations (vectors) relating to different targets of interest as the similarity in their positions in the set, or to perform calculations between set suitability representations (vectors) to derive or generate even more useful information regarding the set.

[0049] In any case, this set suitability expression is an embedded representation of the set, and can be information with high granularity about the set. Although this set suitability expression cannot be directly interpreted by humans, it is expected that it can be used as input data for machine learning models for various applications, and that artificial intelligence can understand the characteristics of the set and, for example, predict the performance of the entire set or personalize services for each member of the set.

[0050] In the specific example of FIG. 1, the "elements" belonging to the "set" are members (people), and the "characteristics" are personalities expressed by the Big Five scores, but the "set," "elements," and "characteristics" are of course not limited to these. For example, the job skills of the members (people) expressed by appropriately set scores may be adopted as the "characteristics." Furthermore, it is also possible to adopt Schwartz's Values ​​proposed in the non-patent document: Shalom H. Schwartz, "A proposal for measuring value orientations across nations." Questionnaire Package of the European Social Survey, 259(290), 261. 2003. In addition, as the "set" to be generated as an expression, various things can be adopted, such as a group of animals, a group of devices used in interaction with each other, or even a group of machine learning models used in interaction with each other, as long as it is a collection of elements that can be expressed by some factor score and have characteristics that allow the good and bad combinations between them to be recognized.

[0051] [Functional configuration of the set suitability expression generation model] Here, a specific functional configuration of the set desirability expression generation model 1 of the present embodiment will be described. As shown in FIG. (A) a first embedded expression generation unit 11; (B) a second embedded expression generation unit 12; The above two functional components are implemented by executing a program installed in a computer. These two functional components will be described below.

[0052] Also in FIG. 1, the first embedded representation generation unit 11 generates a plurality of first embedded data Φ(x1), Φ(X2), ... by performing a dimensionality compression process Φ (e.g., a process of compressing 25 dimensions (channels) to 2 dimensions (channels)) on each of the interaction data x1, x2, ... generated for each combination of N elements (N=2 in FIG. 1, but N≧3 is also possible) that include the target object (member m1) as an element and are possible for the target object and other objects belonging to a set (group A), to generate a plurality of first embedded data Φ(x1), Φ(X2), ...

[0053] The model part that performs the dimensionality reduction process Φ can be configured with a neural network (NN, Neural Networks) algorithm such as DNN (Deep Neural Networks), and in this embodiment, it is pre-trained as the encoding part of an autoencoder. Incidentally, the interaction data used in this pre-training is not limited to the population that is grouped as described later, and can be generated in large quantities from personality data of two people (e.g., Big Five scores) that have already been collected in various fields.

[0054] After that, the set desirability expression generation model 1 including the first embedded representation generation unit 11 is connected to a data estimator (e.g., the evaluation estimator 81 in FIG. 5) that outputs evaluation data related to a predetermined output of the target (member m1), and is trained using learning data including interaction data and correct evaluation data. As a result, the model part that performs the dimensionality reduction process Φ is fine-tuned. By adopting a model part that has been pre-trained in this way, it is possible to construct a set desirability expression generation model 1 that outputs with high accuracy using less limited learning data.

[0055] As a modification, the model part that performs the dimensionality reduction process Φ may be connected to an estimator that receives the first embedding data as input and outputs data related to a predetermined output of the target, and may be pre-trained using training data including interaction data and ground truth data. Here, this training data can also be generated in large quantities from, for example, personality data of two people (e.g., Big Five scores) and data related to the compatibility of the two people (e.g., data on the track record of the pair or data on the degree of connection on SNS) that have already been collected in various fields.

[0056] Also in FIG. 1, the second embedded representation generating unit 12 of this embodiment performs a "pooling process" on a plurality of first embedded data Φ(x1), Φ(X2), ..., and further performs an "activation process" on the pooling process result to generate second embedded data. Here, in this embodiment, the "pooling process" is a process of determining the maximum value (max), minimum value (min), and average value (mean) of each element corresponding to each other in the plurality of first embedded data, i.e., for each channel, and generating vector data with these values ​​as vector elements. In fact, multiple "relationships between the target object and other objects" exist in parallel within a group (set), and the number of such relationships may change depending on the group size. The pooling process is a suitable process for extracting the characteristics of the set suitability from multiple "relationships between the target object and other objects" existing within such a group.

[0057] The above-mentioned "activation process" is a process for increasing the expressiveness of the vector data generated by the pooling process as an embedded representation and aligning the order with other input data. Specifically, the vector data is input to an activation function such as a sigmoid function, a tanh function, or a ReLU (Rectified Linear Unit) and the output is extracted. The "pooling process" is not limited to the above, and may include other statistical representative values ​​as elements, or may generate vector data (second embedded data) having at least one of these statistical representative values ​​as an element. In any case, the second embedded representation generating unit 12 can generate second embedded data having substitution invariance required for the representation (feature amount) of a set by performing such a pooling process.

[0058] The set desirability representation generation model 1 including the above-mentioned second embedded representation generation unit 12 is connected to a data estimator (e.g., the evaluation estimator 81 in FIG. 5) that receives the second embedded data as input and outputs evaluation data related to a predetermined output of the target (member m1), and is trained using learning data including the interaction data and the correct evaluation data. As a result, the second embedded data output by the second embedded representation generation unit 12 becomes a set desirability representation of the set (group A) for the target (member m1).

[0059] [Another embodiment of the set suitability expression generation model] FIG. 2 is a schematic diagram for explaining another embodiment of a set desirability expression generation model according to the present invention.

[0060] The set desirability expression generation model 2 of this embodiment shown in FIG. (D) a first embedded representation generation unit 211 and a second embedded representation generation unit 212 that receive the “first interaction data” as input; (E) A first embedded representation generation unit 221 and a second embedded representation generation unit 222 that receive the second interaction data as input; (F) Set desirability expression generating unit 23 In other words, the functional components make the computer function.

[0061] Of these, the two functional components in (D) above have the same structure (generated by the same algorithm) as the functional components of the same name in the set desirability representation generation model 1 (FIG. 1) and perform the same function. That is, the first embedded representation generation unit 211 receives the "first interaction data" (same as the interaction data in FIG. 1; N=2) of group A for the target of interest (member m1) in the specific example in FIG. 2, and generates and outputs the first embedded data, and the second embedded representation generation unit 212 receives the first embedded data and generates and outputs the second embedded data.

[0062] The two functional components in (E) above also have the same structure (generated by the same algorithm) as the functional components of the same name in the set suitability expression generation model 1 (FIG. 1) and perform the same function. However, unlike the first embedded expression generation unit 211 in (D) above, the first embedded expression generation unit 221 includes the target object (member m1) as an element, and generates an embedded expression for each combination of three elements that can be taken for the target object and other objects. M-1 C 2 The "second interaction data", i.e. x1 = (O1×O2×O3, O1×O2×C3, O1×O2×E3, ...): Interaction data between member m1, member m2, and member m3 x2 = (O1×O2×O4, O1×O2×C4, O1×O2×E4, ...): Interaction data between member m1, member m2, and member m4 x3 = (O1×O2×O5, O1×O2×C5, O1×O2×E5, ...): Interaction data between member m1, member m2, and member m5 ... ... The first embedding data reflecting the second interaction data is generated and output using the above as input.

[0063] 2 explicitly includes the above-mentioned two functional components (D) that receive the first interaction data related to a combination of two elements (two members including member m1) as an input, and the above-mentioned two functional components (E) that receive the second interaction data related to a combination of three elements (three members including member m1). However, the set desirability representation generation model 2 may further include a set of two functional components (a first embedding representation generation unit and a second embedding representation generation unit) that receive the (N'-1)th interaction data related to a combination of N' (N' is an integer of 4 or more) elements (N' members including the target of interest (member m1)) as an input, and generate and output the second embedding data (set desirability representation).

[0064] More generally, when a "pair" of two functional components (a first embedded representation generation unit and a second embedded representation generation unit) is defined, which receive as input the (N-1)th interaction data generated for each combination of N (N is an integer equal to or greater than 2) elements (N members including member m1) including the target object (member m1) and generate and output second embedded data (set desirability representation), it is also preferable to provide a plurality of "pairs" in which the values ​​of N are different between the pairs.

[0065] Also in Fig. 2, the set desirability expression generating unit 23 (F) combines the second embedding data (set desirability expressions) generated and output for each "group" (two "groups" in Fig. 2) and concatenates them in this embodiment to generate new second embedding data, that is, a new set desirability expression for group A for member m1 (target). The set desirability expression generated here is not limited to the first interaction between the target (member m1) and one other member in group A, but is a more preferable expression of set desirability that reflects the (N-1)th interaction (N=3 in Fig. 2) generated for each combination of N (N is an integer equal to or greater than 2) members including the target (member m1).

[0066] For example, in a case where chatting and discussion spaces are appropriately formed among 3 to 5 unspecified members including member m1 (target of interest) in group A and tend to contribute to the activity results of member m1 (target of interest) in group A, it may be particularly preferable to generate a set desirability expression that reflects multiple interactions including the second to fourth interactions (N=3, 4, 5).

[0067] The set desirability representation generation model 2 described above is also connected to a data estimator (e.g., the evaluation estimator 81 in FIG. 5) that receives the second embedded data output from the set desirability representation generation unit 23 and outputs evaluation data related to a predetermined output of the target of interest (member m1), and is trained using learning data including the first interaction data, the second interaction data, and the correct evaluation data. As a result, the second embedded data output from the set desirability representation generation unit 23 becomes the set desirability representation of the set (group A) for the target of interest (member m1).

[0068] [Yet another embodiment of the set desirability expression generation model] FIG. 3 is a schematic diagram for explaining still another embodiment of the set desirability expression generation model according to the present invention.

[0069] The set desirability expression generation model 3 of this embodiment shown in FIG. (G) A first embedded representation generation unit 311 and a second embedded representation generation unit 312 that receive the “interest target interaction data” as input; (H) A third embedded representation generation unit 321 and a fourth embedded representation generation unit 322 that receive “multiple-subject (member) interaction data” as input; (I) A set suitability expression generating unit 33 In other words, the functional components make the computer function.

[0070] Of these, the two functional components in (G) have the same structure (are generated by the same algorithm) as the functional components of the same name in the set desirability representation generation model 1 (FIG. 1) and perform the same function. That is, the first embedded representation generation unit 311 receives the "target interaction data" (same as the interaction data in FIG. 1) of group A for the target (member m1) in the specific example in FIG. 3, and generates and outputs the first embedded data, and the second embedded representation generation unit 312 receives the first embedded data and generates and outputs the second embedded data.

[0071] The two functional components in (H) also have the same structure (generated by the same algorithm) as the first embedded representation generation unit 11 and the second embedded representation generation unit 12 in the set desirability representation generation model 1 (FIG. 1) and perform the same function. However, unlike the first embedded representation generation unit 311 in (G), the third embedded representation generation unit 321 generates embedding expressions for each combination of two elements that can be taken for other objects (member m2, member m3, . . .) other than the target object (member m1). M-1 C 2 Individual "member interaction data", i.e. x1 = (O2×O3, O2×C3, O2×E3, ...): Interaction data between member m2 and member m3 x2 = (O2×O4, O2×C4, O2×E4, ...): Interaction data between member m2 and member m4 ... x1 = (O3×O4, O3×C4, O3×E4, ...): Interaction data between member m3 and member m4 x2 = (O3×O5, O3×C5, O3×E5, ...): Interaction data between member m3 and member m5 ... The fourth embedding representation generating unit 322 receives the third embedding data as input and generates and outputs the third embedding data reflecting the interaction data of other objects (members). Also, the fourth embedding representation generating unit 322 receives the third embedding data as input and generates and outputs the fourth embedding data (set desirability representation).

[0072] In Fig. 3, other-object (member) interaction data for combinations of two elements that other members can take is used, but of course, this is not limited to this. In general, the first embedding data can be generated using other-object (member) interaction data for combinations of N elements (N is an integer equal to or greater than 2) that other objects (member m2, member m3, ...) can take. As a modification, a plurality of "pairs" of the third embedding expression generation unit and the fourth embedding expression generation unit, in which the N values ​​are different between the pairs, may be provided.

[0073] Also in Fig. 3, the set desirability expression generating unit 33 of (I) combines the second embedding data (set desirability expression) generated and output by the second embedding expression generating unit 312 with the fourth embedding data (set desirability expression) generated and output for each "set" (one "set" in Fig. 3), and concatenates them in this embodiment to generate new second embedding data, that is, a new set desirability expression of group A for the target of interest (member m1). The set desirability expression generated here is a more favorable expression of set desirability that also reflects interactions between targets (members) other than the target of interest (member m1).

[0074] For example, it is quite conceivable that the activity results of member m1 (target) in group A may be influenced by the compatibility and relationships between other members. Specifically, if the surrounding members are working in good relationships, it is highly possible that the work of member m1 (target) will also progress due to their influence. Therefore, it is more preferable to generate a set suitability expression that also reflects the interactions between these other members.

[0075] The set desirability expression generation model 3 described above is also connected to a data estimator (e.g., the evaluation estimator 81 in FIG. 5) that receives the second embedded data output from the set desirability expression generation unit 33 and outputs evaluation data related to a predetermined output of the target of interest (member m1), and is trained using learning data including the target of interest interaction data, other target (member) interaction data, and correct evaluation data. As a result, the second embedded data output from the set desirability expression generation unit 33 becomes the set desirability expression of the set (group A) for the target of interest (member m1).

[0076] As another preferred embodiment, the set desirability expression generation model 3 described above is as follows: (a) A first embedding representation generation unit 221 and a second embedding representation generation unit 222 that receive the second interaction data in the set desirability representation generation model 2 (FIG. 2) as input. It is also preferred to further incorporate: (b) A plurality of "pairs" of the first embedded representation generation unit and the second embedded representation generation unit that receive the second interaction data as input, the "pairs" having different N values. In such an embodiment, the set desirability expression generating unit 33 can generate new second embedded data (set desirability expression) in which more detailed (more information-rich) interaction information is embedded by combining (linking) the second embedded data generated and output by the second embedded expression generating unit of (a) or (b) above.

[0077] <Generation of Set Desirability Expressions Using Attention> FIG. 4 is a schematic diagram for explaining a modified embodiment of the set desirability expression generation process in the set desirability expression generation model according to the present invention.

[0078] The set desirability expression generation model 3' shown in Fig. 4 is a model in which a set desirability expression generation unit 33' is adopted instead of the set desirability expression generation unit 33 (Fig. 3) in the set desirability expression generation model 3 described using Fig. 3. This set desirability expression generation unit 33' generates new second embedding data (set desirability expression) by weighting and mixing the second embedding data output from the second embedding expression generation unit 312 and the fourth embedding data output from the fourth embedding expression generation unit 322 using a known attention mechanism.

[0079] Here, the attention mechanism is a mechanism that determines which part of the input data should be paid special attention to, and is currently widely used in the field of machine learning, especially using DNNs. For more information on this attention mechanism, see, for example, non-patent literature: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, AN, Kaiser, L. ukasz, & Polosukhin, I., “Attention is All you Need”, Advances in Neural Information Processing Systems (Vol. 30), 2017,<https: / / proceedings.neurips.cc / paper / 2017 / file / 3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf> This is explained in detail in.

[0080] As also shown in FIG. 4, specifically, in this embodiment, the set desirability expression generating unit 33' functions as a first weighting mixer 331', a scaler 332', a softmax function unit 333' and a second weighting mixer 334'.

[0081] First, the first weighting mixer 331′ calculates (a) second embedding data (vector data) output from the second embedding representation generation unit 312; (b) The fourth embedding data (vector data) output from the fourth embedding representation generation unit 322; are input as Query and Key, respectively, and a function value vector related to the weighted sum of the two (additive attention function) or a function value vector related to the weighted product of the two (multi-plicative attention function) is output. Here, this weighting can be interpreted as indicating the importance of each element in the fourth embedding data (vector data) relative to the second embedding data.

[0082] Next, the scaler 332′ scales the output vector from the first weighting mixer 331′ as follows: (d k ) 0.5 (where d k is normalized by dividing it by the number of dimensions of the fourth embedded representation (key). After that, the softmax function unit 333' applies a known softmax function to maximize the maximum value of the vector elements in this normalized output vector, and outputs a vector whose elements sum to 1.

[0083] Finally, the second weighting mixer 334' outputs a function value vector related to the weighted sum of the output vector from the softmax function unit 333' and the fourth embedding data (vector data) input as Value, or a function value vector related to the weighted product of both. Here, this output vector is new second embedding data (set suitability expression) that takes into account the importance of each element in the fourth embedding data (vector data) to the second embedding data.

[0084] In this way, the set suitability expression generating unit 33' can generate a set suitability expression that is weighted more heavily on interactions between other objects that are particularly important in terms of the set feature (set suitability expression) that reflects interactions between the target object and other objects. For example, it is possible to generate a set suitability expression that more strongly reflects the compatibility and human relationships between other members that are likely to be particularly influenced by the target object member m1.

[0085] The set suitability expression generating unit 33' can be configured with a neural network including a DNN. The set suitability expression generating model 3' including the set suitability expression generating unit 33' is also connected to a data estimator (e.g., the evaluation estimator 81 in FIG. 5) that receives the second embedding data output from the set suitability expression generating unit 33' and outputs evaluation data related to a predetermined output of the target of interest (member m1), and is trained using learning data including the target of interest interaction data, other target (member) interaction data, and correct evaluation data. As a result, the second embedding data output by the set suitability expression generating unit 33' becomes a set suitability expression of the set (group A) for the target of interest (member m1).

[0086] [Intra-set target evaluation model] FIG. 5 is a schematic diagram for explaining an embodiment of a set-in-set target evaluation model that estimates an evaluation result of a target of interest using a set desirability expression output from a set desirability expression generation model according to the present invention.

[0087] The in-set target evaluation model 8 of this embodiment shown in FIG. (A) A set desirability expression (second embedded data) output from one of the set desirability expression generation models 1 (FIG. 1), 2 (FIG. 2), 3 (FIG. 3) and 3' (FIG. 4) described above, and the other forms and modes of these models described above; (a) Factor statistical features generated by normalizing the "All Member (All Target) Factor Statistical Data" and "Other Member (Other Target) Factor Statistical Data" described later; (c) Factor score data (five-dimensional vector for five factors) for member m1 (target) With the input, (D) Evaluation information regarding a specific output of a member m1 (target) in group A (set), which is referred to as an “evaluation score” in this embodiment. This model has a trained rating estimator 81 that outputs the following as a functional configuration unit (making a computer function as the rating estimator 81).

[0088] Here, the all-member factor statistical data in (a) above is statistical data on the factor scores of all members, generated by the "factor statistical data generating unit" 82 using factor score data of all members of group A (set) (e.g., Big Five score data of all M members belonging to group A) as input. Specifically, for example, Mean, variance, maximum, minimum, median, quartiles, and covariance for each factor score of Openness (O), Conscientiousness (C), Extraversion (E), Agreeableness (A), and Emotional Lability (N) For example, the statistical data of all member factors can be at least one of the following: mean value (5-dimensional vector for 5 factors), variance (5-dimensional vector), maximum value (5-dimensional vector), minimum value (5-dimensional vector), and covariance (10(= 5 C 2 )-dimensional vector) for a total of 30 dimensions.

[0089] The other member factor statistical data in (i) above is statistical data on the factor scores of other members, generated by a "factor statistical data generating unit" 82 using as input factor score data of other members other than member m1 (target) in group A (set) (for example, Big Five score data of (M-1) other members). Specifically, it can be vector data containing statistical values ​​similar to the all-member factor statistical data described above. For example, this other member factor statistical data also contains the mean (five-dimensional vector for five factors), variance (five-dimensional vector), maximum value (five-dimensional vector), minimum value (five-dimensional vector), and covariance (10(= 5 C 2 )-dimensional vector) for a total of 30 dimensions.

[0090] In this embodiment, the all member factor statistical data and the other member factor statistical data are normalized by the normalization unit 83 in order to align the order (for example, each vector element is converted to a value between 0 and 1) before being input to the evaluation estimator 81.

[0091] Furthermore, the "evaluation score" in (e) above may be, for example, the average score of regular tests of member m1 (target of interest) in the class if group A (set) is a school class, or, if group A is a company organization, it may be a value representing the work results of member m1 (target of interest) in the company organization, and various other things related to member m1 (target of interest) can be used as the "evaluation score."

[0092] Incidentally, the "evaluation score" adopted in a specific example (FIG. 7) of the grouping method according to the present invention described later is the discussion activity level of the target in group A (determined by a predetermined method) and the individual achievement evaluation score of the target set for a predetermined task in group A. Here, the evaluation estimator 81 (target evaluation model 8 within a set) outputs an estimate of the discussion activity level of the target, and further outputs an estimate of the individual achievement evaluation score of the target. Of course, it is also preferable to provide an individual evaluation estimator 81 for each evaluation score. Note that such an evaluation estimator 81 may be configured as a neural network, for example, a multi-layer perceptron (MLP).

[0093] As described above, the in-set target evaluation model 8 can output a more accurate estimation result of the target evaluation by incorporating the set suitability expression (second embedded data) which is an embedded expression with high information granularity regarding the set suitability (and further by incorporating the factor statistical feature amount and the factor score data of the target of interest). Incidentally, the evaluation estimator 81 (in-set target evaluation model 8) may be input with at least one of the all-target (all-member) factor statistical data, the other-target (other-member) factor statistical data, and the factor score data of the target of interest, and the set suitability expression (second embedded data) described above in (A). It is also possible to input only the set suitability expression (second embedded data).

[0094] The functional configuration of the in-set object evaluation model 8 that estimates the evaluation of the object in the set using the output of the set desirability expression generation model (1,2,3,3') has been described above. Here, training of various operators and parameters in the set desirability expression generation model (1,2,3,3') and the in-set object evaluation model 8 (evaluation estimator 81) can be performed, for example, as follows.

[0095] First, for each of a large number of groups (sets), a large number of teacher data are prepared, which are combinations of factor score data of all members, factor score data of other members (other targets) other than the member of interest (target of interest), interaction data including the member of interest as an element (and interaction data of other members), and the correct evaluation score of the member of interest. Next, the interaction data of the teacher data is input to the set suitability expression generation model (1,2,3,3'), and the factor score data of the teacher data is input to the factor statistical data generation unit 82. The difference between the estimated value of the evaluation score output from the target evaluation model 8 in the set and the correct evaluation score of the teacher data is treated as a loss, and various operators and parameters in each functional configuration unit are learned, for example, by a known backpropagation method.

[0096] Here, the first embedded representation generation unit (and the third embedded representation generation unit) of the set desirability representation generation model (1,2,3,3') is fine-tuned by the above training when it is pre-trained as the encoding unit of the autoencoder as described above. This makes it possible to construct the set desirability representation generation model (1,2,3,3') and the intra-set target evaluation model 8 that output high accuracy with less limited learning data.

[0097] As a modified form of training, the set desirability expression generation model (1,2,3,3') can be placed in front of a separately prepared in-set target evaluation model that outputs estimated values ​​of other evaluation scores, and training can be performed using teacher data tailored to this system to train the set desirability expression generation model (1,2,3,3') and this other in-set target evaluation model. This trained (learned) set desirability expression generation model (1,2,3,3') can then be taken out and placed in front of the in-set target evaluation model 8, and the in-set target evaluation model 8 can then be trained using the large amount of teacher data mentioned above.

[0098] [Set desirability expression generating device, set desirability expression generating method, grouping device, grouping program, grouping method] Fig. 6 is a functional block diagram showing a functional configuration of an embodiment of a set desirability expression generating device according to the present invention, and Fig. 7 is a schematic diagram showing a specific example of a grouping method including a set desirability expression generating method according to the present invention.

[0099] The grouping device 9 shown in Figure 6 as one embodiment of the group suitability expression generating device according to the present invention has a communication interface unit 901, a user attribute information storage unit 902, a performance information storage unit 903, a group evaluation result storage unit 904, a grouping result storage unit 905, a keyboard (KB) 906, a display (DP) 907, and a processor memory.

[0100] Here, the processor memory stores an embodiment of a grouping program including a grouping degree expression generating program according to the present invention, and has a computer function, and executes the grouping program to perform a grouping degree expression generating process and a grouping process. Therefore, the device 9 may be a server, a cloud server, or a dedicated device for grouping, but it may also be, for example, a personal computer (PC), a notebook or tablet computer, or a smartphone equipped with the grouping program according to the present invention.

[0101] Further, the processor memory has a group candidate setting unit 911, a factor score and interaction data generating unit 912, a model constructing unit 913, a group candidate evaluating unit 914, a grouping implementation unit 915, a communication control unit 921, and an input / output control unit 922. These functional components can be regarded as functions of a grouping program stored in the processor memory. Among them, the factor score and interaction data generating unit 912 and a part of the group candidate evaluating unit 914 can also be regarded as functions of a set desirability expression generating program included in the grouping program. Furthermore, the process flow shown by connecting the functional components of the grouping device 9 in FIG. 6 with arrows can also be understood as one embodiment of a grouping method including a set desirability expression generating method according to the present invention.

[0102] Also in the functional block diagram of FIG. 6, the user attribute information storage unit 902 stores and manages user attribute information including Big Five score data for each of a large number of users and attributes such as gender, age, and affiliation, which are acquired by communication from an external user attribute information database (DB) 71 via a communication interface unit 901 and a communication control unit 921. Here, the Big Five score data (personality information) of each user acquired from the user attribute information DB 71 may be data collected by, for example, a questionnaire survey, or may be data estimated using a known technology such as the technology disclosed in Patent Document: JP 2019-197388 A, which includes the first inventor of the present application as an inventor.

[0103] In addition, the performance information storage unit 903 stores and manages evaluation score information relating to evaluation items for each member belonging to each of the multiple groups including group A, which is obtained by communication from an external performance information DB72 via the communication interface unit 901 and the communication control unit 921, in this embodiment, discussion activity level, individual performance evaluation score, etc.

[0104] 6, a group candidate setting unit 911 sets a plurality of group candidates having users included in a group of users to be grouped, which is retrieved from the user attribute information storage unit 902. Here, the user group corresponds to groups to be grouped in various fields, such as a group of students to be classified into classes in the field of education, or a group of employees to be assigned to various companies.

[0105] Specifically, in the specific example shown in FIG. 7, the group candidate setting unit 911 selects group candidates from the 11 captured users u1 to u11 according to the set member number condition: "the number of members is 3 or more and 4 or less." Group candidate 01 = {u1, u2, u3} Group candidate 02={u1, u2, u4} ... Here, u1, u2,... are the identifiers (IDs) of each user, and are managed in association with the group candidate IDs (01, 02,...). For example, if there are 20 users to be grouped, and processing is performed according to the member count condition: "the number of members is a minimum of 3 and a maximum of 4", a total of 5985 group candidates will be generated.

[0106] Returning to the functional block diagram of FIG. 6, the factor score and interaction data generation unit 912 uses the Big Five score data of each user imported from the user attribute information storage unit 902 to generate factor score data (as shown in FIG. 7) for each group candidate, interaction data including the target of interest as an element (each member being the target of interest), and interaction data of other members other than the target of interest.

[0107] Here, when generating these interaction data, it is preferable to exclude interaction terms that are judged to have little effect on the characteristics of the group candidates that are the set for which expressions are to be generated, rather than taking all of the interaction terms (25 in the case of the Big Five score). In fact, if the number of interaction terms, that is, the number of dimensions of the interaction data, increases, the computational cost of the training process of the set suitability expression generation model to be used later increases. Therefore, by selecting the interaction terms and suppressing the number of dimensions of the interaction data to the necessary level, it is possible to reduce the burden of the model construction process.

[0108] Specifically, the factor score and interaction data generating unit 912 first calculates, for each group candidate, which is a set of expression generation targets, the average X_i of each factor score in the group candidate and the average X_ij (i, j=O, C, E, A, N) of the interaction terms of the factor scores. For example, Openness (O) average: X_O = (O1 + O2 + ) / (number of members in the group candidate) Mean of the interaction term between Openness (O) and Conscientiousness (C): X_OC=(O1×C2+O1×C3+···) / (number of combinations of O and C among members) Calculate.

[0109] Next, the factor score and interaction data generating unit 912 calculates a regression equation with k_i, K_ij, and b as coefficients: (1) Σ i (k_i×X_i)+Σ i,j (k_ij × X_ij) + b = (evaluation score (level of discussion activity, etc.)) A multiple regression analysis is performed using and a p-value is calculated for each interaction term (each combination of factor scores).

[0110] Here, the p-value of an interaction is a value that indicates the significance of the interaction, and specifically, it is a value that corresponds to the probability that a difference in evaluation scores will occur under the assumption that the null hypothesis that there is no difference in evaluation scores due to the interaction is correct. Also, in the field of statistical analysis, it is customary that the null hypothesis is rejected if the p-value is 0.05 (5%) or less (as the significance level). Therefore, if the p-value is 0.05 (5%) or less, the corresponding interaction term can be determined to be "significant."

[0111] The factor score and interaction data generating unit 912 excludes from the interaction data any interaction term whose p-value calculated in this way exceeds a predetermined threshold (for example, 0.05 according to the above-mentioned convention). In this case, for example, if the p-value of the interaction term between openness (O) and conscientiousness (C) exceeds a predetermined threshold, the interaction data will be data that does not include the interaction term: O × C. In other words, the factor score and interaction data generating unit 912 generates interaction data that does not include any interaction term whose p-value, which is a value indicating the significance of the interaction between factor scores, exceeds a predetermined threshold.

[0112] The calculation of the p-value explained above is explained in detail in, for example, non-patent literature: "Introduction to Statistics" compiled by the Statistics Department of the College of Arts and Sciences, University of Tokyo, University of Tokyo Press, 30th edition, pp. 175-192, 2010, and non-patent literature: "Linear Regression Analysis" by Chihiro Minoya, Asakura Publishing, 2015. The above multiple regression analysis can be performed, for example, in the published Stats models: <https: / / www.statsmodels.org / stable / generated / statsmodels.regression.linear_model.OLS.html?highlight=ols#statsmodels.regression.linear_model.OLS> It is possible to carry out the method using the following formula:

[0113] 6, the model construction unit 913 generates teacher data for model training from the Big Five score data of each user received from the user attribute information storage unit 902 and the evaluation score information received from the performance information storage unit 903, and performs training processing of any one of the set suitability expression generation models 1, 2, 3, and 3' and the corresponding in-set target evaluation model 8. The training processing performed here can be the processing content already described.

[0114] Similarly, in the functional block diagram of FIG. 6, the group candidate evaluation unit 914 (a) Using the set desirability expression generation model trained and constructed by the model construction unit 913, generate a set desirability expression of each member in each group candidate, and then (b) The generated group suitability representation of each member in each group candidate is input to the in-set target evaluation model 8 trained and constructed by the model construction unit 913, and from the output of this in-set target evaluation model 8, an estimate of each evaluation score for each member of each group candidate is determined, thereby determining the group evaluation score estimation result for each group candidate.

[0115] Specifically, in the example shown in FIG. 7, the group candidate evaluation unit 914 estimates the group evaluation score as follows: Group candidate 01 = [0.9, 0.6, ...] Group candidate 02 = [0.2, 0.4, ...] ... (The first number in [] is the average of the estimated "discussion activity" for each member, the second number is the average of the estimated "individual performance evaluation score" for each member, ...) In this embodiment, the group evaluation score estimation results for each determined group candidate are appropriately stored and managed in the group evaluation result storage unit 904 (FIG. 6).

[0116] Returning to the functional block diagram of Figure 6, the grouping implementation unit 915 uses the group evaluation score estimation results of each group candidate imported from the group evaluation result storage unit 904, and selects multiple of the multiple set group candidates based on the "grouping (constraint) conditions" including the condition that "the sum or average of the estimated values ​​of the evaluation scores for each group candidate becomes large enough to satisfy a specified condition," and sets the grouping results for the user group to be grouped.

[0117] In this embodiment, the grouping implementation unit 915 selects multiple group candidates as grouping results, for example, by using a known greedy search method such as those shown in (a) to (c) below, under the "grouping constraint condition" of maximizing an "index," which is, for example, the average value across members of the sum of the discussion activity level and individual performance evaluation scores, i.e., the average value of multiple group evaluation scores. (A) Each group candidate is sorted by "index," and a number of group candidates that fall within the top s% (s is a preset value, for example, 20) are extracted. (a) From the extracted group candidates, generate “combinations” of all group candidates that satisfy the following conditions 1 and 2. Condition 1: Among multiple group candidates related to the “combination”, there is no overlap in the users belonging to the group candidates. Condition 2: All users included in the user group (subject to grouping) must appear in all of the multiple group candidates related to the “combination”. (c) Among the generated “combinations,” the “combination” for which the average value of the “index” for the multiple group candidates related to that combination is the highest is output as the optimal grouping result.

[0118] In addition, if other conditions than those mentioned above are set as the "grouping constraint condition", such as "each group after grouping must include both females and males" or "each group after grouping must include users under the age of 30", the grouping implementation unit 915 generates "combinations" of all group candidates that also satisfy these conditions in (i) above.

[0119] As a preferred modification, it is also preferable that a priority order is added to each condition included in the "grouping constraint condition." In this case, the grouping implementation unit 915 initially searches for a "combination" of group candidates that satisfies all the conditions included in the "grouping constraint condition." If a "combination" that satisfies all the conditions is not obtained, the grouping implementation unit 915 may remove conditions with low priority orders from the "grouping constraint condition" and search for "combinations" of group candidates again, repeatedly adjusting the conditions according to priority order until a "combination" of group candidates that satisfies the "grouping constraint condition" is obtained.

[0120] In the specific example of grouping processing shown in FIG. 7, the grouping implementation unit 915 receives an estimation result of the group evaluation score for each group candidate (each of group candidates 01, 02, ...) from the group candidate evaluation unit 914, and finally uses this to output group candidate combination 123 (combination with combination ID = 123) as the optimal grouping result.

[0121] This combination 123 is: Group candidate 12 = {u1, u5, u7, u11}, Group candidate 34 = {u2, u4, u6, u9}, and Group candidate 09={u3, u8, u10} That is, the grouping implementation unit 915 has grouped the user group (users u1, u2, . . . , u11) into the above three optimal groups (candidates), in other words, assigned each user of the user group to the above three optimal groups (candidates).

[0122] Returning to the functional block diagram of Fig. 6, the grouping results generated and output by the grouping implementation unit 915 may be appropriately stored and managed in the grouping result storage unit 905, and may be displayed on the display 907 via the input / output control unit 922 as appropriate or in accordance with a user's instruction via the keyboard 906. Furthermore, it is also preferable that the results be transmitted via the communication control unit 921 and the communication interface unit 901 to an external information processing device, for example, a student management server owned by an English conversation school, where they may be used for various purposes, for example, for organizing new classes.

[0123] As described above in detail, the set suitability representation generation model of the present invention receives as input interaction data generated for each combination including the target object as an element, and can generate an embedded representation (set suitability representation) that reflects interactions of a predetermined characteristic between the target object and other objects in the set. In addition, according to the present invention, the generated set suitability representation can be used to estimate the evaluation of the target object in the set taking into account the interaction of the characteristic. Furthermore, using the evaluation estimation result, it is also possible to perform grouping taking into account the interaction of the characteristic.

[0124] This also makes it possible, for example, to make optimal personnel allocations that take into account the compatibility between personnel in various fields, and further, for example, when employees provide services to customers, it becomes possible to provide more effective or satisfying services by adjusting the type of group temporarily formed by employees and customers who have certain attributes.

[0125] In addition, by applying the present invention, for example, it is possible to realize class formation in educational settings that takes into account the compatibility between teachers and students and encourages students to improve their academic ability. In other words, the present invention can contribute to Goal 4 of the United Nations' Sustainable Development Goals (SDGs) "Ensure inclusive and equitable quality education and promote lifelong learning opportunities for all."

[0126] Furthermore, in order to provide adults with decent work (jobs that are rewarding and humane), high-quality jobs that are also good in terms of human relations, it is possible to employ working teams that take into consideration the compatibility between members, as configured by the present invention. In other words, the present invention makes it possible to contribute to Goal 8 of the United Nations-led SDGs, "Promote inclusive and sustainable economic growth, employment and decent work for all."

[0127] Furthermore, in order to provide a consumer with a sustainable consumption environment that is in line with the tastes and consumption trends of the consumer, the present invention can estimate and evaluate the consumption activities (product purchase results) of a group temporarily formed (at a certain sales venue) by a large number of consumers, each with a specific taste and consumption trend, and a group of products, each with various characteristics, and appropriately adjust the contents of the group of products to be provided and the consumers to whom they are provided. In other words, the present invention can also contribute to Goal 12 of the SDGs led by the United Nations, "Ensure sustainable consumption and production patterns."

[0128] Regarding the various embodiments of the present invention described above, various changes, modifications, and omissions within the scope of the technical idea and perspective of the present invention can be easily made by those skilled in the art. The above description is merely an example and is not intended to be restrictive in any way. The present invention is limited only by the scope of the claims and their equivalents. [Explanation of symbols]

[0129] 1, 2, 3, 3' Set Desirability Representation Generation Model 11, 211, 221, 311 First embedded representation generation unit 12, 212, 222, 312 Second embedded representation generation unit 23, 33, 33' Set preference expression generator 321 Third embedded representation generation unit 322 Fourth embedded representation generation unit 331' First weighting mixer 332' Scale section 333' Softmax function part 334' Second weighting mixer 71 User attribute information database (DB) 72 Performance Information DB 8. In-set target evaluation model 81 Rating Estimator 82 Factor statistical data generation section 83 Normalization section 9. Grouping device (set suitability expression generation device) 901 Communication interface section 902 User attribute information storage unit 903 Performance Information Storage Unit 904 Group Evaluation Result Storage Unit 905 Grouping result storage unit 906 Keyboard(KB) 907 Display (DP) 911 Group candidate setting section 912 Factor score and interaction data generator 913 Model Construction Department 914 Group Candidate Evaluation Department 915 Grouping Division 921 Communication Control Unit 922 Input / Output Control Unit

Claims

1. A set suitability expression generation model that causes a computer to function to generate a set suitability expression, which is an embedded expression of a degree to which a set including a plurality of objects having a predetermined characteristic represented by at least one factor score is suitable for a target object belonging to the set, a first embedding representation generation unit that generates a plurality of first embedding data by performing a dimensionality reduction process on each of interaction data generated for each combination of N (N is an integer equal to or greater than 1) elements that include the target of interest as an element and are possible for the target of interest and other targets belonging to the set, the interaction data including interaction terms for the factor scores related to the target of interest and the other targets included in the combination; a second embedding representation generation unit that performs a pooling process on the plurality of first embedding data to generate second embedding data; to make the computer function, The first embedded representation generation unit and the second embedded representation generation unit, which are functional means realized by a computer by the set suitability representation generation model, are trained in such a way that the second embedded representation generation unit is connected to a data estimator that receives the second embedded data as an input and outputs evaluation data related to a predetermined output of a target of interest, and when the first embedded representation generation unit receives the interaction data as learning data as an input, the evaluation data output from the data estimator is compared with the evaluation data of the correct answer as learning data, and the second embedded data is the set suitability representation for the target of interest. A set desirability expression generation model characterized by:

2. causing a computer to function as a plurality of pairs of the first embedded-representation generation unit and the second embedded-representation generation unit, the plurality of pairs having different values ​​of N between the pairs; The computer further functions as a set desirability expression generating unit that combines the second embedding data generated for each set to generate new second embedding data; The new second embedding data is the set preference representation for the target object through the training. The set desirability expression generation model according to claim 1 .

3. a third embedded representation generation unit that performs a dimensionality reduction process on each of interaction data generated for each possible combination of N elements for the other objects belonging to the set, the interaction data including an interaction term for the factor score for the other objects included in the combination, to generate a plurality of third embedded data; a fourth embedding representation generation unit that performs a pooling process on the plurality of third embedding data to generate fourth embedding data; a set desirability expression generating unit that combines the second embedding data and the fourth embedding data to generate new second embedding data; to make the computer function better, The new second embedding data is the set preference representation for the target object through the training.

3. The set desirability expression generation model according to claim 1 or 2.

4. The set desirability expression generation model according to claim 3, characterized in that the set desirability expression generation unit uses an attention mechanism to generate the new second embedding data by weighting and mixing the second embedding data and the fourth embedding data.

5. causing a computer to function as a plurality of pairs of the third embedded-representation generation unit and the fourth embedded-representation generation unit, the plurality of pairs having different values ​​of N between the pairs; The computer further functions as a set desirability expression generating unit that generates new second embedding data by combining the second embedding data and the fourth embedding data generated for each set; The new second embedding data is the set preference representation for the target object through the training. The set desirability expression generation model according to claim 3 .

6. The set suitability expression generation model according to any one of claims 1 to 5, characterized in that the interaction data is data that does not include an interaction term that is determined in advance to have a low significance of the interaction between the factor scores.

7. The set suitability representation generation model according to any one of claims 1 to 6, characterized in that the first embedded representation generation unit is pre-trained as an encoding unit of an autoencoder, or is connected to an estimator that receives the first embedded data as input and outputs data related to a predetermined output, and is pre-trained using training data including the interaction data and ground truth data.

8. A set-in-set object evaluation model, characterized in that it causes a computer to function as a trained evaluation estimator that takes as input the set desirability representation of an object of interest output from the second embedded representation generation unit, which is a functional means realized in a computer by the set desirability representation generation model described in any one of claims 1 to 7, and outputs evaluation information related to a specified output of the object of interest.

9. The in-set target evaluation model according to claim 8, characterized in that the evaluation estimator receives as input at least one of factor score data including the factor score of the target of interest, factor statistical data including statistical values ​​of the set of factor scores related to the target of interest and the other targets, and / or other target factor statistical data including statistical values ​​of the set of factor scores related to the other targets, and outputs evaluation information related to a specified output of the target of interest.

10. a group candidate setting means for setting, from a predetermined object group, a plurality of group candidates whose members are objects included in the object group; a group candidate evaluation means for generating, for each group candidate, a set desirability expression of each member of the group candidate as a focus object, using second embedded data output from the second embedded expression generation unit, which is a functional means realized by a computer using the set desirability expression generation model according to any one of claims 1 to 7, inputting the set desirability expression to a set desirability expression input unit, which is a functional means realized by a computer using the intra-set target evaluation model according to claim 8 or 9, and using evaluation information output from an evaluation information output unit, which is a functional means realized by a computer using the intra-set target evaluation model, to determine an estimate of evaluation information for each member of the group candidate; a grouping implementation means for selecting a plurality of the set group candidates based on a grouping condition including a condition that an estimated value of evaluation information for each member in each group candidate or a statistical value of the estimated value satisfies a predetermined criterion, and setting the selected plurality of group candidates as a grouping result for the predetermined target group; A grouping program for causing a computer to function as follows.

11. A step of setting a plurality of group candidates from a predetermined target group, the group candidates having targets included in the target group as members; a step of generating, for each group candidate, a set desirability expression of each member of the group candidate as a focus object, using second embedded data output from the second embedded expression generation unit, which is a functional means realized by a computer using the set desirability expression generation model according to any one of claims 1 to 7, inputting the set desirability expression to a set desirability expression input unit, which is a functional means realized by a computer using the intra-set target evaluation model according to claim 8 or 9, and determining an estimate of evaluation information for each member of the group candidate, using evaluation information output from an evaluation information output unit, which is a functional means realized by a computer using the intra-set target evaluation model; a step of selecting a plurality of the set group candidates based on a grouping condition including a condition that an estimated value of evaluation information for each member in each group candidate or a statistical value of the estimated value satisfies a predetermined criterion, and setting the selected plurality of group candidates as a grouping result for the predetermined target group; 13. A computer-implemented grouping method comprising:

Images