JP6801281B2 - Promotion support equipment and programs - Google Patents

Description

The present invention relates to a promotion support device and a program.

In recent years, social listening (social reading) has been proposed as a marketing method that is useful for product development and service improvement by collecting and analyzing user's writings on social media. In social listening, unlike the opinions collected in customer surveys, it is thought that more natural and potential demands and demands can be grasped.

In Patent Document 1, the amount of content spread within the group is memorized from the group information of the users held in the SNS and the activity history of each user, and the reward contents of the affiliate type advertisement are obtained from this information. It is stated that it will be decided.

Patent Document 2 describes that if a URL with its own ID as a parameter is written on SNS and another user sees the writing and sees the promotion from the URL, the number of applications of the written user is increased. ..

JP-A-2015-230645 Japanese Unexamined Patent Publication No. 2014-142835

It is effective to analyze and output user trends on the Internet such as SNS word-of-mouth, but it only provides a single piece of data, and users need technology to interpret multiple analysis results. The reality is that experience and sense are required to make an effective promotion plan.

An object of the present invention is to provide a promotion support device and a program that can easily understand what kind of promotion should be performed and what kind of effect can be obtained, and can plan a promotion.

The invention according to claim 1 teaches a storage means for storing promotion history data indicating a promotion history of a product or service, a plurality of reverberation data for the promotion, the promotion history data, and the reverberation data. used as data, and processing means for outputting the simulation results of the promotion using promotion input, and model generation means for generating a learning model for a particular echo of the output, the generated model, said processing The means inputs information that identifies the product or service to be simulated and information that identifies the promotion to be executed, and the information that identifies the product or service includes the name of the product or service and a specific date. The information that includes the sales target value and identifies the promotion to be executed is a promotion support device that includes the type of promotion and the date or period of the promotion.

In the invention according to claim 2, the response data includes sales data indicating the sales amount or sales amount of the goods or services, and the processing means determines the estimated sales amount or sales amount on a specific date as a result of the promotion. The promotion support device according to claim 1, which is output.

The invention according to claim 3 is the promotion support device according to claim 2, wherein the processing means outputs the target sales amount or the degree of achievement of the sales amount on a specific date as a result of the promotion.

The invention according to claim 4 is the promotion support device according to any one of claims 1 to 3 , wherein the model generation means generates a model of a multi-layer neural network.

Invention of claim 5, the model generation unit generates a plurality model of a single model to output a particular echo for each echo, according to claim 1 to 3 for generating an integrated model that integrates multiple models to one another The promotion support device described in any one of them.
The invention according to claim 6 is a storage means for storing promotion history data representing a promotion history of a product or service, a plurality of reaction data for the promotion, and teaching data of the promotion history data and the reaction data. The model is provided with a model generation means for generating a model in which the input is promoted and the output is a specific reaction by learning, and a processing means for simulating and outputting the promotion result using the generated model. The generation means is a promotion support device that generates a single model that outputs a specific reverberation as a plurality of models for each reverberation and generates an integrated model in which the plurality of models are integrated with each other.

The invention according to claim 7 uses a step of storing promotion history data and its reverberation data in a storage device in a computer, and using the promotion history data and its reverberation data as teaching data, promoting an input and a specific reverberation of an output. It is a program that executes a step of generating a model to be obtained by learning and a step of simulating and outputting the promotion result using the generated model , and information that identifies the product or service that is the target of the simulation, and Enter the information that identifies the promotion to be executed, and the information that identifies the product or service includes the name of the product or service and the sales target value on a specific date, and is included in the information that identifies the promotion to be executed. Is a program that includes the type of promotion and the due date or duration of the promotion .
The invention according to claim 8 uses a step of storing promotion history data and its reverberation data in a storage device in a computer, and using the promotion history data and its reverberation data as teaching data, promoting an input and a specific reverberation of an output. It is a program that executes a step of generating a model to be obtained by learning and a step of simulating and outputting the promotion result using the generated model. In the step of generating the model, a specific reaction is output. It is a program that generates a single model for each response and generates an integrated model that integrates the multiple models with each other.

According to the inventions of claims 1 , 6, 7 , and 8, promotion planning for obtaining a desired effect can be supported. Further, according to the inventions of claims 1 and 7, further, the effect of the promotion can be known by inputting the information for specifying the product or service to be simulated and the information for specifying the promotion to be executed. Can be done. Further, according to the inventions of claims 6 and 8, promotion planning can be further supported by using the integrated model.

According to the invention of claim 2, the user can further know the sales volume on a specific date.

According to the invention of claim 3, the user can further know the degree of achievement of the target on a specific date.

According to the invention of claim 4 , it is possible to further support promotion planning by using a multi-layer neural network.

According to the invention of claim 5 , promotion planning can be supported by using the integrated model.

It is a block diagram of the structure of embodiment. It is a functional block diagram of an embodiment. It is a model schematic diagram (No. 1) of an embodiment. It is a model schematic diagram (No. 2) of an embodiment. It is a model explanatory diagram of the amount of reverberation when the promotion is performed a plurality of times in the embodiment. It is explanatory drawing of the promotion history data of an embodiment. It is explanatory drawing of the purchase data of an embodiment. It is explanatory drawing of the learning model of embodiment. It is explanatory drawing of teaching data of embodiment. It is explanatory drawing of the integrated model of embodiment. It is a processing flowchart (the 1) of embodiment. FIG. 2 is a processing flowchart of the embodiment. This is an example of a simulation input screen of the embodiment. This is an example of a simulation output screen (No. 1) of the embodiment. This is an example of a simulation output screen (No. 2) of the embodiment. This is an example of a simulation output screen (No. 3) of the embodiment. This is an example of a simulation output screen (No. 4) of the embodiment. This is an example of a simulation output screen of another embodiment. This is an example of a simulation output screen of still another embodiment.

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

<First Embodiment>
FIG. 1 shows a block diagram of the promotion support device 10 according to the present embodiment. The promotion support device 10 has a configuration as a computer. That is, it includes an input unit 12, a CPU 14, a communication interface (I / F) 16, a display unit 18, a memory 20, and a database 22.

The memory 20 is a program memory such as a ROM for storing a processing program, and a working memory such as a RAM.

The CPU 14 executes each process described later by reading and executing the processing program stored in the program memory. The CPU 14 learns the relationship between the promotion and its result, in particular, based on the past promotion history, response, and purchase data. Further, the CPU 14 executes a promotion simulation using the learning result and outputs the simulation result.

The input unit 12 inputs the past promotion history, response, purchase data, and the like.

The display unit 18 is a liquid crystal display, an organic EL display, or the like, and outputs a simulation result by the CPU 14.

The database 22 stores past promotion history, responses, and purchase data. In addition, the database 22 stores the learning result executed by the CPU 14.

Although the promotion support device 10 is shown as a stand-alone in FIG. 1, the promotion support device 10 may be configured as a server computer in a server client system. The terminal that is the client and the server computer are connected by wire or wirelessly, various data are input from the terminal via the communication I / F16, the promotion and its results are learned, and the simulation result is sent to the terminal via the communication I / F16. Output. The terminal and the server computer may be connected by a dedicated line, or may be connected by a public line such as the Internet.

FIG. 2 shows a functional block diagram of the CPU 14 and the database 22.

The CPU 14 includes a echo learning module 141, a model integration module 142, and a simulation module 143 as functional modules.

On the other hand, the database 22 includes a promotion history storage unit 221 for storing promotion history data, a reaction storage unit 222 for storing reverberation data, a purchase data storage unit 223 for storing purchase data, and a learning result storage unit 224. Specifically, the database 22 is composed of auxiliary storage devices such as a hard disk drive (HDD) and a solid state drive (SSD), and promotion history data, reverberation data, and purchase data are used as teaching data when performing machine learning. Be done.

The reverberation learning module 141 learns the correlation between promotion and reverberation based on the promotion history data, reverberation data, and purchase data accumulated in the database 22. Here, the response is not only the purchase of products or services, but also the PV of the web (page views: the number of times a website has been viewed within a certain period of time), mentions on SNS, reviews, and coverage on TV and magazines. It is a concept that includes the service. Learning is performed on each reaction unit, such as web PV, SNS references, reviews, and coverage on TV and magazines. For promotion, learning is performed by calculating the coefficient of the model (prediction calculation formula) using each echo data. The echo learning module 141 registers the obtained model for each reaction unit in the database 22 as a learning result.

The model integration module 142 integrates the model for each reaction unit obtained by learning with the echo learning module 141 into one model and registers it in the database 22.

The simulation module 143 executes a simulation using the model integrated by the model integration module 142, and simulates the effect of the promotion by outputting the expected result for the input promotion (sales plan).

FIG. 3 schematically shows a model used for learning in this embodiment. In the figure, the horizontal axis represents time and the vertical axis represents the amount of reverberation. When a stimulus such as an advertisement is input (introduced) to the market at a certain timing, the amount of reverberation rises at O2 after the delay of O1, and then the effect is attenuated with a constant attenuation curve and lasts only for the period of O3. is there. This model is a standard model for general promotion.

FIG. 4 is a quantification of the model of FIG. The increase attenuation curve of FIG. 3 is expressed by the reverberations d1, d2, and d3 with respect to the elapsed time from inputting (inputting) a stimulus such as an advertisement to the market.

On the other hand, promotion is usually not performed only once at a certain timing, but multiple times at intervals. The amount of reverberation in this case is a superposition of the amount of reverberation of each promotion.

FIG. 5 shows a model of the amount of reverberation when the promotion is performed a plurality of times. The total amount of reverberation 200 is represented by superimposing the reverberations 100, 102, 104, 106, 108, and 110 for each promotion. Here, there is a reaction such as mention of SNS to the promotion, which affects the purchase performance, and the purchase performance further leads to reviews and other references of SNS. That is, one echo is not independent of another, and one echo affects another.

FIG. 6 shows an example of promotion history data stored in the promotion history storage unit 221 of the database 22. The implementation ID, product ID, promotion method, promotion implementation date, and promotion period are stored. For example, for a product with product ID = AAA, it is shown that the TV commercial was started on 2/1 and the product was promoted for 10 days.

FIG. 7 stores the product ID and the purchase date, which show an example of the purchase data stored in the purchase data storage unit 223 of the database 22. For example, it indicates that the product with the product ID = AAA was purchased on 2/2. This purchase is due to the promotion of TV commercials that started on February 1.

FIG. 8 shows a model used for learning in the echo learning module 141 of the CPU 14. Give a promotion as an input and a response as an output, for example the number of SNS references. For example, when TVCM is continuously performed, the promotion given as input is given 3 days before TVCM amount, 2 days before TVCM amount, 1 day before TVCM amount, and if there is SNS mention number, SNS mention number 3 days ago, SNS mention. Give a few days ago ... etc. Here, the number of SNS references is given to the input because, as described above, when a purchase occurs due to promotion, the reference in SNS increases due to the purchase, which is related to generate a new purchase. is there. Here, the formula of the model is a function output = Σα _k x _{k, where} x is the input element, α is the coefficient (weight), and k is the subscript.
The optimum α is calculated using the input amount and the output amount. Learning using the model is performed for each reaction unit, that is, for each SNS reference or review, featured on TV or magazine, or purchased.

FIG. 9 shows an example of teaching data used for learning. Inputs 1, 2, 3, 4, ... Are given, and their respective outputs are given. For example, as an input
Input 1 = TVCM amount 3 days ago Input 2 = TVCM amount 2 days ago Input 3 = TVCM amount 1 day ago Input 4 = SNS mention number 3 days ago Input ... = SNS mention number 2 days ago Input ... = Sales volume 3 days ago ...
It is the number of SNS mentions as an output. In a simplified example, as two inputs,
Input 1 = 3 days before TVCM amount Input 2 = 2 days before TVCM amount, and the weights are α ₁ and α ₂ , respectively, and the output is output = the number of SNS references.
Σα _k x _k = α ₁ · (TVCM amount 3 days ago) + α ₂ · (TVCM amount 2 days ago) = SNS reference number, using multiple data stored in the promotion history storage unit 221 and the echo storage unit 222 Calculate the optimum α ₁ and α ₂ that satisfy this model. The weight is standardized to 0 ≦ α ≦ 1.

When the model is learned by calculating the coefficient (weight) for each reaction unit as described above, the reverberation learning module 141 stores the learned model in the learning result storage unit 224 of the database 22. The model integration module 142 refers to the learning model stored in the learning result storage unit 224, and generates an integrated model by connecting the same input and output of the learning model (model with a single output) to each other. To do. For example, if the output of one model A is the same as the input of another model B, the model A and the model B are joined to generate one model. More specifically, by repeating the work of connecting other models whose output elements are the input elements of this model to the model, based on the model whose output is the result of the promotion simulation. , Integrate multiple single models into one model. At this time, those having a relatively small coefficient (weight) are removed (pruned) without being joined because the influence can be ignored.

FIG. 10 schematically shows model integration in the model integration module 142. There is a model that outputs the sales volume on a certain base date (this is the parent model), and the inputs are "CM-3d", "CM-2d", "CM-1d", "KW-3d", It is assumed that there are "KW-2d" and "KW-1d". Here, "CM-3d" indicates that the amount of TVCM is 3 days before the reference date. Further, "KW-3d" indicates that the number of SNS mentions is 3 days before the reference date.

Focusing on "KW-2d" among these inputs of the parent model, it is assumed that there is another model whose output is "KW-2d", that is, the number of SNS references two days before the reference date. , Connect the model to "KW-2d". In the figure, if there is a model whose inputs are "CM-3d" and "CM-2d" and whose output is "KW-2d", this model is connected to "KW-2d". "Joining" means taking the output of one model as the input of another model.

Similarly, when focusing on "KW-1d" among the inputs of the parent model, it is assumed that there is another model whose output is "KW-1d", that is, the number of SNS references one day before the reference date. Then, the model is connected to "KW-1d". In the figure, if there is a model whose input is "sales volume-1d" and whose output is "KW-1d", this model is connected to "KW-1d". Further, if there is another model whose input is "KW-3d" and whose output is "sales volume-1d", the model is connected to "sales volume-1d".

By repeating such operations, an integrated model that finally outputs the "sales date (base date)" is created. In addition, the input in the integrated model is a controllable promotion (direct promotion) such as 3 days before the TCCM amount and 2 days before the TVCM amount, and does not include indirect promotion by a third party such as mentioning SNS.

11 and 12 show a processing flowchart in the model integration module 142.

FIG. 11 is an overall processing flowchart. The purchase amount is set as the target item, the default coefficient (weight) is set to 1, and the influence limit, that is, the threshold value for pruning is set to 0.001 near 0 (S101). Then, the model (single model) for each reaction unit stored in the learning result storage unit 224 is read out and integrated processing is performed (S102).

FIG. 12 is a detailed processing flowchart of the integrated processing (S102) in FIG.

First, it is determined whether or not there is data for which the target item is output (S201). When the target item is set to the purchase amount, it is determined whether or not there is data for which the purchase amount is output. If it exists, the process proceeds to the process of creating a new combination model (S202).

That is, the input element of the single model whose output is the target item is acquired after being read from the reference date as the starting point (S203). For example, the input element of the model that outputs the purchase amount is
Sales volume (1 day ago)
Sales volume (2 days ago)
Sales volume (3 days ago)
Therefore, assuming that the base date of the model is 2 days, these input elements are read as the base date as the starting point.
Sales volume (1 day ago) → Sales volume (3 days ago)
Sales volume (2 days ago) → Sales volume (4 days ago)
Sales volume (3 days ago) → Sales volume (5 days ago)
And.

Next, k is set as the input element as the target item, the weight is set to the current weight (1) × the weight (α) of k, the reference date is set to the number of days (y) of the input element (S204), and the weight exceeds the influence limit. (S205), and if the weight exceeds the influence limit, a joining operation is performed (S206). That is, models with the same input and output are connected to each other. A model having the same input and output and capable of being connected to each other is added as the k-th child model of the new combined model created in S202 (S207). For example, among the input elements after replacement
Sales volume (3 days ago)
If there is a model that outputs the sales volume (3 days ago), add it as the first child model of the combined model.

After that, k is incremented by 1 (S208), and the processes of S204 to S208 are repeated until k = n finally. However, n is the number of input elements. As a result, the model is combined with the new united model. At this time, if the weight of the input element is less than the influence limit, the joining operation is not executed (NO in S205, S206 and S207 are skipped), so the model that outputs the input element is added as a child model. It is never pruned. When the models are combined for all the input elements, the combined model is returned (S210). Here, the return is a model that is the standard of integration set at the beginning, and means that the output is replaced with a model that is the purchase amount (base date). Then, the processing after S201 is repeated again. As a result, models whose outputs match the input elements are sequentially added as child models.

On the other hand, if there is no data for which the target item is output (S201), the target item and the reference date are returned without creating the combined model (S211). As a result, only the target item and the base date are combined with the combined model and carried over to the next process.

In this way, when the single model is integrated by going back in time sequentially from the base date, finally, the input element executed as the promotion is input, and the single integrated model based on the purchase date (base date) is obtained. can get. The input of this integrated model is an element executed as a promotion, not an element that the promoter cannot control such as mentioning SNS or being featured in TV / magazine, but a promoter such as posting to TV commercials or magazines. Only elements that can be controlled independently. This integrated model will be the model used in the next simulation. The integrated model is registered in the database 22.

FIG. 13 shows a screen example of the simulation executed by the simulation module 143 displayed on the display unit 18. The simulation module 143 reads the integrated model registered in the database 22, inputs the promotion plan input by the user, and outputs the output, specifically, calculates and outputs the purchase amount on the base date.

Product ID, due date, product attribute, and target value are input to the simulation screen as basic information. The due date is the date on which you want to calculate the purchase amount by simulation and is the base date. The target value is the target purchase amount on the due date. Further, as the promotion information, the promotion type, the start time, and the end date are input. The promotion type is TV commercial, posting to a magazine, posting on the web, etc., and is input by selecting from a pull-down menu, for example. As many promotions as planned are additionally entered. The user, for example,
Date: 4/1
Posting on the web: 2/1
TV commercials: 2/20 to 2/25
And so on. Based on these inputs, the simulation module 143 gives the input elements of the integrated model registered in the database 22 and outputs the purchase amount on the base date. The simulation module 143 outputs the result of the purchase amount obtained by the simulation to the display unit 18.

FIG. 14 shows an example of the simulation result.
Date: 4/1 (until the end of March)
Target value: 1 million sheets Promotion: When the notification is entered on the HP on 2/1. The transition of the sales volume of 500 obtained by the simulation at this time and the transition of each response (number of mentions of PV and SNS, number of reviews, etc.) are displayed in a graph with the horizontal axis as the time axis, and sales on 4/1. The numerical value of the amount is displayed as 10,000 sheets. Since 10,000 sheets are numbers that have not reached the target value, the simulation module 143 compares the magnitude of the simulation result and the target value, and if the simulation result <target value, the message "The target is not reached" is displayed. indicate.

FIG. 15 shows another example of the simulation result.
Date: 4/1 (until the end of March)
Target value: 1 million sheets Promotion: Announcement on HP on 2/1 and TVCM broadcast on 2/20 to 2/25. The transition of the sales volume of 500 obtained by the simulation at this time and the transition of each response (number of mentions of PV and SNS, number of reviews, etc.) are displayed in a graph with the horizontal axis as the time axis, and sales on 4/1. The numerical value of the amount is displayed as 100,000 sheets. Since 100,000 sheets are numbers that have not reached the target value, the simulation module 143 compares the magnitude of the simulation result and the target value, and assumes that the simulation result <target value, and sends a message "The target is not reached". indicate.

FIG. 16 shows yet another example of the simulation results.
Date: 4/1 (until the end of March)
Target value: 1 million sheets Promotion: Announcement on HP on 2/1 and TVCM broadcast on 3 / 1-3 / 31. The transition of the sales volume of 500 obtained by the simulation at this time and the transition of each response (number of mentions of PV and SNS, number of reviews, etc.) are displayed in a graph with the horizontal axis as the time axis, and sales on 4/1. The numerical value of the amount is displayed as 500,000 sheets. Although the simulation result is larger than that in the case of FIG. 14 or FIG. 15, since 500,000 sheets are numbers that have not reached the target value yet, the simulation module 143 compares the simulation result with the target value. , Simulation result <Assuming that the target value, the message "The target is not reached" is displayed in the same way.

FIG. 17 shows yet another example of the simulation results.
Date: 4/1 (until the end of March)
Target value: 1 million sheets Promotion: Announced on 2/1 on HP, and entered a sales promotion campaign on Twitter on 3 / 1-3 / 31. The transition of the sales volume of 500 obtained by the simulation at this time and the transition of each response (number of mentions of PV and SNS, number of reviews, etc.) are displayed in a graph with the horizontal axis as the time axis, and sales on 4/1. The numerical value of the quantity is displayed as 1.01 million sheets. Since 1.01 million sheets are numbers that have reached the target value, the simulation module 143 compares the magnitude of the simulation result and the target value, and if the simulation result> the target value, the message "The target is not reached" is displayed. It does not display, and indirectly notifies the user that the target value is satisfied.

As shown in FIGS. 14 to 17, the user can sequentially input the type and period of promotion, perform a simulation, know the purchase amount (sales amount) on the desired date, and reach the target value. It is possible to make such a promotion plan.

Although the embodiments of the present invention have been described above, the present invention is not limited to this, and various modifications are possible. Hereinafter, other embodiments will be exemplified.

<Second Embodiment>
In the first embodiment, the echo data is stored in the echo storage unit 222 and the purchase data is stored in the purchase data storage unit 223. However, the purchase data is regarded as one of the echo data and stored in the echo storage unit 222. You may. In this case, a learning model is generated using the promotion history data and the echo data.

<Third Embodiment>
In the first embodiment, as a reaction to a certain promotion, mentions on the PV of the web, SNS, reviews, coverage on TV and magazines, etc. are exemplified, but the present invention is not limited to these, and consumers for a certain promotion Any reaction can be a reverberation. Further, when learning with the echo learning module 141, instead of using all the echoes, learning may be performed using only the echoes automatically selected by the device or manually selected by the user. For example, learning is performed using only the reference in SNS and the purchase record as a response. However, it goes without saying that the more data used for learning, the better the learning accuracy and the simulation accuracy.

<Fourth Embodiment>
In the first embodiment, the reverberation learning module 141 generates a single model, and the model integration module 142 integrates these single models to generate an integrated model, but such a single model generation → It is not necessary to go through the step of model integration, the point is that if you use promotion history data and reverberation data, you can generate a model by learning that the input is promotion and the output is a specific reverberation (for example, purchase volume (sales volume)). Good. Such learning also includes machine learning of single-layer or multi-layer neural networks. Multi-layer neural networks are known as excellent computational models with high discriminating ability.

<Fifth Embodiment>
In the first embodiment, as a function used for learning, the input is x and the output is y.
y = Σα _k x _k
Is used, but this is also an example, and as a function generally used for machine learning,
y = φ (Σα _k x _k + b)
May be used. Here, φ is an activation function and b is a bias. Moreover, you may use a ramp function as a function.

<Sixth Embodiment>
In the first embodiment, the promotion support device 10 executes model learning, model integration, and simulation using the integrated model, but these processes may be executed by another computer. For example, computer A implements the reverberation learning module, computer B implements the model integration module, computer C implements the simulation module, or computer A implements the reverberation learning module and model integration module, and the simulation module is computerized. B implements it, etc. These computers A to C are connected to each other so as to be able to send and receive data, and can form a promotion support device as a group of computers.

<7th Embodiment>
The promotion support device of the first embodiment can be constructed as a server computer in the server client system, but promotion history data, response data, purchase data, and the like can be supplied to the server computer from a plurality of terminals. The terminal that supplies the promotion history data and the like and the terminal that inputs various information shown in the screen example of FIG. 13 for performing the simulation and receives the simulation result shown in the screen example of FIG. 14 may be the same or different. May be good.

<8th Embodiment>
In the first embodiment, as shown in FIG. 13, as simulation conditions,
Product ID
Product attribute Date Target value Promotion type Start date End date is entered, but it is not necessary to enter all of these, just enter the information that identifies the target of the simulation and the information that identifies the promotion to be executed. .. Information that identifies the target of the simulation is
Product or service attributes (name, etc.) and sales volume (purchase volume)
Product or service attributes (name, etc.) and SNS mentions Product or service attributes (name, etc.) and web PV
・ ・ ・
And so on. The due date and target value are not essential, but usually, when planning a promotion, it is considered that the target value up to a certain date is set and drafted, which is preferable as an input condition. The information that identifies the promotion to be executed is at least the content of the promotion and the date or period of the promotion.

<9th embodiment>
In the first embodiment, as shown in FIG. 14 and the like, as a simulation result, the sales volume and other reactions are displayed as a graph with the horizontal axis as the time axis, but it is not necessary to display all of them, only the sales volume. Or, only a specific response may be displayed. The content to be displayed may change depending on the target of the simulation. When the simulation target is the sales volume of a certain product, at least the expected sales volume on a specific date needs to be displayed as the simulation result, and the relationship with the target value is additionally displayed as necessary. do it.
Alternatively, as a simulation result, only whether or not the target value has been reached may be displayed. For example, in FIG. 14, only the message "The goal is not reached" is displayed, or in FIG. 17, only the message "The goal can be achieved" is displayed. The degree of achievement of the goal may be displayed instead of whether or not the goal has been reached. For example, "80% of the target", "100% of the target", "120% of the target", etc. Messages such as "not reaching the goal" or "achieving the goal" can be regarded as an example of the degree of achievement of the goal.

FIG. 17 shows a case where the degree of achievement of the target is displayed as the simulation result. The degree of goal achievement is indicated as "target achievement rate 105%", and the user can recognize that the goal can be achieved by this. In FIG. 17, only the degree of goal achievement may be displayed as a message, or only the type and date (period) of the promotion and the degree of goal achievement may be displayed.

<10th Embodiment>
In the first embodiment, basically, the user inputs the information for specifying the promotion to be executed as the simulation condition, but the user inputs only the information for specifying the target of the simulation, and the computer inputs the learning model. A promotion for achieving the target value may be calculated by simulation based on the above and presented to the user as a recommended promotion. For example, when a user inputs a target value of sales volume of a certain product on a certain date, a computer uses a learning model to exemplify and enumerate promotions in which the sales volume of the product on that date can reach the target value. is there.

FIG. 19 shows an example when a computer presents a recommended promotion. As a recommended promotion, the type and date of the promotion are shown. For example, as a recommended promotion
2/1 HP announcement 3 / 1-3 / 31 Spread campaign etc. on SNS. Along with this, the degree of achievement of the goal when the promotion is executed according to the recommended promotion is shown. In this display, a button of "another recommended plan" may be further displayed, and when the user operates the button, another recommended promotion may be presented.

In the present embodiment, machine learning using a multi-layered neural network (multi-layer neural network) can be used as described above, but learning of a multi-layer neural network is known (for example, Japanese Patent Application Laid-Open No. 2015-). 95215 (No. 95215), which will be briefly described below.

すべてのｌ＝１，２，…，Ｌに対して、下式（３）〜（７）の変数を定める。 It is assumed that the number of layers L and the number of units of each layer D _l (l = 1, 2, ..., L) are given in advance. The nth (n = 1, 2, ..., N) th input vector is defined as the following equations (1) and (2).

The variables of the following equations (3) to (7) are defined for all l = 1, 2, ..., L.

なお、ｈは単調増加かつ微分可能な非線形活性関数である。
最適化問題として、下式（８）を最小化する。

を求める。

ここで、

であり、ｔⁿは、次元Ｄ_Lの既知の正解出力ベクトル（スカラー）である。

Note that h is a non-linear activation function that is monotonically increasing and differentiable.
As an optimization problem, the following equation (8) is minimized.

To ask.

here,

In and, t ⁿ is the known correct output vector of dimension D _L (scalar).

For learning, the error back propagation method is used, in which updates by the gradient method are repeated in order from the weight closest to the output layer. The initial value at this time is given by a random number. The gradient method is expressed by the following equation (9).

ｌ＝Ｄ−１，Ｄ−２，…，１の重み層の微分は、下式（１３）〜（１６）の計算で求められる。

The method of calculating the derivative used in the gradient method is as follows. In the following equation, the subscripts i, j, and k are vector and matrix indexes.
First, the derivative of the final weight layer is obtained by the calculation of the following equations (10) to (12).

The derivative of the weight layer of l = D-1, D-2, ..., 1 can be obtained by the calculation of the following equations (13) to (16).

The teacher data used for learning is the data stored in the promotion history storage unit 2221, the reverberation storage unit 222, and the purchase data storage unit 223, and these are the multidimensional vectors that are input to the network and the corresponding neural network. It is given as a set of target values to be output. The neural network in the present embodiment may be a forward propagation type in which signals propagate in order from the input layer to the output layer. The neural network can be realized by GPU (graphics processing unit) or FPGA, or the cooperation between these and the CPU 14, but is not limited thereto.

The promotion in each embodiment exemplifies TC CM, posting / campaign on the web, posting / campaign on SNS, posting / campaign on newspapers / magazines, but is not limited to these, and products and services. Includes any activity / message that raises awareness and interest in and promotes purchasing.

10 Promotion support device, 12 Input unit, 14 CPU, 16 Communication I / F, 18 Display unit, 20 Memory, 22 Database, 141 Reverberation learning module, 142 Model integration module, 143 simulation module, 221 Promotion storage unit, 222 Reverberation storage Department, 223 Purchase data storage unit, 224 Learning result storage unit.

Claims (8)

A storage means for storing promotion history data representing a promotion history of a product or service and a plurality of response data for the promotion.
A model generation means that uses the promotion history data and the echo data as teaching data to generate a model in which the input is promotion and the output is a specific echo by learning.
A processing means that simulates and outputs the promotion result using the generated model, and
Equipped with a,
The processing means inputs information that identifies the product or service that is the target of the simulation and information that identifies the promotion to be executed.
The information that identifies the product or service includes the name of the product or service and the sales target value on a specific date.
The information that identifies the promotion to be executed includes the type of promotion and the date or period of the promotion.
Promotion support device. The response data includes sales data indicating the sales volume or sales amount of the goods or services.
The promotion support device according to claim 1, wherein the processing means outputs a predicted sales amount or a sales amount on a specific date as a result of the promotion. The promotion support device according to claim 2, wherein the processing means outputs the target sales amount or the degree of achievement of the sales amount on a specific date as a result of the promotion. The promotion support device according to any one of claims 1 to 3 , wherein the model generation means generates a model of a multi-layer neural network. The promotion support device according to any one of claims 1 to 3 , wherein the model generation means generates a single model that outputs a specific echo as a plurality of models for each echo, and generates an integrated model in which the plurality of models are integrated with each other. A storage means for storing promotion history data representing a promotion history of a product or service and a plurality of response data for the promotion.
A model generation means that uses the promotion history data and the echo data as teaching data to generate a model in which the input is promotion and the output is a specific echo by learning.
A processing means that simulates and outputs the promotion result using the generated model, and
With
The model generation means generates a single model having a specific echo as an output for each echo, and generates an integrated model in which the plurality of models are integrated with each other.
Promotion support device. On the computer
Steps to store promotion history data and its echo data in a storage device,
Using the promotion history data and its echo data as teaching data, the step of generating a model by learning that the input is promotion and the output is a specific echo,
Steps to simulate and output the promotion results using the generated model,
Is a program for executing,
Enter the information that identifies the product or service that you want to simulate and the promotion that you plan to run.
The information that identifies the product or service includes the name of the product or service and the sales target value on a specific date.
The information that identifies the promotion to be executed includes the type of promotion and the date or period of the promotion.
Program . On the computer
Steps to store promotion history data and its echo data in a storage device,
Using the promotion history data and its echo data as teaching data, the step of generating a model by learning that the input is promotion and the output is a specific echo,
Steps to simulate and output the promotion results using the generated model,
Is a program that executes
In the step of generating the model, a single model having a specific echo as an output is generated for each echo, and a plurality of models are integrated with each other to generate an integrated model.
program.

Images