JP7485867B2 - PROGRAM, INFORMATION PROCESSING APPARATUS, AND INFORMATION PROCESSING METHOD - Google Patents

Description

The present invention relates to a program, an information processing device, and an information processing method.

Technology has been proposed that analyzes the state of a group and its members and gives advice on actions to improve that state (see, for example, Patent Document 1).

International Publication No. 2011/125272

The above technology uses questionnaires to obtain indicators that subjectively and discretely express the characteristics of members that serve as the basis for selecting advice, and has the problem of not being able to objectively express the situations in which each member finds themselves.

The present invention was made in consideration of these points, and aims to provide advice based on quantitative information to help members of the organization to achieve their numerical goals.

The first aspect of the present invention is a program. This program causes a computer to realize the following functions: acquiring a learning model trained by machine learning using a neural network so as to receive a feature value indicating the characteristics of an organization and output activity data, which is data indicating the magnitude of the organization's activity; accepting an input of a target value for the activity data; calculating a first feature value, which is a feature value input to the learning model so that the learning model outputs the target value; acquiring a difference feature value, which is the difference between the first feature value and a second feature value indicating the current characteristics of the organization; and outputting a character string corresponding to the difference feature value to a display unit.

The program may further cause the computer to perform a function of receiving an adjustment amount for adjusting the value of the differential feature amount, and a function of adjusting the differential feature amount based on the adjustment amount to calculate an adjusted feature amount, and the output function may output a character string corresponding to the adjusted feature amount.

The program may further cause the computer to realize a function of inputting the first feature, which has been modified so that the difference between the first feature and the second feature becomes the adjusted feature, into the learning model to calculate adjusted activity data, and the output function may output the adjusted activity data.

The function of calculating the first feature may include a function of setting an initial value of the feature to be input to the learning model, a function of updating the feature until the difference between the value of the activity data, which is the output when the feature is input to the learning model, and the target value becomes less than a predetermined threshold, and a function of outputting the feature updated until the difference from the target value becomes less than the predetermined threshold as the first feature.

The program may further cause the computer to implement a function for accepting settings of constraint conditions related to the feature, and the updating function may update the feature to a degree that satisfies the constraint conditions.

The feature may be information about a user in a social networking service (SNS), and the output function may transmit the character string to the SNS.

The second aspect of the present invention is also a program. This program causes a computer to realize the following functions: a function of acquiring a second learning model that is a learning model trained based on a plurality of combinations of the second feature amount and the difference feature amount calculated by the computer that executes the above-mentioned program and a third feature amount that indicates the feature amount of the organization after the character string is presented, the second feature amount and the difference feature amount are input, and the second learning model is trained to output a value that reflects the deviation between the third feature amount and the first feature amount; a function of generating an evaluation function related to the feature amount that indicates the feature amount of the organization from the learning model and the second learning model; a function of calculating a fourth feature amount that is a feature amount that is an extreme value of the evaluation function; and a function of outputting a character string corresponding to the difference between the fourth feature amount and the second feature amount to a display unit.

The evaluation function may output a larger value the larger the output value of the learning model, and may output a larger value the smaller the output value of the second learning model.

A third aspect of the present invention is an information processing device. This device includes a model acquisition unit that acquires a learning model that has been machine-learned by a neural network so as to receive a feature value indicating the characteristics of an organization and output activity data that is data indicating the magnitude of the organization's activity, a goal acceptance unit that accepts an input of a target value for the activity data, a model inverse calculation unit that calculates a first feature value that is a feature value input to the learning model so that the learning model outputs the target value, a difference acquisition unit that acquires a difference feature value that is the difference between the first feature value and a second feature value indicating the current characteristics of the organization, and an output unit that outputs a character string corresponding to the difference feature value to a display unit.

A fourth aspect of the present invention is an information processing method. In this method, a processor executes the steps of: acquiring a learning model trained by machine learning using a neural network so as to input a feature quantity indicating the characteristics of an organization and output activity data, which is data indicating the magnitude of the organization's activity; accepting an input of a target value for the activity data; calculating a first feature quantity, which is a feature quantity to be input to the learning model so that the learning model outputs the target value; acquiring a difference feature quantity, which is the difference between the first feature quantity and a second feature quantity indicating the current characteristics of the organization; and outputting a character string corresponding to the difference feature quantity to a display unit.

The fifth aspect of the present invention is also a program. This program causes a computer to realize the following functions: acquiring a learning model that has been machine-learned by a neural network so as to receive a feature value indicating the characteristics of an organization and output activity data that is data indicating the magnitude of the organization's activities; acquiring a differential feature value that is the difference between a predetermined first feature value indicating the characteristics targeted by the organization and a second feature value indicating the actual characteristics of the organization; acquiring a second learning model that is a learning model trained based on a plurality of combinations of the second feature value, the differential feature value, and a third feature value indicating the characteristics of the organization after members of the organization act so that the characteristics of the organization become the first feature value, and that is trained to receive the second feature value and the differential feature value and output a value that reflects the deviation between the third feature value and the first feature value; generating an evaluation function related to the feature value indicating the characteristics of the organization from the learning model and the second learning model; calculating a fourth feature value that is the feature value that is the extreme value of the evaluation function; acquiring a differential feature value that is the difference between the fourth feature value and the second feature value; and outputting a character string corresponding to the differential feature value to a display unit.

To provide the above program or to update a part of the program, a computer-readable recording medium having the above program recorded thereon may be provided, or the above program may be transmitted over a communication line.

In addition, any combination of the above components, and any conversion of the present invention between methods, devices, systems, computer programs, data structures, recording media, etc., are also valid aspects of the present invention.

The present invention makes it possible to provide advice based on quantitative information to help members achieve the numerical goals of the organization to which they belong.

FIG. 2 is a diagram for explaining a learning model according to an embodiment. 11 is a diagram for explaining calculation of an input of a learning model and a difference in a feature amount. FIG. FIG. 2 is a diagram illustrating a functional configuration of an information processing device according to an embodiment. 10 is a schematic diagram for explaining a correspondence relationship between a difference feature amount and a character string. FIG. FIG. 2 is a diagram illustrating a functional configuration of a model inverse calculation unit according to an embodiment. 11 is a diagram for explaining a reverse calculation process executed by a model reverse calculation unit according to an embodiment. FIG. 4 is a flowchart illustrating a flow of information processing executed by an information processing device according to an embodiment. FIG. 11 is a diagram showing, in table form, a list of feature amounts handled by an information processing device according to a second modified example.

<Overview of the embodiment>
The information processing device according to the embodiment holds a learning model that has been machine-learned by a neural network so as to output activity data that indicates the magnitude of the activity of an organization when a feature quantity that indicates the characteristics of the organization is input.

Here, the "feature that indicates the characteristics of an organization" can be any quantity that indicates the characteristics of the members that make up the organization or the characteristics of how the members are connected to each other. One example is an adjacency matrix that represents a human network, or a feature that is assigned to each node and edge that makes up the human network.

Furthermore, the "activity data indicating the magnitude of an organization's activity" may be any quantity that clearly indicates whether the activity of the organization is active or not. Examples include the total number of emails exchanged between members of the organization, the total number of interactions of members on a social networking site if the members of the organization participate in that site, or the total number of logins per unit period, and, in the case of a for-profit organization, the sales amount and profit margin of the organization as a whole per unit period.

When an information processing device according to an embodiment accepts any activity data, it calculates the feature amounts to be input to a learning model in order to output the activity data. By comparing the calculated feature amounts with feature amounts indicating the current state of the organization, the information processing device according to an embodiment outputs advice on how the members of the organization and the connections between the members should be so that the organization's activity data becomes the accepted activity data.

FIGS. 1(a)-(c) are diagrams for explaining a learning model according to an embodiment. FIG. 2 is a diagram for explaining the calculation of the input of the learning model and the difference in the feature amount. Below, the learning model will be explained with reference to FIG. 1, and then an overview of the processing executed by the information processing according to the embodiment will be explained with reference to FIG. 2.

Figure 1(a) is a schematic diagram showing the members that make up an organization. For ease of explanation, it is assumed that the members of the organization are three people, member α, member β, and member γ, as shown in Figure 1(a), but the number of members of the organization may be more than three.

Figure 1(b) is a schematic diagram showing a weighted directed graph showing the human network of the three members shown in Figure 1(a) and activity data of an organization composed of the three members. The example shown in Figure 1(b) shows interactions in an SNS operated within the organization to which members α, β, and γ belong.

Specifically, in Figure 1(b), the circled α, β, and γ represent members α, β, and γ, respectively. Furthermore, the arrows connecting members indicate the number of interactions per unit period (e.g., one month) that took place from the member at the start point of the arrow to the member at the end point of the arrow. For instance, the example shown in Figure 1(b) indicates that member α had p interactions with member β per unit period. On the other hand, it indicates that there was no interaction from member β to member α during the unit period.

In addition, in FIG. 1(b), the letters next to each circle are feature quantities that indicate the characteristics of each member. In the example shown in FIG. 1(b), the feature quantities of each member indicate the average working hours of each member per day. For example, it is shown that the average working hours of member β is u hours per day. Furthermore, in FIG. 1(b), the activity data indicates the sales amount per unit period in the organization composed of members α, β, and γ. In the example shown in FIG. 1(b), it is shown that the sales amount of the organization was A yen in one month.

Figure 1(c) is a schematic diagram for explaining the generation of a learning model. Specifically, the learning model shown in Figure 1(c) is generated by machine learning using a neural network based on training data consisting of multiple pairs, each pair being a feature amount of an organization consisting of three members and activity data in the organization. As is known, a human network can be represented by an adjacency matrix. For example, the human network shown in Figure 1(b) is represented by a 3-row, 3-column matrix shown in the following formula (1).

In the adjacency matrix shown in formula (1), each row represents the start point of an arrow in the human network, each column represents the end point of an arrow in the human network, and the values represent the number of interactions per unit period. For example, row 1, column 2 represents the number of interactions (p times) from member α to member β. Since there are no interactions between the same members, the diagonal elements in the adjacency matrix are 0. In addition, the characteristics of each member are represented by the components of a vector in row 1 and column 3 shown in formula (2) below. The elements in each column correspond to the features of each member. For example, the first component in formula (2) represents the characteristics of member α, and indicates that the average working hours of member α are t hours per day.

The information processing device according to the embodiment holds a learning model that has been trained in advance to output activity data of an organization when a feature indicating the characteristics of the organization is input based on a large amount of training data. Here, the feature input to the learning model is x, and the activity data output by the learning model is y. In the example shown in FIG. 1(c), the feature x is information combining a 3-row, 3-column adjacency matrix and a 1-row, 3-column vector, and is information having 12 components. On the other hand, the activity data y is a scalar value. Note that if the learning model is trained to output multiple different activity data (e.g., monthly sales, monthly login count, total number of interactions, etc.), the activity data y is information having multiple components.

The learning model takes feature value x as input and outputs activity data y, so it can be expressed as y = f(x), where f is a multivariate function. In the example shown in FIG. 1(c), the function f is a function with 12 components as arguments. The fact that the learning model has been generated in advance by machine learning means that the function f(x) is fixed. Therefore, the information processing device according to the embodiment calculates the feature value x to be input to the function f in order to output the desired activity data y, by performing a reverse calculation using the function f that represents the learning model.

Figure 2(a) shows that the feature amount for setting the value of activity data y to the target value C is x'. This shows that if the feature amount of an organization becomes x', the activity data of that organization will become C. Also, Figure 2(b) shows the feature amount x of the organization's current state. Therefore, if the feature amount x of an organization can be set to feature amount x', the activity data of that organization can become C.

Figure 2(c) is a schematic diagram showing the difference x'-x between feature quantity x' and feature quantity x. The example shown in Figure 2(c) shows that, for example, increasing the number of interactions from member β to member α by l-c times is one of the conditions for making the organization's activity data C. Similarly, it shows that increasing member α's daily work hours by p-q hours is also one of the conditions for making the organization's activity data C. Note that when l-c or p-q is a negative value, it indicates that the number of interactions from member β to member α should be reduced by |l-c| times, and that member α's daily work hours should be reduced by |p-q| hours, respectively. These are concrete advice for making the organization's activity data the target value C.

The characteristics and activity data of an organization are quantitative and objective information. Furthermore, the learning model can be considered as an aggregate of statistical information of the entire organization. Therefore, the information processing device according to the embodiment can provide advice to members to achieve the numerical goals of the organization to which they belong, using quantitative and objective information and statistical information of the entire organization.

<Functional configuration of information processing device 1 according to embodiment>
FIG. 3 is a diagram showing a schematic functional configuration of an information processing device 1 according to an embodiment. The information processing device 1 includes a storage unit 2, a control unit 3, and a display unit 4. In FIG. 3, arrows indicate main data flows, and data flows not shown in FIG. 3 may exist. In FIG. 3, each functional block indicates a functional unit configuration, not a hardware (device) unit configuration. Therefore, the functional blocks shown in FIG. 3 may be implemented in a single device, or may be implemented separately in multiple devices. Data may be exchanged between functional blocks via any means, such as a data bus, a network, or a portable storage medium.

The storage unit 2 is a large-capacity storage device such as a ROM (Read Only Memory) that stores the BIOS (Basic Input Output System) of the computer that realizes the information processing device 1, a RAM (Random Access Memory) that serves as the working area of the information processing device 1, an HDD (Hard Disk Drive) or an SSD (Solid State Drive) that stores various information such as the OS (Operating System), application programs, and trained learning models that are referenced when the application programs are executed.

The control unit 3 is a processor such as a CPU (Central Processing Unit) or GPU (Graphics Processing Unit) of the information processing device 1, and functions as a model acquisition unit 30, a goal acceptance unit 31, a model inverse calculation unit 32, a difference acquisition unit 33, an output unit 34, an adjustment acceptance unit 35, an adjustment execution unit 36, a model application unit 37, and a constraint condition acceptance unit 38 by executing the programs stored in the memory unit 2.

Note that FIG. 3 shows an example in which the information processing device 1 is configured as a single device. However, the information processing device 1 may be realized by multiple processors, memory, and other computing resources, for example, as in a cloud computing system. In this case, each part constituting the control unit 3 is realized by at least one of multiple different processors executing a program.

The model acquisition unit 30 reads and acquires from the storage unit 2 a learning model that has been machine-learned by a neural network so as to input features that indicate the characteristics of the organization and output activity data that indicates the magnitude of the organization's activities.

The goal receiving unit 31 receives input of a goal value for activity data. Specifically, the goal receiving unit 31 receives input of a goal value from a user of the information processing device 1 via an input interface (not shown), such as a keyboard and mouse, for operating the information processing device 1. The model inverse calculation unit 32 calculates a first feature amount, which is a feature amount to be input to the learning model so that the learning model outputs a goal value. Details of the model inverse calculation unit 32 will be described later.

The difference acquisition unit 33 reads out and acquires a second feature amount indicating the current characteristics of the tissue from the storage unit 2. The difference acquisition unit 33 acquires a difference feature amount, which is the difference between the first feature amount and the second feature amount, by calculation. The output unit 34 outputs a character string corresponding to the difference feature amount to the display unit 4.

Figures 4(a) and 4(b) are schematic diagrams for explaining the correspondence between differential features and character strings. Specifically, Figure 4(a) is a diagram showing an example of differential features, and Figure 4(b) is a diagram showing a schematic data structure of a character string database for converting differential features into character strings. The character string database is managed by the output unit 34 and stored in the storage unit 2.

In FIG. 4(a), #N (N=1, ..., 9) indicates a real value that is an element of the differential feature amount calculated by the model inverse calculation unit 32. For example, #1 indicates some real value, not "1".

As shown in FIG. 4(b), the character string database stores a character string corresponding to #N. The output unit 34 replaces #N contained in the character string database with the actual value of #N and outputs it. For example, the character string corresponding to #1 is "Let's increase (decrease) α's actions toward β by #1 times." Here, if the actual value of #1 is "3," the output unit 34 outputs the character string "Let's increase α's actions toward β by 3 times" to the display unit 4. As another example, if the value of #9 is "-1.2," the output unit 34 outputs the character string "Let's decrease γ's overtime hours by 1.2 hours" to the display unit 4.

In this way, the information processing device 1 can provide advice to members on how to achieve the numerical goals of the organization to which they belong, using quantitative and objective information and statistical information on the entire organization.

Fig. 5 is a diagram showing a schematic functional configuration of the model inverse calculation unit 32 according to the embodiment. The model inverse calculation unit 32 includes an initial value setting unit 320, an update unit 321, and a first feature output unit 322. Fig. 6 is a diagram for explaining the inverse calculation process executed by the model inverse calculation unit 32 according to the embodiment. The inverse calculation process executed by the model inverse calculation unit 32 will be explained below with reference to Figs. 5 and 6.

As described above, when a feature quantity x is input to a learning model and activity data y is output, the relationship between them can be described as y = f(x). For convenience of illustration, the axis showing the feature quantity x in Figure 6 is one-dimensional, but x is usually multidimensional (12 dimensions in the example shown in Figure 1). In other words, the function f representing the learning model can be expressed as a hypersurface in a multidimensional space.

Now, assume that the target value of the activity data accepted by the goal accepting unit 31 is C. The model inverse calculation unit 32 aims to calculate a first feature quantity x1 such that f(x1)=C. In the example shown in FIG. 6, the feature quantity is x1 and the activity data is C, and the points are indicated by black circles. The initial value setting unit 320 sets the initial values xi of the features to be input to the learning model. As an example, the initial value setting unit 320 generates a random number and sets the initial value xi. As another example, the initial value setting unit 320 may add a random number to a second feature quantity x2 indicating the current feature of the organization to set the initial value xi. In the example shown in FIG. 6, the current organization is indicated by a black rectangle.

The update unit 321 calculates the initial value B of the activity data value, which is the output when the initial value x of the feature is input to the learning model. In the example shown in FIG. 6, the point where the feature is x and the activity data is B is indicated by a black triangle. The update unit 321 updates the feature x until the difference between the value of the activity data and the target value C becomes less than a predetermined threshold.

The "predetermined threshold" here is a reference value for stopping updating that the update unit 321 refers to in order to stop the update process. The specific value of the predetermined threshold can be determined through experiments, taking into consideration the balance between the accuracy of the update process by the update unit 321 and the calculation speed. The update unit 321 can sequentially update the feature x by using, for example, a known gradient method.

The update unit 321 may stop the update process if the update process has been repeated until a predetermined upper limit number of updates is reached, even before the difference between the initial value B and the target value C becomes less than a predetermined threshold value. This allows the update unit 321 to avoid a situation in which the update process does not converge and calculations continue.

The first feature output unit 322 outputs the feature updated until the difference from the target value C becomes less than a predetermined threshold as the first feature x1. This allows the model inverse calculation unit 32 to calculate the feature x such that the difference from the target value C becomes less than the predetermined threshold. The output unit 34 may also output the value of the updated activity data to the display unit 4.

Returning to the explanation of FIG. 3, the differential feature amount acquired by the information processing device 1 through calculation can be said to be the feature amount that the organization should have in order for the organization's activity data to become the target value. However, the differential feature amount is an amount mechanically derived based on an existing learning model, and there is no guarantee of its feasibility.

The adjustment receiving unit 35 therefore receives an adjustment amount for adjusting the value of the differential feature amount. Specifically, the adjustment receiving unit 35 receives the adjustment amount from the user of the information processing device 1 via a user interface of the information processing device 1. The adjustment execution unit 36 adjusts the differential feature amount based on the adjustment amount received by the adjustment receiving unit 35 to calculate an adjusted feature amount. The output unit 34 outputs a character string corresponding to the adjusted feature amount to the display unit 4. This allows the information processing device 1 to output advice adjusted by the user of the information processing device 1.

The model application unit 37 also inputs the first feature, which has been corrected so that the difference between the first feature x1 and the second feature x2 becomes the adjusted feature, into the learning model to calculate adjusted activity data. Specifically, the model application unit 37 inputs, into the learning model, as the corrected first feature, the feature obtained by adding the second feature to the adjusted feature adjusted by the adjustment execution unit 36. The output unit 34 outputs the adjusted activity data calculated by the model application unit 37 to the display unit 4.

When the user of the information processing device 1 modifies the differential feature, the first feature x1 is modified as a result. When the modified first feature x1 is input, the output of the learning model may be a value different from the target value, activity data C. The adjustment receiving unit 35 displays the activity data corresponding to the modified first feature x1, so that the information processing device 1 can show the extent to which the modification of the differential feature will affect the final activity data.

After the model inverse calculation unit 32 calculates the differential feature amount, the adjustment receiving unit 35 acquires the adjustment amount of the calculated differential feature amount. Alternatively, to eliminate the need to adjust the differential feature amount, a constraint condition may be set in advance for the calculation of the model inverse calculation unit 32. Specifically, first, the constraint condition receiving unit 38 receives the setting of a constraint condition related to the feature amount x from the user of the information processing device 1 via a user interface. The information processing device 1 sets the constraint condition so that, for example, the implementation of an improvement measure for the activity data is within a realistic range.

The model inverse calculation unit 32 updates the feature quantity x within a range that satisfies the constraint conditions accepted by the constraint condition acceptance unit 38. Specifically, the model inverse calculation unit 32 updates the feature quantity x using a known constrained gradient method. This allows the user of the information processing device 1 to obtain advice for improving the organization's activity data within a range that is practical to implement.

As described above, the information processing device 1 according to the embodiment aims to provide advice for achieving the numerical targets of the organization to which the members belong, and ultimately needs to deliver the advice to the members. Here, if the members belong to an SNS and the feature x is information about the users in the SNS (specifically, at least one of information indicating the connections between users and information indicating the users themselves), the output unit 34 may output a character string corresponding to the difference feature to the SNS instead of or in addition to the display unit 4. Specifically, the output unit 34 can be realized using, for example, a known BOT function. This allows the information processing device 1 to save the user of the information processing device 1 the trouble of notifying the members of the organization of the advice.

<Processing flow of information processing method executed by information processing device 1>
7 is a flowchart for explaining the flow of information processing executed by the information processing device 1 according to the embodiment. The processing in this flowchart starts, for example, when the information processing device 1 is started.

The model acquisition unit 30 receives as input feature quantities indicative of the characteristics of the organization and acquires a learning model trained to output activity data, which is data indicative of the magnitude of the organization's activities (S2). The goal acceptance unit 31 accepts input of a goal value for the activity data from the user of the information processing device 1 (S4).

The model inverse calculation unit 32 calculates a first feature amount, which is a feature amount to be input to the learning model so that the learning model can output a target value (S6). The difference acquisition unit 33 acquires a difference feature amount, which is the difference between the first feature amount and a second feature amount indicating the current characteristics of the organization (S8).

The output unit 34 outputs a character string corresponding to the differential feature to the display unit 4 (S10). When the output unit 34 outputs the advice, the process in this flowchart ends.

<Effects of the information processing device 1 according to the embodiment>
As described above, the information processing device 1 according to the embodiment can provide advice for achieving the numerical targets of the organization to which the members belong, based on quantitative information.

Although the present invention has been described above using embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes are possible within the scope of the gist of the invention. For example, all or part of the device can be configured by distributing or integrating functionally or physically in any unit. In addition, new embodiments resulting from any combination of multiple embodiments are also included in the embodiments of the present invention. The effect of the new embodiment resulting from the combination also has the effect of the original embodiment.

<First Modification>
In the above, the case where the human network of members shown in Fig. 1(b) is a weighted directed graph has been described. However, the human network is not limited to a weighted directed graph, and may be an unweighted directed graph or an undirected graph.

<Second Modification>
In the above, a case has been described in which the adjustment execution unit 36 adjusts the differential feature amount based on the adjustment amount received from the user, in order to ensure the feasibility of the differential feature amount acquired by the information processing device 1 through calculation, or a constraint condition is set in advance so that advice as an improvement measure for activity data is within a realistic range. Instead of or in addition to this, advice with a high expected value for the intervention effect may be provided by modeling whether or not members of an organization followed advice given in the past, that is, the acceptability of the advice, while taking into account the success rate of the intervention effect obtained by back-calculation.

Below, the information processing investment 1 relating to the second modified example will be explained, but matters common to the information processing device 1 relating to the above-mentioned embodiment will be omitted or simplified as appropriate.

Figure 8 is a diagram showing, in table form, a list of features handled by the information processing device 1 according to the second modified example. As shown in Figure 8, the information processing device 1 according to the second modified example handles a first feature, a second feature, a third feature, a fourth feature, and a difference feature.

As described above, when the feature amount is x, the activity data is y, and f is a fixed multivariate function representing a learning model, it can be formulated as y=f(x). When aiming to make the activity data a target value C, the first feature amount x1 that realizes this is x ₁ =f ^-1 (C). Since the first feature amount x1 is a feature amount that makes the activity data an ideal target value C, for convenience of explanation, in the information processing device according to the second modified example, the first feature amount x1 is described as the first feature amount x _ideal hereinafter. At this time, when the feature amount of the organization before providing advice to the organization is the second feature amount x _act , the difference feature amount x _diff between the first feature amount x _ideal and the second feature amount x _act is x _diff =x _ideal -x _act , and the content corresponds to advice for making the activity data of the organization a target value C.

The first feature x _ideal and the second feature x _act are node features that represent the adjacency matrix representing communication between members of the organization and the features of the members themselves. The character string that is the advice acquired by the information processing device 1 through calculation is performed in the hope of such transformation, but there is no guarantee that each member will act as intended according to the advice, and it may happen that the expected effect is not obtained by providing the advice. Therefore, a receptivity model R that represents the receptivity of members of the organization to advice is defined as follows from the first feature x _ideal for achieving the target value C of the activity data and the third feature x' _act of the organization after the advice.

The deviation between the first feature x _ideal and the third feature x' _act is defined as r = |x' _act -x _ideal | using any distance function (for example, a general norm such as the Frobenius norm). This indicates that the smaller r is, the more the member acts in accordance with the advice.

A function R that measures r in a plurality of different cases and outputs an estimated value of r corresponding to the input of the first feature x _ideal and the difference feature x _diff is calculated using a known learning method such as a neural network or a linear regression model. That is, the acceptability model R is a second learning model different from the learning model acquired by the model acquisition unit 30, and is trained to input the second feature x _act and the difference feature x _diff and output a value r reflecting the deviation between the third feature x' _act and the first feature x _ideal . The second learning model is expressed by the following formula (3).
R( _xact , _xdiff )=r (3)

The smaller the output of the acceptability model R, the higher the acceptability of the organization to the advice. The information processing device 1 according to the second modification is designed to provide advice with a large amount of activity data of the organization and high acceptability. In order to confirm the relationship between the advice and the acceptability of the organization, the difference feature does not necessarily have to be the difference between the first feature x _ideal and the second feature x _act , and a maximization feature that maximizes the output of the learning model f may be used instead of the first feature x _ideal for learning. That is, in the information processing device 1 according to the second modification, the second learning model is generated in advance by learning and stored in the storage unit 2. When the target value C is set as the maximum value of the learning model f, the first feature x _ideal is x _ideal = argmaxF(x).

The model inverse calculation unit 32 according to the second modification calculates a feature quantity x that optimizes an evaluation function G composed of the learning model F(x) acquired by the model acquisition unit 30 and an acceptability model R(x _act , x _diff ). The evaluation function G is expressed by the following formula (4).

In formula (4), the arguments F, R, and _xact of the evaluation function G are known, and the only variable is x. As shown in formula (4), the evaluation function G is designed to output a larger value as the output value of the learning model f increases, and to output a larger value as the output value of the second learning model R decreases. Instead of the evaluation function G shown in formula (4), the evaluation function G' shown in the following formula (5) may be used.

The model acquisition unit 30 according to the second modified example reads and acquires the learning model f and the receptivity model R from the storage unit 2. The model inverse calculation unit 32 generates an evaluation function G shown in formula (4) from the learning model f and the receptivity model R. The model inverse calculation unit 32 updates and calculates a fourth feature amount _xr , which is a feature amount that sequentially maximizes the evaluation function G, for example, by using a known gradient method. That is, the fourth feature amount _xr is a feature amount that is a target of the organization in consideration of the receptivity of the organization, and is expressed by the following formula (6).

The difference acquisition unit 33 acquires the difference between the fourth feature amount _xr and the second feature amount indicating the current feature of the organization, and the output unit 34 outputs a character string corresponding to the difference to the display unit. In this way, the information processing device 1 according to the second modified example can provide advice for improving the activity data of the organization while taking into consideration the receptivity of the organization to the advice.

The information processing device 1 according to the second modified example can collect learning data for the acceptability model R by providing advice to an organization and acquiring new information about the acceptability of its members. This allows the information processing device 1 according to the second modified example to continuously update the acceptability model R and improve its accuracy.

Reference Signs List 1: Information processing device 2: Memory unit 3: Control unit 30: Model acquisition unit 31: Goal acceptance unit 32: Model reverse calculation unit 320: Initial value setting unit 321: Update unit 322: First feature quantity output unit 33: Difference acquisition unit 34: Output unit 35: Adjustment acceptance unit 36: Adjustment execution unit 37: Model application unit 38: Constraint condition acceptance unit 4: Display unit

Claims (11)

On the computer,
A function of acquiring a learning model that has been machine-learned by a neural network so as to input a feature quantity indicating the characteristics of an organization and output activity data that is data indicating the magnitude of the organization's activities;
A function of receiving an input of a target value of the activity data;
A function of calculating a first feature amount which is a feature amount to be input to the learning model so that the learning model outputs the target value;
a function of acquiring a difference feature amount that is a difference between the first feature amount and a second feature amount that indicates a current feature of the tissue;
a function of outputting a character string corresponding to the differential feature amount to a display unit;
A program to achieve this. The computer includes:
a function of receiving an adjustment amount for adjusting the value of the differential feature amount;
a function of adjusting the difference feature amount based on the adjustment amount to calculate an adjusted feature amount,
The output function outputs a character string corresponding to the adjusted feature.
The program according to claim 1. The computer includes:
a function of inputting the first feature amount corrected so that a difference between the first feature amount and the second feature amount becomes the adjusted feature amount into the learning model to calculate adjusted activity data;
The outputting function outputs the adjusted activity data.
The program according to claim 2. The function of calculating the first feature amount is
A function of setting initial values of features to be input to the learning model;
a function of updating the feature amount until a difference between the value of the activity data, which is an output when the feature amount is input to the learning model, and the target value becomes less than a predetermined threshold value;
a function of outputting, as the first feature amount, the feature amount that has been updated until a difference between the feature amount and the target value becomes less than a predetermined threshold value;
The program according to claim 1 , further comprising: The computer includes:
Further realizing a function of accepting settings of constraint conditions related to the feature amount,
The updating function updates the feature amount within a range that satisfies the constraint condition.
The program according to claim 4. The feature amount is information about a user in a social networking service (SNS),
The output function transmits the character string to the SNS.
The program according to any one of claims 1 to 5. On the computer,
a function of acquiring a second learning model trained based on a plurality of combinations of a second feature amount and a difference feature amount calculated by a computer that executes the program according to claim 1 and a third feature amount indicating a feature of a tissue after presentation of the character string, the second learning model being trained so as to output a value reflecting a deviation between the third feature amount and the first feature amount by using the second feature amount and the difference feature amount as input;
A function of generating an evaluation function related to a feature quantity indicating a feature of an organization from the learning model and the second learning model;
A function of calculating a fourth feature amount which is an extreme value of the evaluation function;
a function of acquiring a difference feature amount that is a difference between the fourth feature amount and the second feature amount;
a function of outputting a character string corresponding to the differential feature amount to a display unit;
A program to achieve this. The evaluation function outputs a larger value as the output value of the learning model is larger, and outputs a larger value as the output value of the second learning model is smaller.
The program according to claim 7. a model acquisition unit that acquires a learning model that has been machine-learned by a neural network so as to input a feature quantity indicating an organization's characteristics and output activity data that is data indicating the magnitude of the organization's activities;
a goal receiving unit that receives an input of a goal value of the activity data;
a model inverse calculation unit that calculates a first feature amount that is a feature amount to be input to the learning model so that the learning model outputs the target value;
a difference acquisition unit that acquires a difference feature amount that is a difference between the first feature amount and a second feature amount that indicates a current feature of the tissue;
an output unit that outputs a character string corresponding to the differential feature to a display unit;
An information processing device comprising: The processor:
A step of acquiring a learning model that is machine-learned by a neural network so as to input a feature quantity indicating a characteristic of an organization and output activity data that is data indicating the magnitude of the activity of the organization;
A step of receiving an input of a target value of the activity data;
calculating a first feature amount which is a feature amount to be input to the learning model so that the learning model outputs the target value;
acquiring a difference feature amount that is a difference between the first feature amount and a second feature amount indicating a current feature of the tissue;
outputting a character string corresponding to the differential feature amount to a display unit;
An information processing method for performing the above. On the computer,
A function of acquiring a learning model that has been machine-learned by a neural network so as to input a feature quantity indicating the characteristics of an organization and output activity data that is data indicating the magnitude of the organization's activities;
a function of acquiring a difference feature amount that is a difference between a predetermined first feature amount indicating a target feature of the tissue and a second feature amount indicating an actual feature of the tissue;
a function of acquiring a second learning model trained based on a plurality of combinations of the second feature amount, the difference feature amount, and a third feature amount indicating a characteristic of an organization after members of the organization have acted so that the characteristic of the organization becomes the first feature amount, the second learning model being trained to output a value reflecting a deviation between the third feature amount and the first feature amount, using the second feature amount and the difference feature amount as input;
A function of generating an evaluation function related to a feature quantity indicating a feature of an organization from the learning model and the second learning model;
A function of calculating a fourth feature amount which is an extreme value of the evaluation function;
a function of acquiring a difference feature amount that is a difference between the fourth feature amount and the second feature amount;
a function of outputting a character string corresponding to the differential feature amount to a display unit;
A program to achieve this.

Images