JP2021131640A - Feature extraction device and feature extraction method - Google Patents

Abstract

To perform exhaustive analysis of a supply chain without omission in consideration of consumption behavior of consumers without persons' understanding of business features.SOLUTION: A feature extraction device 1 is, for example, a computer comprising a CPU, a memory, and so on and comprises: a data collaboration platform 2 in which data relating to purchase, sales and consumption activities between a plurality of affiliated companies as well as purchase, sales, and consumption of consumers are stored is stored; and an analysis engine device 3 which combines two of pieces of data stored in the data collaboration platform 2 to construct a matrix structure and constructs a neural network showing a relation between the matrix structure and key performance indicator data 42 to perform analysis.SELECTED DRAWING: Figure 1

Description

The present invention relates to a feature amount extraction device and a feature extraction method utilizing a data linkage platform. Specifically, it is a key performance indicator (KPI) that is sales, etc., based on all the data that can be obtained in the process from product production to distribution, reaching consumers through stores, and being consumed. : Key Performance Indicator) is related to the technology that can extract the feature quantity to improve and efficiently recommend the business improvement measures.

The connection between raw material manufacturers, manufacturers, wholesalers, and retailers, including retailers such as supermarkets and convenience stores, is generally understood by the concept of supply chain. And in the case of BtoC (Business to Consumer) business, the supply chain is ultimately connected to consumers.

FIG. 13 is a diagram showing an outline of the supply chain.
The supply chain is composed of raw material manufacturers / producers 90, product manufacturers 91, trading companies / wholesalers 92, retailers 93, and consumers 94, respectively. The raw material maker / producer 90 receives the production plan from the product maker 91, and sends the procurement distribution to the product maker 91 according to the production plan.

The product maker 91 receives the sales plan from the trading company / wholesale 92, and consolidates the products to the co-delivery center of the trading company / wholesale 92 according to the sales plan. “Optimization of SCM (Supply Chain Management)” indicated by the arrow means improving the flow of goods from the raw material maker / producer 90 to the trading company / wholesale 92.
The trading company / wholesaler 92 receives the expected order from the retailer 93 and delivers the product to the retailer 93 stores accordingly. The “efficiency of store / logistics operations” indicated by the arrows means improving the flow of orders from retail 93 to product maker 91 and the flow of products from product maker 91 to retail 93.
The retailer 93 senses the demand / preference of the consumer 94, and sells the product by delivering the product to the doorway of the consumer 94, for example, according to the demand / preference. Here, improvement of the level of delivery service from retail 93 to consumer 94 and measures represented by advertisements and campaigns are issues.

Marketing is about improving this entire supply chain. Coordination of physical distribution (inventory) information and efficiency of business processes are to streamline the flow of goods from raw material manufacturers / producers 90 to retail 93.

At present, the flow of goods and money is that each company only owns its own data. Therefore, it is difficult for the retail 93 and the raw material manufacturer / producer 90 to cooperate with each other and for the raw material manufacturer / producer 90 to hold the data of the retail 93. Therefore, although there is a concept of supply chain, it is difficult to actually manage the connection of the entire supply chain.

When deriving the required amount of procurement and the amount of marketing implemented within a company, it was often the case that different models were created and the required amount was derived in order to match the nature of the required amount to be implemented and the feature amount to be analyzed. Here, as a method of creating a model, for example, an ARIMA (Autoregressive Integrated Moving Average) model is often applied in product demand forecasting, and a sales tendency analysis is often applied in marketing.

When marketing a certain product, it is considered that the sales volume of this product will increase more than before when marketing is not carried out. An increase in the sales volume of this product causes, for example, an increase in the order volume of this product or its raw materials. Changes that occur in any of the supply chains in this way are each relevant to the entire supply chain. Therefore, changes in the supply chain should be immediately reflected in the model of the entire supply chain.
However, at present, since a model is created for each company that constitutes the supply chain, there is a problem that changes in the supply chain are not reflected until the next modeling is carried out.

Therefore, inventions as described in Patent Document 1 have been published. The invention described in Patent Document 1 is an inexpensive system that can be used for various businesses in general, and can efficiently optimize a large-scale corporate collaboration type without requiring a high degree of specialized knowledge. It is a company-linked optimization support system.

Japanese Unexamined Patent Publication No. 2004-127259

Claim 1 of Patent Document 1 states that "a system that supports corporate-linked optimization in order to improve corporate value in a business that provides products to customers, and is subject to (1) optimization. Means for describing and designing business processes according to the type of corporate collaboration, (2) means for inputting restrictions on procurement, production, delivery, and sales of products and intermediate products, resources, and cost data, and (3) the above. A company-linked optimization support system characterized by having a means for carrying out "P1 optimization" based on the above-mentioned business process while considering the constraint conditions, resources, and cost data of the above. " ..

However, in the method of Patent Document 1, it takes time and man-hours to understand the processes before and after. Also, it is difficult for a person to notice all the constraints and describe the constraint equations. Furthermore, it requires specialized knowledge to understand the characteristics of business and construct constraint expressions, and it is difficult to automate by the system.

In addition, the supply chain can be further improved by considering the consumption behavior of consumers and the consumption data of consumers. For example, it is conceivable to take measures in consideration of the consumer's consumption behavior, such as advertising the product or other related products when the consumer consumes the product. That is, consumer behavior can be relevant to the entire supply chain, although it is not a component of the supply chain. However, in the invention described in Patent Document 1, consumer behavior is not taken into consideration.

Therefore, it is an object of the present invention to enable a person to comprehensively analyze the supply chain without understanding the characteristics of business, taking into consideration the consumption behavior of consumers.

The feature extraction device of the present invention combines two of the data related to purchase and sale in a company, the data linkage platform that stores the purchase information and consumption information in the consumer, and the data stored in the data linkage platform. It is equipped with an analysis engine that constructs and analyzes a matrix structure.

Further, the feature extraction method of the present invention includes a process of accumulating data related to purchasing, sales and consumption behavior in a company, and purchasing and consumption behavior in an individual in a data linkage platform, and data accumulated in the data linkage platform. A step of constructing a matrix structure in which two of the two are combined and a step of analyzing using the matrix structure are executed.
Other means will be described in the form for carrying out the invention.

According to the present invention, it is possible to perform a comprehensive supply chain analysis without omission, taking into consideration the consumer behavior of consumers without understanding the characteristics of business.

It is a figure which shows the configuration example of the feature extraction system in this embodiment. It is a figure which shows the structural example of the feature extraction apparatus. It is a figure which shows the hardware configuration of the feature extraction apparatus. It is a flowchart which shows the estimation process of the purchase amount of each consumer. It is a flowchart which shows the estimation process of the consumption amount of each consumer. It is a figure (the 1) which shows an example of the processing of the feature extraction apparatus. It is a figure (the 2) which shows an example of the processing of a feature extraction apparatus. It is a figure which shows the operation which creates the heat map image by combining two variables. It is a figure which shows the IoT data which concerns on consumption behavior. It is a figure which shows an example of the constructed neural network. It is a figure which shows an example of the operation which searches the constructed neural network and discovers a feature quantity. It is a figure which shows an example of the operation of back-converting a feature quantity into an original variable. It is a figure which shows an example of the operation which estimates the behavior of a consumer by searching the constructed neural network. It is a figure which shows the outline of the supply chain.

Hereinafter, with respect to the embodiment of the present invention, each figure will be used as an example of the case where business improvement measures are recommended in the process from the production of the product to the distribution and the process of reaching and consuming the product through the store. This will be described in detail.

FIG. 1 is a diagram showing a configuration example of the feature extraction system S in the present embodiment.
The feature extraction system S is configured by connecting a feature extraction device 1, an electronic commerce server 97, a consumer server 96, and a living data server 50 installed in a data center or the like via a network N.

The electronic commerce server 97 is a server on which an electronic commerce site for retailers to sell to consumers is installed. The electronic commerce server 97 stores purchase history information of each consumer.

The consumer server 96 collects consumption activity data and living data by a plurality of consumers from each living data server 50. The living data server 50 is installed in a consumer's home or the like, detects the movement of each consumer, generates living data and consumption activity data, and uploads them to the consumer server 96.

A refrigerator 51, a microwave oven 52, a camera 54, an AI speaker 55, an activity meter 56, a smartphone 57, a wearable sensor 58, and a smart watch 59 are connected to the living data server 50 by wire or wirelessly. The home electric appliance connected to the living data server 50 is not limited to the one shown in FIG.

The living data server 50 collects and accumulates IoT data from these home appliances, processes and transforms these data to generate living data and consumption activity data, and transfers them to a higher-level consumer server 96. As a result, the life data server 50 can capture the movement of the consumer and its persona.

A sensor 511 such as an imaging device is installed in the refrigerator 51, and can detect the inventory status of foodstuffs. Not limited to this, for example, if a wireless tag is attached to each product and the refrigerator 51 has a mechanism to detect the wireless tag, which store the product was purchased from, when it was put in the refrigerator, etc. You can see if it was cooked with the ingredients.
A sensor 521 such as an image pickup device is installed in the microwave oven 52, and can detect the type and amount of cooking ingredients. The IoT data related to the consumption activity detected by the refrigerator 51 and the microwave oven 52 is transmitted to the living data server 50 and collected / accumulated. Here, the consumption activity includes, for example, "cooking according to a predetermined recipe", "consumption of food and drink", and the like.

The camera 54 and the AI speaker 55 are installed in, for example, a living room, and can detect daily activities such as cooking and eating by consumers. If the camera 54 is installed in the kitchen, the consumer can grasp the contents of the dish including the amount.

The activity meter 56, the smartphone 57, the wearable sensor 58, and the smart watch 59 can detect IoT data related to the daily activities of each consumer. The consumer IoT data detected by these devices is sent to the living data server 50. The living data server 50 processes and transforms the IoT data to generate living data related to the living behavior of the consumer. Here, the consumer's daily behavior includes the timing and time of daily events such as going out, going to bed, eating, taking medicine, and contacting others, and the physical and mental status of the consumer associated with those behaviors. ..

FIG. 2 is a diagram showing a configuration example of the feature extraction device 1 according to the present embodiment.
The feature extraction device 1 of the present embodiment includes a data linkage platform 2, an analysis engine device 3, a data storage unit 4, a policy devising unit 111, and an external system interface unit 19, and a computer that implements these in a series. Is.

The data linkage platform 2 stores information starting from data on materials purchased by a product manufacturer, and extending to a production process, product wholesale, distribution, store / EC site, and consumption. Since the data linkage platform 2 stores data related to purchasing, sales and consumption behavior in connected companies, and purchasing and consumption behavior in individual consumers on the same platform, it is possible to handle such data in a unified manner. It is supposed to be. The data linkage platform 2 further accumulates data related to the lives of individual consumers.

The data linkage platform 2 includes raw material maker data 21, product maker data 22, trading company / wholesale data 23, store data 241 and EC (Electronic Commerce) site data 242, consumer data 25, consumer behavior data 27, living data 29, and ISM. (In-store Merchandising) Stores data 246. The data linkage platform 2 further stores sales promotion history data 263, SNS (Social Networking Service) data 262, external data 261 and customer management data 243, IoT data 244, sensor data 245, and the like.

The raw material maker data 21 is information on the production amount and timing of the raw material of the product.
The product maker data 22 is information on the production amount of each product and information on the shipping time of these products, and is based on the purchased raw materials. These raw material manufacturer data 21 and product manufacturer data 22 are, for example, acquired from a manufacturing execution system (MES) in each manufacturer.

The trading company / wholesale data 23 is information on the distribution amount and timing of the produced products (commodities).
The store data 241 is information on the number of purchases and the timing at the actual store.
The EC site data 242 is information regarding the action history on the EC site.
The consumer data 25 is information held by the retail business in order to grasp the behavior of the consumer, for example, the number of page views in the access history to the EC site, the online migration history, and the like. Further, the customer management data 243 is product purchasing information and marketing information on the downstream side one below or below the company connected by the supply chain acquired from CRM (Customer Relationship Management). Customer management data 243 is consumer information for the retail industry.

The consumption behavior data 27 is data indicating the consumption behavior of the product by each consumer, such as "cooking gratin" and "consumption of milk". The consumption behavior data 27 determines what kind of consumption activity the life data server 50 corresponds to, for example, based on the cooking motion image of the kitchen taken by the refrigerator 51, the microwave oven 52, and the camera 54 shown in FIG.

The living data 29 is data related to the life of each consumer, such as "having a supper" and "having a breakfast". The living data 29 determines what kind of living behavior the living data server 50 corresponds to, for example, based on the IoT data collected by the activity meter 56, the smartphone 57, the wearable sensor 58, and the smart watch 59 shown in FIG. judge.

The ISM data 246 is display / arrangement information in the store, and refers to, for example, the number of shelves, lighting, the presence / absence of POP (Point of Purchase advertising), the distance from the entrance, and the like. The sales promotion history data 263 is historical information related to sales promotion (campaign) to the downstream side of the supply chain by raw material manufacturers, product manufacturers, trading companies, wholesalers, and retailers. The SNS data 262 is written information and aggregated information transmitted by the SNS. The external data 261 includes weather information, position information by GPS (Global Positioning System) or mesh, existence information of competing stores, day of the week information, event information, and the like.

IoT data 244 is information on consumer movement lines and contact behavior with product shelves using human flow sensors, etc., store and outside temperature information using temperature sensors, and stores and outside using humidity sensors. Humidity information, other information detected by sensors such as pollen amount, wind speed, rain amount, etc., storage and removal of the refrigerator to grasp the time when the consumer purchased and consumed the product, and cooking motion image of the kitchen. The sensor data 245 is behavior data of staff and consumers in the store using a wearable sensor, and includes steps, behavior estimation, activity, and the like. The IoT data 244 is data of a high-performance refrigerator 51, a microwave oven 52, and a washing machine. The microwave oven 52 can be used as it is because it is possible to know what was made by selecting the cooking menu. The IoT data 244 may be data on the usage status of the detergent related to the operation of the washing machine.

The data for each business and each data collection unit is hereinafter referred to as a "group", and the data items that are composed thereof are hereinafter referred to as "variables". For example, when multiple companies such as a milling maker and a dairy maker provide raw materials for confectionery, these multiple companies belong to the same group of raw material makers. Further, in the case of multi-stage wholesale that flows from the manufacturer to wholesale B via wholesale A and is connected to the retail store, these wholesale A and wholesale B have the same role and belong to the same group of wholesale. The processing of the analysis engine device 3 of the present embodiment is performed in group units. In addition, a persona is set for each group, which is a unit.

The analysis engine device 3 includes a data preprocessing unit 31, a network structure construction unit 32, a feature amount estimation unit 33, a product recommendation planning unit 34, a data interpretation unit 35, and a consumption behavior estimation unit 36.
The data preprocessing unit 31 preprocesses various data included in the data linkage platform 2 to generate a feature amount.

The network structure construction unit 32 constructs a network structure that associates the persona, each feature amount generated by the data preprocessing unit 31, and a key performance indicator by deep learning technology. A persona is a typical user image of user information (attributes) and services / products. Key performance indicators are either sales, gross profit, net profit, number of shipments, sales volume, number of purchases, purchase frequency, joint sales, or consumption.

The feature amount estimation unit 33 estimates the feature amount that contributes to the improvement of the key performance evaluation index based on the network structure constructed by the network structure construction unit 32. The feature amount estimation unit 33 estimates the causal relationship from the input layer to the output layer via a human-understandable intermediate layer.
The product recommendation planning unit 34 recommends a product that contributes to the improvement of the key performance evaluation index based on the feature amount estimated by the feature amount estimation unit 33.

The data interpretation unit 35 reversely transforms the feature amount estimated by the feature amount estimation unit 33 into a form that is easy for humans to interpret. The retail marketer, the product maker marketer, and the product developer capture the changes in the factors for each persona from the various data inversely converted by the data interpretation unit 35. Furthermore, each person in charge or developer can mean and understand the change in the tendency of the middle class due to the persona.

The consumption behavior estimation unit 36 estimates the consumer's living behavior and consumption behavior that affect the improvement of the important performance evaluation index based on the network structure constructed by the network structure construction unit 32. The consumption behavior estimation unit 36 estimates a causal relationship from the output layer to the input layer via a human-understandable intermediate layer.

The data storage unit 4 stores the preprocessed images 72a, 72b, ... Which are the feature quantities generated by the data preprocessing unit 31, the persona data 41 including each attribute of the persona, the important performance evaluation index data 42, and the like. ..

The measure devising unit 111 calculates the optimum combination of two variables from the feature amount (normalized) estimated by the feature amount estimation unit 33 to further improve the key performance indicator, and measures this. Invent as.

The external system interface unit 19 is an interface for linking with the external system. The external system interface unit 19 is an interface for distributing and coordinating the sales promotion measures planned by the analysis engine device 3 to the external system.

The feature extraction device 1 of the present embodiment utilizes the data linkage platform 2 to catch and correct the tendency change of each model even between models having different properties. The analysis engine device 3 of the feature extraction device 1 decomposes each of the complicated elements and features into the factors that are the components, and estimates the relationship between the persona and the key performance indicator. As a result, the measure devising unit 111 can devise a measure for improving the important performance evaluation index. Further, the marketer of the retail industry, the marketer of the product maker, and the product developer can extract the characteristics of the consumer by the feature extraction device 1 and use it as a reference for studying product recommendation measures and product development. Here, the consumer is a general consumer, and means a user, a member, etc. from the retailer's point of view.

For example, a retail marketer collects not only data of his / her own business but also data related to manufacturing and distribution by utilizing the data linkage platform 2. The marketer can further discover new features across industries and consider sales promotion measures from a new perspective using the analysis engine device 3. With new feature quantities across industries, it is possible to grasp consumer preferences, simultaneous consumption, and product consumption timing that could not be discovered in the past, so the person in charge can take efficient sales promotion measures (for example, products). (Recommendation, distribution of coupons, appropriate distribution measures of products), and thus the important performance evaluation index (KPI) can be improved.

The person in charge of the manufacturer grasps the persona and the consumption pattern of the consumer by using the feature extraction device 1. The person in charge of the manufacturer makes use of the consumer's persona and consumption pattern in product development, and plans to collaborate and acquire with the manufacturer that manufactures the product that is consumed at the same time and the product that is consumed by many consumers. Furthermore, the person in charge of the manufacturer can improve the key performance indicator (KPI) by catching the fashion at an early stage and knowing the consumer's preference.

FIG. 3 is a diagram showing a hardware configuration of the feature extraction device 1.
The feature extraction device 1 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, an auxiliary storage device 14, a communication device 15, a display device 16, an input device 17, and a media reading device. 18 is provided.

The CPU 11 is a unit that executes various arithmetic processes. The CPU 11 embodies the analysis engine device 3 by, for example, reading the analysis engine program 141 stored in the auxiliary storage device 14 into the RAM 13 and executing the program.
The analysis engine program 141 can be stored in a computer on a communication network in a downloadable state, stored in a portable storage medium such as a BD (Blu-ray Disc), and distributed. Further, the analysis engine program 141 can be installed in the feature extraction device 1 via the communication device 15 or the media reading device 18.

The ROM 12 is a storage device that stores a startup program and the like of the feature extraction device 1. The RAM 13 is a storage device that stores a program executed by the CPU 11 and data used for executing the program.

The auxiliary storage device 14 is, for example, a storage device such as an SSD (Solid State Drive) using an HDD (Hard Disk Drive), a flash ROM, or the like, and stores the analysis engine program 141.
The communication device 15 is a device that communicates with an external computer or device.
The display device 16 is, for example, a device such as an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence).
The input device 17 is, for example, a device such as a keyboard or a similar device, a mouse, a touch panel, or a microphone.
The media reading device 18 is a device that reads and writes information to and from a portable recording medium such as a BD.

When the feature extraction device 1 functions as the analysis engine device 3, the data storage area of the data preprocessing unit 31 is realized by the RAM 13 or the auxiliary storage device 14. Further, a part or all of the data storage area of the data preprocessing unit 31 may be realized, for example, in storage on a communication network connected via the communication device 15.
Each functional unit (data storage unit 4, analysis engine device 3, external system interface unit 19) is realized by, for example, at least one of a CPU 11, a RAM 13, and an auxiliary storage device 14. That is, the processing of each functional unit can be realized, for example, by the CPU 11 loading the analysis engine program 141 stored in the auxiliary storage device 14 into the RAM 13 and executing the processing. The analysis engine device 3 may be realized by coordinating a plurality of computers.

FIG. 4 is a flowchart showing an estimation process of the purchase amount of each consumer. This estimation process is executed by the data preprocessing unit 31 of the feature extraction device 1.
First, the data preprocessing unit 31 acquires the purchase history of each consumer from the electronic commerce server 97 (S40). Then, the data pre-processing unit 31 totals the purchase amount of each product by each consumer (S41), and ends the process of FIG.

FIG. 5 is a flowchart showing an estimation process of the consumption amount of each consumer. This estimation process is executed by the life data server 50 shown in FIG. 1, for example, at a predetermined cycle.
First, the CPU of the life data server 50 receives IoT data sensed by the consumer's consumption behavior from each device used by the consumer (S50). Then, the CPU of the life data server 50 analyzes the received IoT data, estimates and stores the consumption amount of each product by the consumer (S51). This clarifies the relationship between each consumer's consumption and each consumption behavior.

Next, the CPU of the life data server 50 receives IoT data that senses the life behavior of the consumer from each device that monitors the consumer (S52). Then, the CPU of the life data server 50 analyzes the received IoT data, estimates and stores the life data by the consumer (S53). When the process of step S53 is completed, the CPU of the living data server 50 ends the process of FIG.

6A and 6B are flowcharts showing an example of processing of the analysis engine device 3.
The processes shown in FIGS. 6A and 6B may be executed according to, for example, a timing instructed by a retail marketer, for example, a periodic or a predetermined schedule, and the product maker's marketer / product developer develops the product. It may be executed at the timing when a certain product is consumed by the consumer.

First, the data preprocessing unit 31 of the analysis engine device 3 acquires the data required for analysis from the data linkage platform 2 (S10). For example, the data preprocessing unit 31 acquires planning and actual information such as production, shipping, and inventory from the manufacturing execution system of each manufacturer, and uses them as raw material manufacturer data 21 and product manufacturer data 22, respectively. Next, the data preprocessing unit 31 acquires information such as inventory and distribution plans in factories and distribution warehouses from supply chain management in distribution and uses them as trading company / wholesale data 23.

Further, the data preprocessing unit 31 acquires product purchase information, marketing information, etc. from CRM (Customer Relationship Management) as customer management, and uses it as customer management data 243. The data preprocessing unit 31 acquires the number of page views, the net migration history, and the like from the access history to the EC site, and uses it as the consumer data 25.

Next, the data preprocessing unit 31 determines a keyword to be an aggregation key from each operation on the data linkage platform 2, and aggregates the data stored in each operation for each primary key (S11). In this case, the aggregation unit may be, for example, a unit such as one week or one month, an arbitrary cycle, or an arbitrary schedule unit.

The data preprocessing unit 31 repeats the loop from steps S12 to S16 by the combination of variables.
First, the data preprocessing unit 31 combines arbitrary two variables from, for example, the partition table 241d generated for each industry (S13).
Next, the data preprocessing unit 31 creates a heat map image that visualizes a variable matrix in which two variables of each industry are combined (S14).

FIG. 7 is a diagram showing an operation of creating a heat map image by combining two of all variables related to the amount ordered from the manufacturer.
As shown in FIG. 7, here, the heat map is created by combining the variable # 1 (70a) and the variable # 4 (70d). The variable # 1 (70a) and the variable # 4 (70d) are a combination of two of the quantitative data and the qualitative data held in the data linkage platform 2. The two combined variables are normalized from 0.0 to 1.0 and heat-mapped to create preprocessed images 71b, 71c, ... The preprocessed images 71b, 71c, ... Have a matrix structure in which two data are combined for each unit.
Hereinafter, when the preprocessed images 71b, 71c, ... Are not distinguished, they are simply described as the preprocessed image 71. Here, if there are N variables related to the quantity ordered from the manufacturer, _n C ₂ types of preprocessed images 71 are created.

As described above, since the feature extraction device 1 creates the heat map image in which the combination of the two variables is normalized, the existing neural network library for determining the similarity of the images can be diverted. Furthermore, the feature extraction device 1 can easily identify an image that ignites (activates) as the key performance indicator (KPI) improves by learning the heat map image with such a neural network.

Here, in the image that ignites (activates) with the improvement of the important performance evaluation index, the difference between the maximum value (peak) and the minimum value (bottom) of the heat map image increases with the improvement of the important performance evaluation index. At this time, there is a significant relationship between any of the two variables related to the heat map image and the key performance indicator.

Returning to FIG. 6A again, the processing after step S15 will be described. The data preprocessing unit 31 is a preprocessed image 72 that combines two of all variables related to the quantity ordered from the manufacturer and a preprocessed image 72 that combines two of all the variables related to the quantity ordered from the wholesaler. Create a visualized image by combining 74, a preprocessed image 76 that combines two of all variables related to retail purchases, and two of the preprocessed images of all variables related to consumption. , Saved in the data storage unit 4 (S15). In step S16, if the data preprocessing unit 31 repeats the combination of two of all variables in all industries, the data preprocessing unit 31 ends this loop processing and proceeds to the processing of step S17.

The network structure construction unit 32 has preprocessed the persona 80, the preprocessed images 71a to 71d, the preprocessed images 73a to 73e, the preprocessed images 75a to 75e, the preprocessed images 77a to 77e, and the preprocessed images 77a to 77e. The neural network 8 is constructed by using the images 79a to 79e, the preprocessed images 60a to 60e, the preprocessed images 61a to 61e, and the key performance indicator 89 (S17). At this time, the input data is, for example, the persona 80 which is the attribute information of the consumer, and the output data is the key performance evaluation index 89 such as the sales amount and the number of shipments to the consumer.

FIG. 8 is a diagram showing an example of IoT data 244 for estimating consumption behavior and living behavior.
The IoT data 244 is a data table including a device column 2441, a data column 2442, and a consumption amount column 2443.

In the device column 2441, the name of the home electric appliance from which the data is collected is stored. In the data column 2442, data related to consumption behavior collected from home appliances and living behavior estimated from sensors are stored. The consumption amount column 2443 stores the consumption amount estimated by the living data server 50 of FIG.

The first line of the IoT data 244 is the data collected from the microwave oven 52 of FIG. In the data field 2442, it is stored that the consumer has selected the menu of "for one gratin". In the consumption amount column 2443, it is stored that 400 ml of milk, 50 g of macaroni, 75 g of chicken, 1/3 of onions and 30 g of flour are estimated consumption amounts.

The second line of the IoT data 244 is the data collected from the refrigerator 51 of FIG. The data column 2442 stores that the consumer has executed the "milk pack removal". The consumption column 2443 stores that a 1 liter pack of milk is the estimated consumption.

The third line of the IoT data 244 is the data collected from the camera 54 of FIG. In the data column 2442, it is stored that the consumer's living behavior is "dinner". The life data server 50 of FIG. 1 determines from the image taken by the camera 54 that the consumer is having a supper. Since each of the data in the first to third lines is executed at a close time, they are related.

FIG. 9 is a diagram showing an example of the constructed neural network 8.
The neural network 8 consumes a persona 80, an order amount 81 from a manufacturer, an order amount 82 from a wholesaler, a purchase amount 83 at retail, a consumer purchase amount 84, a consumer consumption amount 85, and consumption. It is a network composed of consumer behavior 86 and consumer life 87. Persona 80 includes elements that indicate consumer attributes. Persona 80 is, for example, gender, age, occupation, income, hobbies, and the like. Each of these elements constitutes a partial network with the preprocessed images 71a to 71d that constitute the order quantity 81 from the manufacturer. In addition, the order amount 81 from the manufacturer, the order amount 82 from the wholesaler, the purchase amount 83 at the retail, the purchase amount 84 from the consumer, the consumption amount 85 from the consumer, and the consumption behavior 86 from the consumer. The consumer life 87 constitutes each layer of the neural network 8.

Each element of Persona 80 is associated with neurons in the input layer of the partial network. Preprocessed images 71a-71d and the like are associated with neurons in the output layer of the partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer by the neurons in the hidden layer, it is possible to obtain a partial network that outputs preprocessed images 71a to 71d and the like by using each element of these personas 80 as an input. ..

Here, each node of the partial network has a weight indicating the influence from each node of the previous layer and a bias which is a constant that does not affect the node of the previous layer. When training this network, the network structure construction unit 32 initializes the weight and bias of each node, calculates the value of each layer from the input data, and sets the calculated value of the output layer and the preprocessed images 71a to 71d. The loss function showing the divergence is calculated, and the partial differential of the weight and bias of each layer with respect to this loss function is calculated. Then, the network structure construction unit 32 corrects the weight and the bias by using the steepest descent method while calculating the partial differential with respect to the total loss from the partial differential with respect to the loss function for each input. This makes it possible to obtain a partial network that outputs the preprocessed images 71a to 71d when each element of the persona 80 is given to the input layer.

Similarly, the preprocessed images 71a to 71d and the like constituting the order quantity 81 to the manufacturer form a partial network with the preprocessed images 73a to 73e and the like constituting the order quantity 82 to the wholesaler. The preprocessed images 71a-71d and the like that make up the order quantity 81 from the manufacturer are linked to the neurons in the input layer of this partial network. Preprocessed images 73a-73e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 73a to 73e by using these preprocessed images 71a to 71d as inputs. Can be done.

The preprocessed images 73a to 73e and the like constituting the wholesale order quantity 82 form a mutual partial network with the preprocessed images 75a to 75e and the like constituting the retail purchase quantity 83. The preprocessed images 73a-73e and the like that make up the wholesale order quantity 82 are tied to the neurons in the input layer of this partial network. Preprocessed images 75a-75e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 75a to 75e by taking these preprocessed images 73a to 73e as inputs. Can be done.

The preprocessed images 75a to 75e and the like constituting the retail purchase amount 83 form a partial network with each other with the key performance indicator 89. Preprocessed images 75a-75e and the like that make up retail purchases 83 are tied to neurons in the input layer of this partial network. Key performance indicators 89 are linked to neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer by the neurons in the hidden layer, it is possible to obtain a partial network that outputs the important performance evaluation index 89 by using these preprocessed images 75a to 75e as inputs. ..

The key performance indicator 89 is not limited to the preprocessed images 75a to 75e constituting the retail purchase amount 83, and may be linked to neurons in any layer.

The preprocessed images 75a-75e and the like that make up the retail purchase amount 83 further form a mutually partial network with the preprocessed images 77a-77e that make up the consumer purchase amount 84. Preprocessed images 75a-75e and the like that make up retail purchases 83 are tied to neurons in the input layer of this partial network. Preprocessed images 77a-77e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 77a to 77e by taking these preprocessed images 75a to 75e as inputs. Can be done.

The preprocessed images 77a to 77e constituting the consumer purchase amount 84 further form a mutual partial network with the preprocessed images 79a to 79e constituting the consumer consumption amount 85. Preprocessed images 77a-77e and the like that make up the consumer purchase amount 84 are tied to neurons in the input layer of this partial network. Preprocessed images 79a-79e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 79a to 79e by taking these preprocessed images 77a to 77e as inputs. Can be done.

The preprocessed images 79a to 79e constituting the consumer consumption 85 are further mutual with the preprocessed images 60a to 60e constituting the consumer consumption behavior 86 through the consumption amount related to each product. Configure a partial network. Preprocessed images 79a-79e and the like that make up the consumer's consumption 85 are tied to neurons in the input layer of this partial network. Preprocessed images 60a-60e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 60a to 60e by taking these preprocessed images 79a to 79e as inputs. Can be done.

The preprocessed images 60a to 60e constituting the consumer behavior 86 further form a mutual partial network with the preprocessed images 61a to 61e constituting the consumer consumption behavior 86. Preprocessed images 60a-60e and the like that make up the consumer's consumption behavior 86 are linked to neurons in the input layer of this partial network. Preprocessed images 61a-61e and the like are associated with neurons in the output layer of this partial network. There is a hidden layer between the input layer and the output layer, and the result of calculation with the value of the previous layer is stored. Then, by learning the relationship between the input layer and the output layer, the neurons in the hidden layer obtain a partial network that outputs the preprocessed images 61a to 61e by using these preprocessed images 60a to 60e as inputs. Can be done.

The neural network 8 is a feedforward neural network in which all the nodes belong to one of the layers and all the nodes are connected only to the nodes of the adjacent layers. That is, the neural network 8 does not have a connection between layers or a connection across layers. As a result, the feature amount can be obtained by the back propagation method.

Returning to FIG. 6B, the description will be continued. In step S20, the feature amount estimation unit 33 fires (activates) the structure of the neural network 8 when the key performance indicator (KPI) is improved by using a technique such as a back propagation method. And its network relationship are discovered, and these features are stored in the data storage unit 4. That is, the feature amount estimation unit 33 estimates the feature amount that contributes to the key performance indicator from the neural network 8.

FIG. 10 is a diagram showing an example of a method of searching the constructed neural network 8 to find a feature amount.
Here, the feature amount estimation unit 33 estimates that the feature amount that ignites when the key performance indicator (KPI) is improved is the preprocessed image 75a. It is presumed that the neuron in the previous stage that fires the determined preprocessed image 75a neuron is the preprocessed image 73c. Here, as the difference between the maximum value (peak) and the minimum value (bottom) of the preprocessed image 75a increases, the difference between the maximum value (peak) and the minimum value (bottom) of the preprocessed image 73a increases. At this time, there is a significant relationship between any of the two variables related to the preprocessed image 75a and the preprocessed image 73a. Similarly, it is presumed that the neuron in the previous stage that fires the neuron of the preprocessed image 73c is the preprocessed image 71b.

Here, the backpropagation method will be described. When estimating the feature amount for improving the important performance evaluation index in the neural network 8, the feature amount estimation unit 33 gives the neural network 8 a sample for learning. Next, the feature amount estimation unit 33 obtains an important performance evaluation index output by the neural network 8 and obtains an error in the output layer of the neural network 8. The feature amount estimation unit 33 uses the error to calculate the error for each neuron. That is, the feature amount estimation unit 33 calculates the expected output value and the magnification of each neuron, and calculates the difference between the requested output and the actual output to obtain a local error.

The feature amount estimation unit 33 adjusts the weight of each neuron so that this local error becomes small. Then, it is determined that the neurons in the previous stage connected with a larger weight are responsible for the local error. In other words, the neurons in the previous stage, which are determined to be responsible for local errors, are features that contribute to changes in key performance indicators. The feature amount estimation unit 33 performs the same processing on the neuron group in the previous stage of the neuron in the previous stage determined as such.

There are actually a plurality of hidden layers between the key performance indicators (KPI) and the preprocessed images 76a to 76e, but by estimating the features of these hidden layers in order, the front It is possible to calculate which of the processed images 76a to 76e is a feature amount that contributes to a key performance indicator (KPI).

Returning to FIG. 6B, the description will be continued. In step S21, the product recommendation planning unit 34 creates a product recommendation (recommendation) based on the feature amount estimated by the feature amount estimation unit 33, and notifies the outside via the external system interface unit 19 or the like. That is, the product recommendation planning unit 34 recommends a product related to the feature amount.
In step S22, the data interpretation unit 35 inversely transforms the normalized variable combination into the original data in order to interpret the extracted features. This makes it easier for people to understand when planning measures, and it is possible to formulate suitable target numerical values and attributes.

FIG. 11 is a diagram showing an example of an operation of inversely converting a feature quantity into an original variable.
The preprocessed image 71b shown in FIG. 11 is a heat map created by combining the variable # 1 (70a) and the variable # 4 (70d). Since the two combined variables are normalized from 0.0 to 1.0, they are inversely transformed to obtain the original variables # 1 (70a) and variables # 4 (70d).

The explanation will be continued by returning to FIG. 6B. In step S23, the consumption behavior estimation unit 36 estimates the feature amount affecting the key performance indicator 89 from the neural network 8. Here, the consumption behavior estimation unit 36 obtains a neuron as a feature quantity that affects the key performance indicator 89 in the retail purchase amount 83, and by tracing this neuron in the opposite direction, the influence on the key performance indicator 89 is exerted. A neuron with a high feature quantity is estimated (S24). The consumption behavior estimation unit 36 further reversely converts the feature amount into the consumer's consumption behavior and living behavior (S25), and recommends the inversely converted consumer's consumption behavior and living behavior (S25). This makes it easier for people to understand when planning measures.

FIG. 12 is a diagram showing an example of a method of searching the constructed neural network 8 to discover consumption behavior and living behavior.
Here, the feature amount estimation unit 33 estimates that the feature amount that ignites when the key performance indicator (KPI) is improved is the preprocessed image 75a. It is presumed that the neuron in the subsequent stage that fires the neuron of the determined preprocessed image 75a is the preprocessed image 77d. It is presumed that the neuron in the subsequent stage that fires the neuron of the preprocessed image 77d is the preprocessed image 79d. It is estimated that the neuron in the subsequent stage that fires the neuron of the preprocessed image 79d is the preprocessed image 60b. It is presumed that the neuron in the subsequent stage that fires the neuron of the preprocessed image 60b is the preprocessed image 61b.

The explanation will be continued by returning to FIG. 6B. In step S26, the measure devising unit 111 devises a measure for improving the key performance indicator (KPI) by associating the interpreted feature amount, consumption activity and / and living behavior with the data of the data linkage platform 2.

For example, the feature amount estimation unit 33 divides the region of the preprocessed image into neurons, and adjusts the weights for each neuron so that the local error of the key performance indicator (KPI) becomes small. Then, it is determined that the region of the preprocessed image in the previous stage connected with a larger weight is responsible for the local error. That is, the area of the preprocessed image determined to be responsible for the local error is a feature quantity that contributes to the change of the key performance indicator. The measure devising unit 111 devises the change in the area of the preprocessed image estimated in this way as a measure.

In step S27, the people involved in the supply chain feed back the measures devised by the measure devising unit 111 to the business. After that, the feature extraction device 1 returns to the process of step S10 of FIG. 6A. In this way, people involved in the supply chain can efficiently operate the PDCA (Plan, Do, Check, Action) cycle of business by using the feature extraction device 1.

<< Effect of this embodiment >>
The feature extraction device 1 of the present embodiment is a feature quantity for changing a key performance indicator (KPI) in consideration of all processes from product production to distribution and reaching consumers through stores. To discover. The feature extraction device 1 further discovers consumer behavior and lifestyle behavior that affect key performance indicators (KPI). This will allow people involved in the supply chain to formulate measures to improve key performance indicators (KPIs).

In addition, it will be possible to perform mechanical analysis without understanding the business characteristics of the industry, and it will be possible to build business processes without the need for specialized knowledge. In addition, it is possible to clarify the features that could not be discovered by humans, the consumption behavior and living behavior of consumers, and to grasp the relationship between hidden pre- and post-processes, which makes it possible to realize effective measures. And.

(Modification example)
The present invention is not limited to the above-described embodiment, and includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to the one including all the described configurations. It is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add / delete / replace a part of the configuration of each embodiment with another configuration.

Each of the above configurations, functions, processing units, processing means, and the like may be partially or wholly realized by hardware such as an integrated circuit. Each of the above configurations, functions, and the like may be realized by software by the processor interpreting and executing a program that realizes each function. Information such as programs, tables, and files that realize each function can be placed in a recording device such as a memory, hard disk, SSD (Solid State Drive), or a recording medium such as a flash memory card or DVD (Digital Versatile Disk). can.
In each embodiment, the control lines and information lines indicate what is considered necessary for explanation, and do not necessarily indicate all the control lines and information lines in the product. In practice, it can be considered that almost all configurations are interconnected.

Examples of modifications of the present invention include the following (a) to (f).
(A) The feature extraction device of the present invention may be configured by a server on the cloud (data center), and may be further configured by an on-premises (owned by the company) server, and is not limited.
(B) Each neuron constituting the neural network is not limited to the heat map image composed of two variables, and may be each variable itself.
(C) The present invention is not limited to the supply chain between a plurality of connected companies. You may analyze purchasing and sales data in a single company and devise measures to improve key performance indicators (KPIs). In this case, the persona is not the end consumer, but the consumer of the company.
(D) The data stored in the data linkage platform 2 is not limited to that shown in FIG. 1, and other data may be stored.
(E) The method for obtaining the feature amount is not limited to the back-propagation method, and any method may be used.
(F) The matrix structure is not limited to a two-dimensional image, and may be a multidimensional object having three or more dimensions. For example, by setting the matrix structure to a three-dimensional object, an existing neural network library that handles the three-dimensional object can be diverted.

S Feature extraction system 1 Feature extraction device 11 CPU
12 ROM
13 RAM
14 Auxiliary storage device 141 Analysis engine program 15 Communication device 16 Display device 17 Input device 18 Media reader 111 Measure devising department 19 External system interface unit 2 Data linkage platform 21 Raw material manufacturer data 22 Product manufacturer data 23 Trading company / wholesale data 241 Store data 242 EC site data 243 Customer management data 244 IoT data 245 Sensor data 246 ISM data 25 Consumer data 261 External data 262 SNS data 263 Sales promotion history data 3 Analysis engine device (analysis engine)
31 Data preprocessing section 32 Network structure construction section 33 Feature estimation section 34 Product recommendation planning section 35 Data interpretation section 36 Consumption behavior estimation section (estimation section)
4 Data storage unit 41 Persona data 42 Important performance evaluation index data 50 Life data server 51 Refrigerator 511 Sensor 52 Microscope 521 Sensor 54 Camera 55 AI speaker 56 Activity meter 57 Smartphone 58 Wearable sensor 59 Smart watch 60, 60a to 60e Preprocessing Finished image (matrix structure)
61, 61a to 61e Preprocessed image (matrix structure)
70a variable # 1
70d variable # 4
71, 71a to 71d Preprocessed image (matrix structure)
73, 73a to 73e Preprocessed image (matrix structure)
75,75a-75e Preprocessed image (matrix structure)
77, 77a-77e Preprocessed image (matrix structure)
79, 79a-79e Preprocessed image (matrix structure)
8 Neural Network 80 Persona 81 Orders from Manufacturers 82 Orders to Wholesale 83 Purchases at Retail 84 Consumer Purchases 85 Consumer Consumption 86 Consumer Consumption Behavior 87 Consumer Life 89 Important Performance Evaluation Indicator 90 Raw material maker / producer 91 Product maker 92 Trading company / wholesale 93 Retail 94 Consumer 96 Consumer server 97 Electronic commerce server

Claims (9)

A data linkage platform that accumulates data related to purchasing, sales and consumption behavior in companies, and purchasing and consumption behavior in individuals,
An analysis engine that constructs and analyzes a matrix structure that combines two of the data accumulated in the data linkage platform, and
A feature extraction device characterized by being equipped with. The feature extraction device according to claim 1, wherein data related to the life of an individual is further accumulated in the data linkage platform. The analysis engine constructs the matrix structure in which two data are combined for each unit, and analyzes the relationship between the data and the key performance indicator for each unit.
The feature extraction device according to claim 1 or 2. The key performance indicator is any one of sales, gross profit, net profit, number of shipments, sales volume, number of purchases, frequency of purchase, and consumption behavior in a company, and number of purchases, frequency of purchase, and consumption behavior in an individual. Is,
The feature extraction device according to claim 3. The data linkage platform accumulates data related to purchasing, sales and consumption behaviors between a plurality of connected companies, and individual purchasing and consumption behaviors.
The feature extraction device according to claim 1. The analysis engine
A network structure construction unit that builds a neural network that associates the matrix structure that combines two data, the key performance indicator, and the persona that is the attribute of the consumer.
A feature amount estimation unit that estimates a feature amount that contributes to the key performance indicator from the neural network,
The feature extraction device according to claim 3, further comprising. The analysis engine
A network structure construction unit that builds a neural network that associates the matrix structure, the key performance indicators, and the consumer behavior of consumers by combining the two data.
A consumption behavior estimation unit that estimates consumption behavior that contributes to the key performance indicator from the neural network,
The feature extraction device according to claim 3, further comprising. The network structure construction unit associates the consumer's life data with the neural network,
A life estimation unit that estimates life data that contributes to the key performance indicators from the neural network,
The feature extraction device according to claim 7, wherein the feature extraction device is provided. The process of accumulating data related to purchasing, sales and consumption behavior in companies, and purchasing and consumption behavior in individuals on the data linkage platform,
The process of constructing a matrix structure that combines two of the data accumulated in the data linkage platform, and
The process of analysis using the matrix structure and
A feature extraction method characterized by executing.

Images