Prediction device, prediction method, and program

Abstract

To provide a prediction device, a prediction method, and a program that predict effective measures.SOLUTION: A prediction device (1) comprises a storage part (14) which stores measure implementation information (144) representing measure effects on a first object having measures taken, and a control part (13) which predicts measure effects on a second object having no measure taken on the basis of the measure implementation information, and the control part structures a first graph (450A to 450D) consisting of a plurality of nodes including at least one first node related to the first object and at least one second node related to the second object and a plurality of links connecting the nodes on the basis of similarities between nodes, decides an extent of measure effects on the first node on the basis of the measure implementation information, and then propagates the extent of measure effects on the second node from the first node as a starting point in the first graph.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a prediction device, a prediction method, and a program for predicting the effectiveness of a policy. [Background technology]

[0002] Patent Document 1 discloses a purchase prediction analysis system. The purchase prediction analysis system performs cluster analysis of regions using multiple factors and calculates the product purchase rate for each cluster. The purchase prediction analysis system decides whether to adopt or reject the calculated product purchase rate based on predetermined criteria. The purchase prediction analysis system generates a prediction model, which is a calculation formula for multiple regression analysis, using the product purchase rate of the region in the cluster for which the product purchase rate was adopted as the objective variable and the factor scores of that region as the explanatory variables. The purchase prediction analysis system uses the generated calculation formula to calculate and predict the product purchase rate for all regions from the explanatory variables for the region. This makes it possible to predict an appropriate product purchase rate even in regions where an appropriate product purchase rate cannot be obtained. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2016-6621 Summary of the Invention [Problem to be solved by the invention]

[0004] The present disclosure provides a prediction device, a prediction method, and a program for predicting effective measures. [Means for solving the problem]

[0005] The prediction device of the present disclosure includes a storage unit that stores policy implementation information that indicates the effectiveness of a policy in a first target for which a policy has been implemented, and a control unit that predicts the effectiveness of a policy in a second target for which a policy has not been implemented based on the policy implementation information. The control unit constructs a first graph that includes a plurality of nodes, including at least one first node associated with the first target and at least one second node associated with the second target, and a plurality of links that connect the nodes based on the similarities between the nodes, determines the degree of the policy effectiveness in the first node based on the policy implementation information, and propagates the degree of the policy effectiveness to the second node in the first graph, starting from the first node.

[0006] These general and specific aspects may be realized by a system, a method, and a computer program, as well as combinations thereof. [Effects of the Invention]

[0007] According to the prediction device, prediction method, and program disclosed herein, the degree of effectiveness of a policy is propagated using a graph including nodes and links, making it possible to predict effective policies for targets for which the policy has not been implemented. [Brief explanation of the drawings]

[0008] [Figure 1] Block diagram showing the configuration of a prediction device and a terminal device [Figure 2] FIG. 10 is a diagram showing the functional configuration of the control unit and data stored in the storage unit during learning. [Figure 3] An example of ID-POS data [Figure 4] An example of customer data [Figure 5] An example of district data [Figure 6] An example of purchase data before the implementation of the measures [Figure 7] An example of purchasing data during the implementation of the campaign [Figure 8] FIG. 1 is a diagram schematically illustrating an example of prediction data. [Figure 9] Flowchart showing the operation of the prediction device to generate prediction data [Figure 10] Flowchart showing the process of setting effect ranks for areas where implementation has been completed [Figure 11] A diagram to explain the calculation of the policy effectiveness value and the setting of effectiveness ranks for areas where the policy has been implemented [Figure 12] A diagram to explain the setting of rank probabilities for areas where implementation has been completed [Figure 13] A diagram to explain the construction of similarity graphs for each district characteristic [Figure 14] Flowchart showing the process of predicting the effectiveness rank for non-implemented areas [Figure 15] A diagram for explaining the construction of a prediction graph for each effect rank [Figure 16] A diagram for explaining the synthesis of similarity graphs based on importance. [Figure 17] Schematic diagram of the propagation of effect ranks [Figure 18] Diagram to explain the predicted effectiveness rank for non-implemented areas [Figure 19] A diagram showing the functional configuration of the control unit and data stored in the storage unit at the time of recommendation. [Figure 20] Flowchart showing the recommendation operation of the prediction device [Figure 21] Diagram to explain the completion of purchasing data in new areas [Figure 22] Diagram to explain how trade area effectiveness ranks are determined [Figure 23] A diagram explaining recommendations and non-recommendations according to effectiveness rank [Figure 24] Flowchart showing the operation of updating prediction data of the prediction device DETAILED DESCRIPTION OF THE INVENTION

[0009] (Findings that formed the basis of this disclosure) It is desirable to select and recommend an effective measure from among multiple candidate measures to a target for which the measure is to be implemented. The target for which the measure is to be implemented may be, for example, a region or a store. However, in a purchase prediction analysis system such as that disclosed in Patent Document 1, when the effect of a measure is known in only a few regions, it is difficult to predict the effect of the measure in a region where the effect is unknown. As a result, it is not possible to recommend an effective measure to a region where the measure is not being implemented.

[0010] The present disclosure provides a prediction device that can predict and recommend effective measures to other targets that have not implemented the measures, even if the targets that have actually implemented the measures are small.

[0011] (Embodiment) An embodiment will be described below with reference to the drawings. In this embodiment, an example will be described in which a store is targeted for implementing a measure and the measure is recommended to the store. In this embodiment, a trade area, which is an area capable of attracting customers to the store, includes one or more districts. The prediction device of this embodiment uses label propagation, a type of machine learning, to propagate an effectiveness rank indicating the degree of effectiveness of each measure from districts within the trade area of ​​a store that has implemented the measure to districts within the trade area of ​​a store that has not implemented the measure. This allows the degree of effectiveness of each measure in a store that has not implemented the measure to be predicted and highly effective measures to be recommended to the store. For example, based on the results of measures implemented within the trade area of ​​a small number of retail chain stores, a measure predicted to be highly effective out of multiple measures is recommended to a store in a different trade area. Examples of measures implemented in a store include POP, island displays, LED signs, in-store vision, receipt coupons, and point increases.

[0012] In this specification, an area within the trade area of ​​a store that has implemented a measure is also referred to as an "implemented area." An area within the trade area of ​​a store that has not implemented a measure, excluding the implemented area, is also referred to as an "unimplemented area."

[0013] 1. Configuration of prediction device and terminal device Fig. 1 shows the configuration of a prediction device 1 and a terminal device 2. The prediction device 1 and multiple terminal devices 2 constitute a prediction system 100. The prediction system 100 uses data from stores that have implemented measures to predict and recommend effective measures for stores that have not implemented measures.

[0014] The prediction device 1 is a server. The terminal device 2 is one of various information processing devices such as a POS (Point of Sales) register, a personal computer, a tablet terminal, and a smartphone. For example, the prediction device 1 is a cloud server, and the terminal device 2 is installed in a store. In this case, the prediction device 1 and the terminal device 2 are connected via the Internet.

[0015] The prediction device 1 includes an input unit 11, a communication unit 12, a control unit 13, a storage unit 14, and a bus 15.

[0016] The input unit 11 is a user interface for inputting various operations by the user, and can be realized by a touch panel, a keyboard, a button, a switch, or a combination of these.

[0017] The communication unit 12 includes a circuit for communicating with external devices in accordance with a predetermined communication standard. Examples of the predetermined communication standard include LAN, Wi-Fi (registered trademark), Bluetooth (registered trademark), USB, and HDMI (registered trademark). The communication unit 12 acquires data related to each store from terminal devices 2 located in multiple stores. In this embodiment, the acquired data related to the stores includes ID-POS data, customer data, and area data. The area data may be acquired from the terminal device 2 or from another external device. The communication unit 12 transmits measure recommendation information indicating the recommended measures to the terminal device 2.

[0018] The control unit 13 can be realized by a semiconductor element or the like. The control unit 13 can be configured by, for example, a microcomputer, a CPU, an MPU, a GPU, a DSP, an FPGA, or an ASIC. The functions of the control unit 13 may be configured by hardware alone, or may be realized by combining hardware and software. The control unit 13 realizes predetermined functions by reading data and programs stored in the storage unit 14 and performing various arithmetic processing.

[0019] The storage unit 14 is a storage medium that stores programs and data necessary to realize the functions of the prediction device 1. The storage unit 14 can be realized by, for example, a hard disk (HDD), an SSD, a RAM, a DRAM, a ferroelectric memory, a flash memory, a magnetic disk, or a combination of these.

[0020] The bus 15 is a signal line that electrically connects the input unit 11, the communication unit 12, the control unit 13, and the storage unit .

[0021] The terminal device 2 includes an input unit 21, a communication unit 22, a control unit 23, a storage unit 24, a display unit 25, and a bus 26.

[0022] The terminal device 2 acquires ID-POS data, customer data, and area data through the input unit 21 or the communication unit 22.

[0023] The input unit 21 can be realized by a barcode reader, a card reader, a touch panel, a keyboard, a button, a switch, or a combination thereof.

[0024] The communication unit 22 includes a circuit for communicating with external devices in accordance with a predetermined communication standard. Examples of the predetermined communication standard include LAN, Wi-Fi (registered trademark), Bluetooth (registered trademark), USB, and HDMI (registered trademark). The communication unit 22 transmits ID-POS data, customer data, and area data to the prediction device 1. The communication unit 22 acquires policy recommendation information from the prediction device 1.

[0025] The control unit 23 can be realized by a semiconductor element or the like. The control unit 23 can be configured by, for example, a microcomputer, a CPU, an MPU, a GPU, a DSP, an FPGA, or an ASIC. The functions of the control unit 23 may be configured by hardware alone, or may be realized by combining hardware and software. The control unit 23 realizes predetermined functions by reading data and programs stored in the storage unit 24 and performing various arithmetic processing.

[0026] The storage unit 24 is a storage medium that stores programs and data necessary to realize the functions of the terminal device 2. The storage unit 24 can be realized by, for example, a hard disk drive (HDD), an SSD, a RAM, a DRAM, a ferroelectric memory, a flash memory, a magnetic disk, or a combination of these.

[0027] The display unit 25 is, for example, a liquid crystal display or an organic EL display, and displays, for example, a recommendation statement for the measure indicated by the measure recommendation information.

[0028] The bus 26 is a signal line that electrically connects the input unit 21, the communication unit 22, the control unit 23, the storage unit 24, and the display unit 25.

[0029] 2. Behavior of the predictor during training 2.1 Functional configuration of the prediction device during training The functions of the prediction device 1 during learning will be described with reference to Figures 2 to 8. Figure 2 shows the functional configuration of the control unit 13 and the data stored in the memory unit 14 during learning of the prediction device 1. Figure 3 shows an example of ID-POS data 141. Figure 4 shows an example of customer data 142. Figure 5 shows an example of district data 143. Figure 6 shows an example of purchase data 144A before the implementation of a campaign. Figure 7 shows an example of purchase data 144B during the implementation of a campaign. Figure 8 schematically shows an example of prediction data 145. In this embodiment, districts a, b, c, etc. shown in Figure 8 correspond to districts such as "Moriguchi," "Minami-Kadoma," and "Kita-Kadoma" shown in Figures 5 to 7.

[0030] As shown in FIG. 2, the control unit 13 of the prediction device 1 includes a data aggregation unit 131, an effect rank setting unit 132, a prediction data generation unit 133, and an update determination unit .

[0031] The storage unit 14 of the prediction device 1 stores ID-POS data 141, customer data 142, and area data 143 acquired via the communication unit 12 from the terminal devices 2 in a plurality of stores.

[0032] The ID-POS data 141 is data that indicates sales of products. In the example of Fig. 3, the ID-POS data 141 includes the date and time when the product was purchased, the ID of the customer who purchased the product, the purchased product, the unit price and quantity of the product purchased, and the total amount of the purchased product.

[0033] The customer data 142 is data related to a customer. In the example of Fig. 4, the customer data 142 includes a customer ID, a gender, a postal code of the customer's residence, and a birth month.

[0034] The district data 143 is data indicating the characteristics of a district. In the example of Fig. 5, the district data 143 includes the postal code and name of the district, the population in the district, the number of households, and the male-female ratio of the population.

[0035] The data aggregation unit 131 aggregates the ID-POS data 141, the customer data 142, and the district data 143 to generate purchase data 144. The data aggregation unit 131 stores the generated purchase data 144 in the storage unit 14.

[0036] As shown in FIGS. 6 and 7, the purchase data 144 includes purchase data 144A before the implementation of the campaign and purchase data 144B during the implementation of the campaign. In the examples of FIGS. 6 and 7, the purchase data 144A and 144B include the postal code and name of the district, the population in the district, the number of households, the gender ratio of the population, sales in the district, and average customer spending. The data aggregation unit 131 calculates sales and average customer spending for districts within the store's trade area based on, for example, ID-POS data 141, customer data 142, and district data 143. When generating the purchase data 144, if each district is associated with the trade area of ​​multiple stores, i.e., if each district is associated with multiple stores, the sales of all stores associated with each district may be summed. When generating the purchase data 144, each district may be associated with only one store. The purchase data 144 is an example of campaign implementation information that indicates the effectiveness of the campaign. For example, the difference in sales between the purchase data 144A and 144B indicates the effectiveness of the campaign.

[0037] The effect rank setting unit 132 determines the effect rank of a measure for an implemented district, which is a district within the trade area of ​​a store where the measure has been implemented. Specifically, the effect rank setting unit 132 calculates a measure effect value based on purchase data 144A before the measure is implemented and purchase data 144B during the measure is implemented. The effect rank setting unit 132 compares the measure effect value with a predetermined threshold value to classify the degree of effect of the measure into multiple effect ranks. The multiple effect ranks are, for example, ranks A, B, C, and D.

[0038] The predicted data generation unit 133 predicts the effectiveness rank of a measure for an unimplemented area, which is an area within the trade area of ​​a store where the measure has not been implemented. Specifically, the predicted data generation unit 133 constructs a predicted graph for each measure and propagates the effectiveness rank from the node of the implemented area to the node of the unimplemented area using the label propagation method. As a result, the predicted data generation unit 133 generates predicted data 145 indicating the effectiveness rank of each measure for each area and stores it in the storage unit 14.

[0039] The prediction data 145 includes, for example, information indicating a prediction graph 450 as shown in FIG. 8. The information indicating the prediction graph 450 includes, for example, importance, which is a weighting used when combining multiple similarity graphs. The importance will be described in detail later. The prediction graph 450 is composed of nodes 451, links 452 connecting the nodes 451, and labels 453 assigned to the nodes 451. In this embodiment, the nodes 451 correspond to districts. The links 452 correspond to the similarities between districts. In other words, the links 452 have strengths corresponding to the similarities between districts. For example, the prediction data 145 includes information indicating the strengths of all the links 452 in the prediction graph 450. The labels 453 correspond to the effectiveness ranks of measures. The prediction data 145 includes, for example, data obtained by adding the effectiveness ranks of each measure to the district data 143, as shown in FIG. 22.

[0040] The update determination unit 134 determines whether to update the predicted data 145. For example, the update determination unit 134 determines whether to update the predicted data 145 based on a change in the effect rank for the implemented district set by the effect rank setting unit 132. The update determination unit 134 may determine whether to update the predicted data 145 based on at least one of the ID-POS data 141, the customer data 142, the district data 143, and the purchase data 144. When the update determination unit 134 determines to update the predicted data 145, it instructs the predicted data generation unit 133 to update the predicted data 145. As a result, the predicted data generation unit 133 reconstructs the predicted graph 450 and updates the predicted data 145.

[0041] 2.2 Overall operation during learning FIG. 9 shows the operation of generating predicted data by the control unit 13 of the prediction device 1.

[0042] The data aggregation unit 131 acquires ID-POS data 141, customer data 142, and area data 143 (S1). For example, the data aggregation unit 131 acquires ID-POS data 141 corresponding to a predetermined period before the campaign is implemented, and customer data 142 and area data 143 corresponding to the predetermined period, from the terminal devices 2 of multiple stores. Furthermore, the data aggregation unit 131 acquires ID-POS data 141 corresponding to a predetermined period during the campaign, and customer data 142 and area data 143 corresponding to the predetermined period, from the terminal devices 2 of the store that implemented the campaign. The predetermined period is, for example, one month. In step S1, the data aggregation unit 131 may read from the storage unit 14 the ID-POS data 141, customer data 142, and area data 143 that were previously acquired from the terminal devices 2 and stored in the storage unit 14.

[0043] The data aggregation unit 131 generates purchase data 144A before the implementation of the campaign and purchase data 144B during the implementation of the campaign based on the ID-POS data 141, customer data 142, and district data 143 (S2).

[0044] The effect rank setting unit 132 sets an effect rank for the measure for the district within the trade area of ​​the store that has already implemented the measure (S3).

[0045] The prediction data generation unit 133 calculates the similarity between districts for each district characteristic and generates a similarity graph (S4). District characteristics include, for example, population, number of households, male-female ratio of the population, sales, and average customer spending. For example, a population similarity graph, a number of households similarity graph, a male-female ratio similarity graph, a sales similarity graph, and an average customer spending similarity graph are generated (see FIG. 13).

[0046] The predicted data generation unit 133 generates a predicted graph by combining similarity graphs for each district characteristic and predicts the effectiveness rank of districts within the trade area of ​​stores that have not implemented the measures (S5). This generates predicted data 145.

[0047] The prediction device 1 performs the processes of steps S1 to S5 for each measure.

[0048] 2.3 Setting effectiveness rankings for implemented areas Setting of the effect rank of an implemented district, which is a district within the trade area of ​​a store where a measure has been implemented, will be described with reference to Figures 10 to 12. Figure 10 shows the operation of setting the effect rank of an implemented district (details of step S3 in Figure 9). Figure 11 shows an example of a measure effect value and an effect rank. Figure 12 shows an example of a rank probability.

[0049] The effect rank setting unit 132 calculates a measure effect value based on the purchase data 144A before the measure is implemented and the purchase data 144B during the measure is implemented (S301). For example, the measure effect value is calculated by "measure effect value = sales during the measure is implemented / sales before the measure is implemented × 100".

[0050] The effect rank setting unit 132 determines an effect rank based on the measure effect value (S302). For example, the effect rank setting unit 132 compares the measure effect value with three thresholds and assigns it one of ranks A, B, C, and D.

[0051] The effect rank setting unit 132 sets the probability of each effect rank according to the determined effect rank (S303). Specifically, the probability of a ranked effect rank is set to 1.0, and the probability of an unranked effect rank is set to 0. For example, as shown in Fig. 12, for Moriguchi, who has been determined to be effect rank A, the probability of rank A is set to 1.0, and the probabilities of ranks B, C, and D are set to 0.

[0052] 2.4 Construction of Similarity Graph FIG. 13 is a diagram illustrating the generation of a similarity graph 45 for each district characteristic in step S4. District characteristics include, for example, population, number of households, male-to-female ratio, sales before the implementation of the policy, and average customer spending. In step S4, the forecast data generation unit 133 generates, for example, a population similarity graph 45a, a number of households similarity graph 45b, a male-to-female ratio similarity graph 45c, a sales similarity graph 45d, and an average customer spending similarity graph 45e based on the purchase data 144. When no distinction is made between the population similarity graph 45a, the number of households similarity graph 45b, the male-to-female ratio similarity graph 45c, the sales similarity graph 45d, and the average customer spending similarity graph 45e, they are collectively referred to as a similarity graph 45. Each node 45N in each similarity graph 45 represents a district. Links 45L between the population similarity graph 45a, the number of households similarity graph 45b, the male-female ratio similarity graph 45c, the sales similarity graph 45d, and the average customer spending similarity graph 45e each have link strengths that represent the similarities between the population, number of households, male-female ratio, sales, and average customer spending.

[0053] Specifically, in step S4, the predicted data generating unit 133 calculates the link strength A kij is calculated using formula (1).

[0054]

number

[0055] In equation (1), k is the characteristics of the district such as population, sales, and average customer spending, and i and j are the nodes that represent the districts. kij is the strength of the link 45L between node i and node j in the similarity graph 45 for district characteristic k, and v ki is the value of characteristic k of node i, v kj is the value of characteristic k of node j, and σ is a positive definite parameter. kij Specifically, matrix A k The (i,j) component of the link strength A kij In the calculation of A k 1 n =1 nHere, n is the total number of districts, and A k is the link strength of the similarity graph 45 for the district characteristic k, and the link strength A k Specifically, v is a matrix. ki The process of normalizing with respect to k is called A kij This is performed before the calculation of the similarity graph in FIG. 16. As a result, the sum of each row 161 indicating the link strength in the similarity graph in FIG.

[0056] 2.5 Predicting the probability of effectiveness ranking in non-implemented areas With reference to Figures 14 to 18, the prediction of the effectiveness rank of non-implemented areas, which are areas within the trade areas of stores where no measures are being implemented, will be described. Figure 14 shows the operation of predicting the effectiveness rank of non-implemented areas (details of step S5 in Figure 9). Figure 15 is a diagram for explaining the generation of a prediction graph for each effectiveness rank. Figure 16 is a diagram for explaining the synthesis of similarity graphs 45. Figure 17 schematically shows the propagation of effectiveness ranks. In Figure 17, areas indicated by solid lines indicate areas for which an effectiveness rank has been determined, and areas indicated by dashed lines indicate areas for which an effectiveness rank has not been determined. Figure 18 is a diagram for explaining the prediction of the effectiveness rank for non-implemented areas.

[0057] The prediction data generation unit 133 performs steps S501 to S503 in FIG. 14 for each effectiveness rank. As a result, as shown in FIG. 15, for example, for measure 1, prediction graphs 450A, 450B, 450C, and 450D are constructed, each labeled with ranks A, B, C, and D. The nodes of prediction graphs 450A, 450B, 450C, and 450D are districts, and the links have strengths that combine the similarities between the districts' characteristics. In prediction graphs 450A, 450B, 450C, and 450D, the effectiveness ranks are propagated from districts where the measures have been implemented to districts where the measures have not been implemented. When there is no need to distinguish between prediction graphs 450A, 450B, 450C, and 450D, they are collectively referred to as prediction graph 450.

[0058] Specifically, the prediction data generation unit 133 calculates the importance of each characteristic of the area (S501). The prediction data generation unit 133 uses, for example, an EM (Expectation Maximization) algorithm to calculate the importance using equation (2). The EM algorithm consists of an E step and an M step. The E step calculates the plausible importance, and the M step updates the probability of the effect rank so that the expected value of the importance calculated in the E step is maximized.

[0059] (E step)

number

[0060] In equation (2), TIFF0007811719000003.tif88 is the updated importance of characteristic k, f is the predicted probability of the effect rank, L k is graph A k The graph Laplacian of ν,β net is a positive definite parameter, and n is the total number of districts.

[0061] Graph Laplacian L k can be calculated using equation (3).

[0062]

number

[0063] The prediction data generation unit 133 calculates the importance u k Based on the above, the respective similarity graphs 45 are combined to construct the predicted graph 450 (S502). Specifically, the predicted data generating unit 133 calculates the link strengths A of the predicted graphs 450A to 450D using the formula (4). int Calculate the link strength A int Specifically, is a matrix.

[0064]

number

[0065] For ease of explanation, Figure 16 shows the link strength A of the similarity graph 45 of population, number of households, and average customer spending. k From the predicted graph, the link strength A of 450 int As shown in Equation (4) and FIG. 16, the link strength A of the similarity graph 45 is calculated. k The importance of each district characteristic is calculated as u k By multiplying the values ​​by 1 and calculating the sum, the link strength A of the predicted graphs 450A to 450D is calculated. int are calculated respectively.

[0066] The predicted data generation unit 133 calculates the link strength A of the predicted graphs 450A to 450D. int In accordance with the above, the prediction data generation unit 133 calculates the probability of the effectiveness rank of a neighboring node of a node whose effectiveness rank has been determined (S503). That is, the prediction data generation unit 133 propagates the effectiveness rank to a neighboring node of the node whose effectiveness rank has been determined. For example, in a prediction graph 450A of effectiveness rank A shown in FIG. 17, the prediction data generation unit 133 calculates the probability of effectiveness rank A for a neighboring node of a node 451 whose effectiveness rank has been determined, based on the strength of a link 452. For example, the prediction data generation unit 133 calculates the probability of effectiveness rank A for districts a and e which are neighboring district d, and districts f and g which are neighboring district h.

[0067] Specifically, the prediction data generating unit 133 calculates the probability of the effect rank using equation (5).

[0068] (M step)

number

[0069] In equation (5), TIFF0007811719000007.tif96 is the predicted value of the probability of the effectiveness rank after updating, f is the predicted value of the probability of the effectiveness rank before updating, y is the probability of the effectiveness rank in the implemented area, G is the diagonal matrix for calculation, I n is the n-dimensional identity matrix, L int is the graph Laplacian, βy ,β bias ,β net is a positive definite parameter. The graph Laplacian L int is calculated using equation (6).

[0070]

number

[0071] The diagonal matrix G for calculation is as follows: where l is the number of districts where the program has been implemented, and n is the total number of districts.

[0072]

number

[0073] The EM algorithm uses equations (2) and (5) to repeatedly calculate predicted values ​​of the importance and effectiveness rank probabilities, and when the amount of change compared to the value before the update falls below a threshold, the predicted value of the effectiveness rank probability is determined.

[0074] 2.6 Determining the effectiveness rank of non-implemented areas The prediction data generation unit 133 determines whether or not calculation of the probabilities of all effect ranks has been completed for nodes adjacent to the node for which the effect rank has been determined (S504). For example, if all probabilities of effect ranks A, B, C, and D have not been calculated, the process returns to step S501, and steps S501 to S503 are performed for the effect ranks that have not been calculated. When all probabilities of effect ranks A, B, C, and D have been calculated, the process proceeds to step S505.

[0075] The prediction data generation unit 133 determines the effect rank for the neighboring node for which the probability of each effect rank has been calculated, based on the probability of each effect rank (S505). For example, as shown in Fig. 18, if the probabilities of effect ranks A, B, C, and D for the Kyobashi district are calculated as "0.8", "0.3", "0.4", and "0.1", the effect rank A with the highest probability of "0.8" is determined to be the effect rank for the Kyobashi district.

[0076] The prediction data generation unit 133 determines whether the effect ranks of all the districts in the prediction graph 450 have been determined (S506). If the effect ranks of all the districts have not been determined (No in S506), the process returns to step S501. As a result, as shown in FIG. 17, the effect ranks are propagated from the districts whose effect ranks have been determined to the districts whose effect ranks have not been determined.

[0077] When the effectiveness ranks of all the districts are determined (Yes in S506), the prediction data generation unit 133 stores the prediction data 145 in the storage unit 14 (S507). As described above, the prediction data 145 includes information indicating, for example, a prediction graph 450 as shown in FIG. 8, that is, prediction graphs 450A, 450B, 450C, and 450D. Specifically, for example, the prediction data 145 includes information indicating a district that is a node 451, a strength A of a link 452, int , importance u k , which contains the effect rank and the probability of the effect rank for each district, labeled 453.

[0078] As described above, the prediction device 1 performs steps S1 to S5 shown in Fig. 9 for each measure. The prediction device 1 performs steps S501 to S503 shown in Fig. 14 for each effectiveness rank. That is, the prediction device 1 generates prediction graphs 450 in numbers corresponding to "number of measures x number of ranks" from multiple similarity graphs 45, and calculates the probability of effectiveness rank for each effectiveness rank for each measure. The prediction device 1 determines the effectiveness rank with the highest probability for each district as the effectiveness rank for that district.

[0079] 3. Prediction device behavior during recommendation 3.1 Functional configuration of the prediction device for recommendation FIG. 19 shows the functional configuration of the control unit 13 and data stored in the storage unit 14 when the prediction device 1 makes a recommendation.

[0080] The control unit 13 of the prediction device 1 includes a trade area setting unit 135 , a purchase data complementing unit 136 , a prediction data updating unit 137 , and a measure recommendation unit 138 .

[0081] The commercial area setting unit 135 acquires, from the input unit 11 or the communication unit 12, store information indicating the target store for which the effect of a policy is to be predicted, and sets the commercial area.

[0082] The purchase data complementation unit 136 calculates the sales and average customer spending of the new store based on the sales and average customer spending of the existing store, and complements the purchase data 144. The forecast data update unit 137 uses the complemented purchase data 144 to reconstruct the forecast graph 450 and update the forecast data 145.

[0083] The measure recommendation unit 138 determines a measure to be recommended to the store based on the prediction data 145. The measure recommendation unit 138 transmits measure recommendation information indicating the determined measure to the terminal device 2 via the communication unit 12.

[0084] 3.2 Recommendation behavior Recommendation of measures will be described with reference to Figs. 20 to 23. Fig. 20 shows the operation of the control unit 13 of the prediction device 1 when making a recommendation. Fig. 21 shows an example of complementing the purchase data 144. Fig. 22 is a diagram for explaining the determination of the effectiveness rank of a commercial area. Fig. 23 shows an example of recommending and not recommending measures according to the effectiveness rank.

[0085] 20, when the commercial area setting unit 135 acquires store information indicating the target store for which the policy effect is to be predicted from the input unit 11 or the communication unit 12, it sets the commercial area of ​​that store based on the store information (S601). The commercial area setting unit 135 identifies the residential district of store visitors based on, for example, the ID-POS data 141 and customer data 142 acquired from the target store for prediction, and sets the commercial area. The commercial area setting unit 135 may set the commercial area within a predetermined distance from the target store for prediction.

[0086] The trade area setting unit 135 determines whether the districts within the set trade area include a new district (S602). A new district is a district not included in the prediction data 145. For example, if the store being predicted is a new store that will open soon, the effectiveness rank of the districts within the trade area of ​​the new store is not included in the prediction data 145 generated during learning. In this case, the new district is included in the districts within the trade area set in step S601. If the new district is included in this way (Yes in S602), proceed to step S603. If the effectiveness rank of the districts within the set trade area is included in the prediction data 145, that is, if the new district is not included (No in S602), proceed to step S606.

[0087] The purchasing data complementing unit 136 acquires district data 143 for the new district (S603). For example, the purchasing data complementing unit 136 acquires district data 143 for the new district from the terminal device 2 or another external device and stores it in the storage unit 14. Alternatively, the purchasing data complementing unit 136 reads district data 143 for the new district that was previously acquired and stored in the storage unit 14 from the storage unit 14. The purchasing data complementing unit 136 complements the purchasing data 144 based on the district data 143 for the new district and the purchasing data 144 for the existing district (S604). For example, as shown in FIG. 21 , the purchasing data complementing unit 136 adds data 43 included in the district data 143 for the new district to the purchasing data 144, and calculates sales and average customer spending 44B for the new district based on sales and average customer spending 44A for the existing district. For example, the purchasing data complementing unit 136 sets the average values ​​of sales and average customer spending for the existing district to the sales and average customer spending for the new district. The purchase data supplementation unit 136 may calculate the sales and average spending per customer in the new district from the sales and average spending per customer in the existing district by regression analysis.

[0088] The forecast data update unit 137 reconstructs the forecast graph 450 for each measure using the supplemented purchase data 144 (S605). The forecast data update unit 137 updates the forecast data 145 in the storage unit 14 based on the reconstructed forecast graph 450. Step S605 of reconstructing the forecast graph 450 corresponds to steps S4 and S5 in FIG. 9.

[0089] The policy recommendation unit 138 predicts the effectiveness ranks of all policies in the commercial area based on the prediction data 145 (S606). For example, if the commercial area set in step S601 includes multiple districts, the population of each district is tallied for each effectiveness rank for each policy, and the effectiveness rank with the highest population is set as the effectiveness rank for the commercial area. Specifically, for example, if a commercial area P including "Moriguchi," "Minami-Kadoma," "Kita-Kadoma," "Hirakata," and "Kyobashi" shown in FIG. 22 is set in step S601, when determining the effectiveness rank of policy 1 for commercial area P, the total population for each effectiveness rank of policy 1 is first calculated. For policy 1, the population for effectiveness rank A is 1500 (=1200+300), the population for effectiveness rank B is 1100 (=1000+100), the population for effectiveness rank C is 0, and the population for effectiveness rank D is 600. Therefore, for policy 1, the effectiveness rank A with the highest population is set as the effectiveness rank for commercial area P. When determining the effectiveness rank of measure 2 for trade area P, the total population for each effectiveness rank of measure 2 is calculated. For measure 2, the population for effectiveness rank A is 300, the population for effectiveness rank B is 1200, the population for effectiveness rank C is 1600 (=1000 + 600), and the population for effectiveness rank D is 100. In this case, for measure 2, effectiveness rank C, which has the largest population, is set as the effectiveness rank of trade area P. If the trade area set in step S601 includes only one district, for example, in the example of Figure 22, if trade area Q is set, the effectiveness rank of the Minojima district included in trade area Q is set as the effectiveness rank of trade area Q.

[0090] The policy recommendation unit 138 determines the policy to be recommended according to the predicted effect rank (S607). For example, as shown in FIG. 23, the policy recommendation unit 138 determines the policies with effect ranks A and B as the policies to be recommended. The policy recommendation unit 138 transmits policy recommendation information indicating the determined policies to the terminal device 2 of the store that is the prediction target and that was specified in step S601 via the communication unit 12. For example, the policy recommendation information includes a recommendation statement for the policy. The terminal device 2 displays the recommendation statement on the display unit 25 based on the policy recommendation information.

[0091] As described above, the prediction device 1 uses the label propagation method to propagate the effectiveness rank of a measure from areas within the trade area of ​​a store that has implemented the measure to areas within the trade area of ​​a store that has not implemented the measure. Based on the propagated effectiveness rank of the measures, the prediction device 1 predicts the effectiveness rank of all measures in the trade area of ​​the store being predicted. Based on the predicted effectiveness rank, the prediction device 1 determines which measures to recommend. This makes it possible to recommend effective measures for the store being predicted, even if only a few stores have implemented the measure.

[0092] 4. Update forecast data FIG. 24 shows an example of an update operation of the prediction data 145 by the control unit 13 of the prediction device 1. When generating the prediction data 145 for the first time, the prediction device 1 performs the process shown in FIG. 9. When the prediction data 145 is already stored in the storage unit 14, the prediction device 1 performs the process shown in FIG. 24. Steps S11, S12, S13, S16, and S17 in FIG. 24 perform the same processes as steps S1, S2, S3, S4, and S5 in FIG. 9, respectively. In this embodiment, steps S14 and S15 determine whether or not the prediction data 145 needs to be updated. Steps S14 and S15 are an example of an update determination.

[0093] The data aggregation unit 131 newly acquires ID-POS data 141, customer data 142, and district data 143 (S11). Based on the newly acquired ID-POS data 141, customer data 142, and district data 143, the data aggregation unit 131 generates purchase data 144A before the implementation of the campaign and purchase data 144B during the implementation of the campaign (S12). The effect rank setting unit 132 sets the current effect rank of the campaign for the district within the trade area of ​​the store where the campaign has been implemented (S13). The update determination unit 134 compares the current effect rank with the past effect rank (S14). The update determination unit 134 determines whether the effect rank has changed (S15). For example, the update determination unit 134 determines whether the average effect rank of all districts is different between the past and the present. If the average effect rank of all districts has changed, the update determination unit 134 determines that the prediction data 145 needs to be updated and proceeds to step S16. The predicted data generation unit 133 calculates the similarity between districts for each district characteristic and generates a similarity graph 45 (S16). The predicted data generation unit 133 reconstructs a predicted graph 450 by combining the similarity graphs 45 for each district characteristic, and predicts the effectiveness rank of districts within the trade area of ​​stores that have not implemented the measures (S17). This updates the predicted data 145.

[0094] Specifically, when the update determination unit 134 determines that the prediction data 145 should be updated, the prediction data generation unit 133 reconstructs the prediction graph 450 by reapplying the calculation described above in "2.5 Predicting the probability of the effectiveness rank in unimplemented areas" to update the predicted values ​​of the importance and the probability of the effectiveness rank.

[0095] As described above, by updating the prediction data 145 when the effectiveness rank changes, it becomes possible to make recommendations that are suited to the season, for example.

[0096] 5. Effects and Supplements The prediction device 1 of this embodiment includes a storage unit 14 that stores purchase data 144 representing the effects of a campaign implemented at a store, and a control unit 13 that predicts the effects of a campaign implemented at a store based on the purchase data 144. The store that implemented the campaign is an example of a first target that has implemented the campaign. The store that has not implemented the campaign is an example of a second target that has not implemented the campaign. The purchase data 144 is an example of campaign implementation information representing the effects of the campaign. The control unit 13 constructs prediction graphs 450A-450D, each of which includes a plurality of nodes, each of which includes at least one district node associated with the first target store and at least one district node associated with the second target store, and a plurality of links connecting the nodes based on the similarity between the nodes. The at least one district node associated with the first target store is an example of a first node. The at least one district node associated with the second target store is an example of a second node. The prediction graphs 450A-450D are examples of first graphs. The control unit 13 determines the degree of effectiveness of a measure at a node in an implemented area based on the purchase data 144, and propagates the degree of effectiveness of the measure to nodes in unimplemented areas in the prediction graphs 450A-450D, starting from the node in the implemented area. This makes it possible to predict effective measures in unimplemented areas even if there are few stores that have implemented the measure. Since the label propagation method propagates the degree of effectiveness of a measure in descending order of similarity between areas where the measure has been implemented and those where the measure has not been implemented, it is possible to predict effective measures for those unimplemented areas even if the similarity between the implemented area and the unimplemented area is low.

[0097] Specifically, the degree of the policy effectiveness includes multiple effectiveness ranks, and the control unit 13 propagates each effectiveness rank to the second node and calculates the probability of each effectiveness rank at the second node. The control unit 13 determines the effectiveness rank with the highest probability among the multiple effectiveness ranks as the effectiveness rank of the second node. This makes it possible to accurately predict the effectiveness rank of the policy for each district.

[0098] The multiple nodes are related to multiple characteristics. The multiple characteristics are, for example, population, average customer spending, male-female ratio, and sales. The control unit 13 generates a similarity graph 45 for each characteristic, which is composed of multiple nodes and multiple links connecting the nodes based on the similarity of each characteristic. The similarity graph 45 is an example of a second graph. The control unit 13 calculates the importance of each characteristic for each rank, and combines the similarity graphs 45 for each characteristic into one based on the importance to generate prediction graphs 450A to 450D. As a result, the effectiveness rank is propagated according to the characteristics of the area, making it possible to accurately predict the effectiveness rank of measures for each area.

[0099] The control unit 13 determines whether to recommend a measure to the second target store based on the degree of effectiveness of the measure at the second node. For example, the control unit 13 recommends measures ranked A and B to the store. This allows only effective measures to be recommended to the store.

[0100] If the store to be predicted is associated with a trade area that includes two or more districts, the control unit 13 determines the effectiveness rank of the trade area according to the effectiveness rank of each district within the trade area. A trade area is an example of a group. The control unit 13 determines whether to recommend a measure to the store to be predicted according to the effectiveness rank of the trade area. This makes it possible to recommend effective measures to the store.

[0101] The control unit 13 generates a prediction graph 450 for each of the multiple measures, propagates the degree of measure effectiveness, and determines which measure to recommend from among the multiple measures based on the degree of measure effectiveness. This makes it possible to recommend an effective measure from among the multiple measures to the store.

[0102] (Other embodiments) As described above, the above embodiment has been described as an example of the technology disclosed in the present application. However, the technology in the present disclosure is not limited to this, and can be applied to embodiments in which appropriate modifications, substitutions, additions, omissions, etc. are made. Therefore, other embodiments will be described below as examples.

[0103] In the above embodiment, an example has been described in which the terminal device 2 is connected to the prediction device 1 via the Internet. However, the prediction device 1 may be installed in each store together with the terminal device 2 and connected to the terminal device 2. In the above embodiment, the prediction system 100 is configured by the prediction device 1 and the terminal device 2. However, all of the functions of the prediction system 100 may be realized by a single device. Some of the functions of the prediction device 1 described in the above embodiment may be performed by a separate prediction device. For example, a prediction device including the data aggregation unit 131, effect rank setting unit 132, predicted data generation unit 133, and update determination unit 134, which are functions during learning, and a prediction device including the trade area setting unit 135, purchase data complementation unit 136, predicted data update unit 137, and measure recommendation unit 138, which are functions during recommendation, may be separate devices.

[0104] In the above embodiment, the effectiveness rank for the implemented area was set based on the change in sales before and during the implementation of the campaign. However, the effectiveness rank may be set using other methods. For example, the effectiveness rank may be set based on any one of the following before and during the implementation of the campaign: sales, number of store visitors, average customer spending, reach rate, purchase rate, and store visit rate. Here, reach rate = number of people reaching the shelf / number of store visitors, purchase rate = number of product purchasers / number of store visitors, and store visit rate = number of store visitors in the area / area population. Furthermore, instead of using the actual values ​​of sales, number of store visitors, average customer spending, etc. obtained from the store, values ​​obtained by adjusting the obtained values ​​using a seasonal adjustment method or the like may be used. For example, the values ​​may be adjusted using the Census Bureau method, the MITI method, the monthly average method, the Parsons method, or the 12-month moving average method.

[0105] In the above embodiment, an example has been described in which, in the prediction graph 450 that propagates the effectiveness rank of a measure, the node 451 is a district, and the link 452 is the similarity of the district's population, number of households, average customer spending, etc. However, the node 451 and the link 452 are not limited to the above embodiment. For example, the node 451 may be a store, and the link 452 may be the similarity of sales between stores or sales ratios for each category. The node 451 may be a customer, and the link 452 may be the similarity of the customer's purchase amount or purchase ratio for each category.

[0106] In the above embodiment, when making a recommendation, the total population for each effect rank was calculated, and the effect rank with the largest population was set as the effect rank of the commercial area. However, the criteria for setting the effect rank of a commercial area are not limited to the total population. For example, the number of households, sales volume, etc. may be added up for each effect rank, and the effect rank with the largest value may be set as the effect rank of the commercial area.

[0107] In the above embodiment, the past effect rank of the area where the policy was implemented is compared with the current effect rank to determine whether the forecast data 145 needs to be updated. However, the update determination is not limited to the above embodiment. The update determination unit 134 may determine whether the forecast data 145 needs to be updated based on at least one of the newly acquired ID-POS data 141, customer data 142, and area data 143. The update determination unit 134 may determine whether the forecast data 145 needs to be updated based on newly generated purchase data 144. For example, the update determination unit 134 may reconstruct the forecast graph 450 and update the forecast data 145 when the purchase rate of a specific product or product category changes by more than a predetermined value. The update determination unit 134 may determine whether the forecast data 145 needs to be updated when the policy effect value changes by more than a predetermined value. The update determination unit 134 may generate data representing the economic situation, such as sales volume, from the purchase data 144, and reconstruct the forecast graph 450 and update the forecast data 145 when an economic change is detected. The update determination unit 134 may exclude areas with a bias in purchased products from the construction of the prediction graph 450. The prediction graph 450 may be reconstructed when the proportion of occupations or the proportion of foreign residents in the area changes. Depending on the characteristics of the residents of the area, for example, areas with a high proportion of specialized occupations may be excluded from the reconstruction of the prediction graph 450. The necessity of an update may be determined based on the amount of fluctuation in residents' usage time by media. Examples of usage time by media include average TV viewing time, internet usage time, smartphone usage time, and newspaper subscription rate. The update determination unit 134 may determine the necessity of an update based on time information indicating the month, season, day of the week, or year. For example, the prediction graph 450 may be reconstructed monthly. The update determination unit 134 may update the prediction data 145 at a timing specified by the user via the input unit 11 or the communication unit 12. The prediction data 145 may be updated based on a prediction of the effect of sales of seasonal products, etc., and the reconstruction of the prediction graph 450 may be terminated when the prediction error becomes smaller than a predetermined threshold. Reconstructing the prediction graph 450 in accordance with the seasonality enables seasonal recommendations.

[0108] In the above embodiment, an example has been described in which a prediction graph 450 predicting the effect of a measure for each district is generated and a measure is recommended to a store, but the configuration of the prediction graph 450 and the items recommended are not limited to those in the above embodiment. As shown below, a prediction graph with a different configuration may be generated. Items other than the measures may also be recommended. Variation 1: A prediction graph may be generated for customers, with links based on similarities such as customer purchase amounts or purchase ratios by category, and product purchasing trends as labels. This prediction graph may be used to recommend products at retail stores. Variation 2: A prediction graph may be generated for products, with links based on similarities such as sales, sales volume, and purchase rate, and labels based on the effectiveness of measures for each product. This prediction graph may be used to recommend measures for category areas such as fruit and vegetables and beverages at retail stores. Variation 3: For a factory or logistics base (logistics sorting facility), a prediction graph may be generated in which links are derived from demographics such as employee age and gender, the age of the building, climatic conditions, equipment specifications, and other similarities, and labels are derived from the rank of the effect of changes in production efficiency or throughput. This prediction graph may be used to recommend measures to improve operational efficiency at the factory or logistics base. Measures to improve operational efficiency include, for example, layout changes and work systems. Variation 4: A prediction graph may be generated for a department, with links based on similarities in the age, gender, length of service, etc. of employees within the department, and labels representing the effectiveness of changes in work efficiency when a system is updated. This prediction graph may be used to recommend updates to information systems in companies or local governments. Variation 5: For entertainment facilities, a prediction graph may be generated in which links are based on similarities in visitor age, gender, nationality, number of visitors, etc., and labels are based on effectiveness rankings based on changes in visitor attraction rates or sales increase rates. This prediction graph may be used to recommend measures to entertainment facilities such as zoos or aquariums. Variation 6: A prediction graph may be generated in which the target is a resident or a town, the population, the gender ratio of the population, or the number of accidents, etc., are used as links, and the effectiveness of accident prevention campaigns or crime prevention campaigns are used as labels. This prediction graph may be used to recommend measures to companies or local governments. Other variations: Instead of measures, recommendations may be made to a) driving methods for car drivers, b) schools to take entrance exams at, c) travel destinations or travel plans based on residential characteristics, d) news sites based on application usage rates, e) commercial placements or content based on viewer characteristics, f) advertisement placements or content based on passerby characteristics, g) exercise based on daily exercise characteristics or geographical information, or h) behaviors such as eating, sleeping, or walking based on physical condition, biosignals, or the surrounding environment.

[0109] (Outline of the embodiment) (1) A prediction device disclosed herein includes a memory unit that stores policy implementation information representing the effectiveness of a policy in a first target for which a policy has been implemented, and a control unit that predicts the effectiveness of a policy in a second target for which a policy has not been implemented based on the policy implementation information. The control unit constructs a first graph consisting of a plurality of nodes including at least one first node associated with the first target and at least one second node associated with the second target, and a plurality of links connecting the nodes based on the similarity between the nodes. The control unit determines the degree of policy effectiveness in the first node based on the policy implementation information, and propagates the degree of policy effectiveness to the second node in the first graph using the first node as a base point.

[0110] This makes it possible to predict effective measures for second targets to which no measures have been implemented, even if the number of first targets to which measures have been implemented is small.

[0111] (2) In the prediction device of (1), the control unit may determine whether to recommend the measure to the second target based on the degree of effectiveness of the measure at the second node.

[0112] This allows for effective measures to be recommended.

[0113] (3) In the prediction device of (1) or (2), the control unit may determine whether or not the prediction needs to be updated based on at least one characteristic among the multiple characteristics, and if it decides to update, may recalculate the link strength of the first graph and re-determine the degree of effectiveness of the policy.

[0114] (4) In any of the prediction devices (1) to (3), the degree of effectiveness of a policy may include multiple ranks, and the control unit may propagate each rank to the second node, calculate the probability of each rank at the second node, and determine the rank with the highest probability among the multiple ranks as the rank of the second node.

[0115] This makes it possible to predict effective measures based on the rank, making it possible to recommend multiple measures.

[0116] (5) In the prediction device of (4), the plurality of nodes are related to a plurality of characteristics, and the control unit may generate a second graph for each characteristic, the second graph being composed of a plurality of nodes and a plurality of links connecting the nodes based on the similarity of each characteristic, calculate the importance of each characteristic for each rank, and combine the second graphs for each characteristic into one based on the importance to generate a first graph.

[0117] This allows ranks to be propagated based on the similarity of multiple properties.

[0118] (6) In the prediction device of (4), when the second target is associated with a group including two or more nodes, the control unit may determine the rank of the group according to the rank of each node in the group, and may determine whether to recommend a measure to the second target according to the rank of the group.

[0119] This makes it possible to recommend an effective measure when the second target is related to multiple nodes.

[0120] (7) In any of the prediction devices (1) to (4), the first target and the second target may be a store, and the node may correspond to any of the store, a district within the store's trade area, or a customer visiting the store.

[0121] This makes it possible to predict effective measures for stores.

[0122] (8) In the prediction device of (7), the nodes correspond to districts, and the similarity between the nodes may be similarity regarding at least one of the population, number of households, gender ratio of the population, sales, and average customer spending within the district.

[0123] (9) In the prediction device of (8), the control unit may determine the degree of effectiveness of the measures at the first node based on the difference between before and during the implementation of the measures in at least one of sales, number of visitors, and average customer spending.

[0124] (10) In the prediction device of (8), the trade area of ​​the second target store includes one or more districts, and the control unit may determine the degree of effectiveness of the measures in the trade area according to the degree of effectiveness of the measures in the districts included in the trade area, and may determine whether to recommend measures to the second target store according to the degree of effectiveness of the measures in the trade area.

[0125] (11) In the prediction device of (2), the control unit may generate a first graph for each of a plurality of measures, propagate the degree of measure effectiveness, and determine a measure to recommend from among the plurality of measures based on the degree of measure effectiveness.

[0126] This makes it possible to recommend multiple measures.

[0127] (12) In any of the prediction devices (1) to (4), the first target and the second target may be a factory, and the multiple nodes may correspond to any of the factory, the age of the building, weather conditions, equipment specifications, and employees working in the factory.

[0128] (13) In any of the prediction devices (1) to (4), the first target and the second target may be a logistics base, and the multiple nodes may correspond to any of the logistics base, the age of the building, weather conditions, equipment specifications, and employees working at the logistics base.

[0129] (14) The prediction method disclosed herein is a prediction method that uses a calculation unit to predict the effectiveness of a measure in a second target for which a measure has not been implemented, based on measure implementation information that represents the effectiveness of the measure in a first target for which the measure has been implemented. The prediction method includes the steps of: constructing a graph (S502) consisting of a plurality of nodes including at least one first node associated with the first target and at least one second node associated with the second target, and a plurality of links connecting the nodes based on the similarity between the nodes; determining the degree of the measure effectiveness in the first node based on the measure implementation information; and propagating the degree of the measure effectiveness to the second node in the graph, starting from the first node (S503).

[0130] The prediction device and prediction method described in all claims of the present disclosure are realized by the cooperation of hardware resources, such as a processor, a memory, and a program. [Industrial Applicability]

[0131] The prediction device of the present disclosure is useful, for example, as a device that recommends effective measures to stores that are not implementing measures. [Explanation of symbols]

[0132] 1 Prediction device 2. Terminal Device 11,21 Input section 12,22 Communications Department 13,23 Control section 14,24 Storage section 15,26 Bus 25 Display section 100 Prediction System 131 Data Collection Department 132 Effect Rank Setting Section 133 Prediction Data Generation Unit 134 Update determination section 135 Trade Area Setting Section 136 Purchasing Data Completion Department 137 Forecast Data Update Unit 138 Policy Recommendation Department

Claims

1. A prediction device including a storage unit and a control unit, The storage unit policy implementation information including numerical data indicating a state before the policy and numerical data indicating a state during the policy with respect to a first target for which the policy has been implemented; statistical data associated with each of the first subject and a second subject to which the measure is not implemented; Store each of the plurality of objects being associated one-to-one with one of the plurality of nodes; the plurality of objects includes the first object associated with a first node and the second object associated with a second node; each of the plurality of nodes is connected to one or more other nodes by a link; The control unit determining a rank of the first target among a plurality of ranks representing the degree of the measure effect by comparing a measure effect value calculated based on a difference between numerical data indicating the state before the measure and numerical data indicating the state during the measure, which is included in the measure implementation information, with a predetermined threshold value; calculating the probability of each rank in the second node associated with the second object by performing a calculation using a label propagation method in which the rank of the first object is used as the label of the first node associated with the first object and the similarity between the statistical data related to each of the plurality of objects is used as the link strength between the plurality of nodes; Prediction device.

2. the control unit determines to recommend the policy to the second target when a rank of the second node determined by a probability of each rank in the second node is higher than a predetermined standard; outputting the measure as a recommended measure; The prediction device according to claim 1 .

3. the control unit determines that re-execution of the calculation by the label propagation method is necessary when the fluctuation of the statistical data is equal to or greater than a predetermined value, and when deciding to re-execute, re-calculates the link strength based on the statistical data after the fluctuation, and re-executes the calculation by the label propagation method based on the recalculated link strength, thereby re-determining the rank of the second node. The prediction device according to claim 1 or 2.

4. the first target and the second target are stores, The plurality of nodes correspond to any of the store, a district within the trade area of ​​the store, and customers who visit the store. The prediction device according to any one of claims 1 to 3.

5. The plurality of nodes are associated with the district; The link strength between the nodes is at least one of link strengths of population, number of households, male-to-female ratio, sales, and average customer spending between the district that is the first node and the district that is the second node. The prediction device according to claim 4 .

6. the control unit determines a difference between a first target and a second target in at least one of sales, number of customers, and average customer spending as a degree of the measure effectiveness of the first node; The prediction device according to claim 5 .

7. The trade area of ​​the second target store includes one or more districts; the control unit determines the degree of the policy effect of the commercial area by propagating the degree of the policy effect of the district included in the commercial area; For each node associated with each of the plurality of districts included in the commercial area, the rank with the highest probability among the plurality of ranks is determined as the effectiveness rank of each node; Selecting a node showing a rank with the highest probability from the effect ranks determined for each node; determining that the measure to be taken on the selected node is to be recommended to the second target store; The prediction device according to claim 5 .

8. the control unit determines the rank with the highest probability among a plurality of ranks as the rank of the second node; outputting the degree of the policy effect represented by the rank of the second node as a predicted result of the policy effect on the second target; The prediction device according to claim 1 .

9. the first object and the second object are factories; The plurality of nodes are associated with any of the factory, the age of the building, weather conditions, equipment specifications, and employees working in the factory. The prediction device according to any one of claims 1 to 3.

10. the first target and the second target are logistics bases, The plurality of nodes are associated with any of the logistics base, the age of the building, the weather conditions, the specifications of the equipment, and the employees working at the logistics base. The prediction device according to any one of claims 1 to 3.

11. A method for predicting the effect of a measure, which is executed by a computer including a storage unit and a control unit, The storage unit policy implementation information including numerical data indicating a state before the policy and numerical data indicating a state during the policy with respect to a first target for which the policy has been implemented; statistical data associated with each of the first subject and a second subject to which the measure is not implemented; Store each of the plurality of objects being associated one-to-one with one of the plurality of nodes; the plurality of objects includes the first object associated with a first node and the second object associated with a second node; each of the plurality of nodes is connected to one or more other nodes by a link; a step of determining a rank of the first target among a plurality of ranks representing the degree of the measure effect by the control unit by comparing a measure effect calculated based on a difference between numerical data indicating the state before the measure and numerical data indicating the state during the measure, which is included in the measure implementation information, with a predetermined threshold value; a step of calculating a probability of each rank in a second node associated with the second object by performing a calculation using a label propagation method in which the rank of the first object is used as a label of a first node associated with the first object and the similarity between the statistical data related to a plurality of objects is used as a link strength between the nodes, by the control unit; A prediction method, including:

12. A program that causes a computer to execute the prediction method described in claim 11.

Images