JP2022076421A - Demand prediction server and method for predicting demand - Google Patents

Abstract

To provide a mechanism of increasing the accuracy of predicting a sale by conducting a plurality of pre-processing while switching the processing before calculating a sale prediction.SOLUTION: A demand prediction server includes: a sale data acquiring unit for acquiring sale data of an item; cleansing processing unit for calculating post-cleansing data by executing a combination of cleansing processing to correct data for predicting a sale, on the acquired sale data; a model learning processing unit for generating sale prediction models, by a mechanical learning, with the post-cleansing data as an input and with the sale prediction data as an output, calculating the accuracy of the generated sale prediction models, and selecting a sale prediction model with an accuracy of at least a predetermined value; and a sale prediction processing unit for outputting sale prediction data on the basis of the selected sale prediction model.SELECTED DRAWING: Figure 20

Description

The present invention relates to a demand forecast server and a demand forecast method.

As a background technique in this technical field, there is Japanese Patent Application Laid-Open No. 2016-115157 (Patent Document 1). In this gazette, "In order to prevent an increase in the load when providing highly accurate sales forecasts for each product, the sales forecast program provision method of the embodiment is such that the computer has different sales for each of a plurality of products. The sales forecast program corresponding to each of the plurality of products generated by using the prediction model generation algorithm is displayed in an identifiable state and in a selectable state as to which of the plurality of products is supported. The process is executed. "

Japanese Unexamined Patent Publication No. 2016-115157

However, in Patent Document 1, it is not considered at all to switch and execute a plurality of preprocessing before calculating the sales forecast.

Therefore, the present invention provides a mechanism for improving the accuracy of sales forecast by switching and executing a plurality of preprocesses before calculating the sales forecast.

In order to solve the above problems, for example, the configuration described in the claims is adopted.
The present application includes a plurality of means for solving the above problems. For example, a sales data acquisition unit for acquiring product sales data and a sales forecast for the acquired sales data. A cleansing processing unit that calculates post-cleansing data by executing a combination of multiple cleansing processes that modify the data, and a sales forecast model that inputs the post-cleansing data and outputs sales forecast data are generated by machine learning. , A model learning processing unit that calculates the accuracy of the generated sales forecast model and selects the sales forecast model whose accuracy is equal to or higher than a predetermined value, and outputs sales forecast data based on the selected sales forecast model. It is characterized by having a sales forecast processing unit and a sales forecast processing unit.

According to the present invention, it is possible to provide a mechanism for improving the accuracy of sales forecast by switching and executing a plurality of preprocesses before calculating the sales forecast.
Issues, configurations and effects other than those described above will be clarified by the following description of the embodiments.

It is an example of the block diagram of the whole sales forecast system 1. This is an example of the hardware configuration of the sales forecast server 101. This is an example of the hardware configuration of the user terminal 102. This is an example of the hardware configuration of the administrator terminal 103. This is an example of product management information 500. This is an example of sales information 600. This is an example of sales total information 700. This is an example of the feature amount additional sales information 800. This is an example of accuracy evaluation information 900. It is an example of another accuracy evaluation information 1000. It is an example of the figure explaining the data flow of the sales forecast system 1. This is an example of the stylized processing flow 1200. This is an example of the cleansing process flow 1300. This is an example of the cleansing model generation processing flow 1400. This is an example of the feature amount generation processing flow 1500. This is an example of the sales forecast model learning process flow 1600. This is an example of the sales forecast processing flow 1700. This is an example of the cleansing process correction / change process flow 1800. This is an example of the data display screen 1900 after cleansing. This is an example of the sales forecast display screen 2000. This is an example of the cleansing process determination model generation process flow 2100.

Hereinafter, examples will be described with reference to the drawings.

FIG. 1 is an example of a configuration diagram of the entire sales forecast system 1.
The sales forecasting system 1 includes a plurality of user terminals 102 and a plurality of administrator terminals 103, each of which is connected to the sales forecasting server 101 via a network. The network may be wired or wireless, and each terminal can send and receive information via the network.

Each terminal of the sales forecast system 1 and the sales forecast server 101 may be, for example, a mobile terminal (mobile terminal) such as a smartphone, a tablet, a mobile phone, or a mobile information terminal (PDA), or may be a glasses type, a wristwatch type, or a clothing type. It may be a wearable terminal such as. It may also be a stationary or portable computer, or a server located in the cloud or on a network. Further, the function may be a VR (Virtual Reality) terminal, an AR (Augmented Reality) terminal, or an MR (Mixed Reality) terminal. Alternatively, it may be a combination of these plurality of terminals. For example, a combination of one smartphone and one wearable terminal can logically function as one terminal. Further, it may be an information processing terminal other than these.

Each terminal of the sales forecast system 1 and the sales forecast server 101 have a processor that executes an operating system, an application, a program, etc., a main storage device such as a RAM (Random Access Memory), an IC card, a hard disk drive, and an SSD (SSD). Solid State Drive), auxiliary storage devices such as flash memory, communication control units such as network cards, wireless communication modules, and mobile communication modules, touch panel, keyboard, mouse, voice input, input by motion detection by image capture of camera unit, etc. It is equipped with an input device and an output device such as a monitor and a display. The output device may be a device or a terminal for transmitting information for output to an external monitor, display, printer, device, or the like.

Various programs and applications (modules) are stored in the main memory, and each functional element of the entire system is realized by executing these programs and applications by the processor. In addition, each of these modules may be implemented by hardware by integrating them. Further, each module may be an independent program or application, but may be implemented in the form of a part of a subprogram or a function in one integrated program or application.

In this specification, each module is described as a subject (subject) that performs processing, but in reality, a processor that processes various programs, applications, etc. (module) executes processing.
Various databases (DBs) are stored in the auxiliary storage device. A "database" is a functional element (storage unit) that stores a data set so that it can handle arbitrary data operations (for example, extraction, addition, deletion, overwriting, etc.) from a processor or an external computer. The method of implementing the database is not limited, and may be, for example, a database management system, spreadsheet software, or a text file such as XML or JSON.

The user terminal 102 is a terminal installed in a company or the like that uses sales forecast information. Using sales forecast information includes not only using sales forecast information by oneself, but also providing the information to other companies.
The administrator terminal 103 is a terminal used by an administrator or the like of the sales forecast system 1.
The sales forecast server 101 receives input of various information such as product information and sales information necessary for performing sales forecast from the user terminal 102, and stores these in the company DB 221.

FIG. 2 is an example of the hardware configuration of the sales forecast server 101.
The sales forecast server 101 is composed of, for example, a server arranged on the cloud.
The main storage device 201 stores programs and applications such as a stylized module 211, a cleansing module 212, a feature quantity generation module 213, a sales forecast model learning module 214, a sales forecast module 215, and a visualization module 216. Each functional element of the sales forecast server 101 is realized by executing the program or application by the processor 203.

The stylized module 211 performs a process of converting information about products and sales in different formats sent from each company and stored in the company DB 221 into a stylized format.
The cleansing module 212 performs cleansing processing on the stylized sales data. The sales data of the company DB221 may have data problems such as input mistakes, design mistakes, peculiar orders, and not in a format suitable for sales forecasts. By modifying it, the data will be changed to a format suitable for machine learning, and the accuracy of sales forecast by machine learning will be improved.

The feature amount generation module 213 performs processing for adding a feature amount to data such as sales that has undergone cleansing processing, and changes the format to a format suitable for inputting a machine learning model. Data such as sales that have undergone cleansing processing may be referred to as sales data after cleansing, and data to which features are further added may be referred to as sales data after cleansing. The feature amount generation module 213 stores the sales data after cleansing in the post-cleansing data DB 223.

The sales forecast model learning module 214 generates one or a plurality of sales forecast models by inputting the cleansed sales data to which the feature amount is added. Further, the sales forecast model learning module 214 evaluates these plurality of sales forecast models and selects the champion model with the highest accuracy. It should be noted that the configuration is not limited to the one with the highest accuracy, and may be configured to select one having an accuracy of a certain threshold value or more.
The sales forecast module 215 executes the sales forecast based on the champion model or the selected sales forecast model, and stores the result in the forecast result DB 224.
The visualization module 216 displays the sales data after cleansing by the cleansing module 212, the sales forecast result by the sales forecast module 215, and the like on the user terminal 102 and the administrator terminal 103.

The auxiliary storage device 202 stores the company DB 221 and the standard data DB 222, the post-cleansing data DB 223, and the prediction result DB 224.
The company DB 221 stores product data and sales data transmitted from a user terminal 102 of a company or the like.
The standard data DB 222 stores sales data and the like after the standardization process.
The post-cleansing data DB 223 stores sales data and the like after the cleansing process.
The forecast result DB 224 stores the information of the sales forecast result.

FIG. 3 is an example of the hardware configuration of the user terminal 102.
The user terminal 102 is composed of, for example, a stationary computer, a notebook computer, a tablet, or the like.
The cleansing result display module 311 and the sales forecast display module 312 are stored in the main storage device 301, and each functional element of the user terminal 102 is realized by executing these programs and applications by the processor 303. ..

The cleansing result display module 311 displays the sales data after the cleansing process and the sales data before the cleansing process on the user terminal 102, the administrator terminal 103, etc. in a form that can be compared, and the cleansing process is performed by the user or the like. Check if is done properly and accept corrections to data and settings.
The sales forecast display module 312 cooperates with the visualization module 216 of the sales forecast server 101 to display the received sales forecast result.
The user terminal data 321 of the auxiliary storage device 302 stores information related to the user, such as sales data of the company to which the user belongs, and the sales forecast result transmitted from the sales forecast server 101.

FIG. 4 is an example of the hardware configuration of the administrator terminal 103.
The administrator terminal 103 is composed of, for example, a stationary computer or a notebook computer.
The management module 411 is stored in the main storage device 401, and each functional element of the administrator terminal 103 is realized by executing these programs and applications by the processor.
The management module 411 manages the entire sales forecast system 1. Further, the management module 411 accesses the sales forecast server 101, and sets and manages the sales forecast server 101.
The administrator terminal data 421 of the auxiliary storage device 402 stores information for managing the sales forecast system 1.

FIG. 5 is an example of the product management information 500.
The product management information 500 is information for managing products, and stores information such as product ID 501, group ID 502, product name 503, JAN code 504, category 505, and list price 506.
The product ID 501 is a unique ID automatically generated for each product.
The group ID 502 is an ID generated for each group composed of a plurality of (corresponding) products having different product identification numbers but substantially the same type of products.
The product name is stored in the product name 503.

The JAN code 504 is one of the product identification information that identifies which product of which business operator. In Japan, products are generally managed by JAN (Japanese Article Number) code, but internationally used EAN (European Article Number) code may be used, and other product identification information. You may.
Product classification information is stored in category 505.
The price of the product is stored in the list price 506.

FIG. 6 is an example of sales information 600.
The sales information 600 is information for managing sales for each product, and stores information such as product ID 601, group ID 602, date 603, store code 604, list price 605, selling price 606, and number of units sold 607.
Product information is stored in the product ID 601 and the group ID, and the value of the product management information 500 is referred to.
The date on which the product was sold is stored in the date 603.
The store code 604 stores information that identifies the store that sold the product.
The list price 605, the selling price 606, and the number of units sold 607 are information about the sold products. The list price 605 may be configured to refer to the value of the list price 506 of the product management information 500.

FIG. 7 is an example of the sales total information 700.
The sales total information 700 is information obtained by totaling the sales information 600 of the product, and stores information such as the product ID 701, the group ID 702, the product name 703, and the sales data 704.
The sales total information 700 stores information on weekly sales in a period from, for example, 4/8 to 6/23.
In this embodiment, sales information such as sales information 600 and sales summary information 700 may be collectively referred to as sales data.

FIG. 8 is an example of the feature amount additional sales information 800.
The feature amount addition sales information 800 is information added by generating a feature amount for improving the accuracy of the machine learning model while making the sales data into a format suitable for inputting a machine learning model.
The feature amount additional sales information 800 stores information such as product ID 801, category 802, date 803, list price 804, selling price 805, number of units sold 806, month 807, day 808, discount rate 809, and the like. In the example of FIG. 8, variables such as product IDs and categories are replaced with the categories of numbers 0, 1, and 2, and the month / day is added from the date, and the discount rate is added as the feature amount from the list price and the selling price.

FIG. 11 is an example of a diagram illustrating a data flow of the sales forecast system 1.
The stylized module 211 generates stylized sales data and the like by stylizing processing of product data and sales data in various formats stored in the company DB 221 and stores them in the stylized data DB 222.
The cleansing module 212 acquires the standardized sales data from the standard data DB 222 and performs the cleansing process, and the visualization module 216 uses the sales data before the cleansing process and the sales data after the cleansing in a comparable form. It is output to the personal terminal 102 and the administrator terminal 103. The visualization module 216 has a Query API (Application Programming Interface) that receives inquiries from the user terminal 102 and the administrator terminal 103, and a Web APP (Web APP) that outputs the answers to the inquiries and the data of the sales forecast server 101 to the user terminal 102. Web Application) and.

The user of the user terminal 102 or the like accepts confirmation of whether or not there is an error in the data correction by the cleansing process, and stores the confirmed sales data after cleansing in the cleansing data DB 223. In addition, the cleansing process settings may also be output, and corrections and confirmations of the settings may be accepted.
The feature quantity generation module 213 performs a process of adding a feature quantity to the cleansing sales data acquired from the cleansing data DB 223, and the sales forecast model learning module 214 generates a sales forecast model for this data. do.
The sales forecast module 215 performs sales forecast processing using a sales forecast model whose accuracy is equal to or higher than a predetermined threshold among the sales forecast models, and stores the result in the forecast result DB 224.
The sales forecast result is displayed on the user terminal 102 and the administrator terminal 103 via the visualization module 216.

FIG. 12 is an example of the stylized processing flow 1200.
The stylized processing flow 1200 is a process in which the stylized module 211 changes product data and sales data in various formats into standard sales data suitable for machine learning.
The stylized module 211 acquires sales data and the like from the company DB 221 (step 1210).
The stylized module 211 changes the acquired sales data and the like into a stylized data format suitable for machine learning (step 1220).
The stylized module 211 stores stylized sales data and the like in the stylized data DB 222.
In this embodiment, the stylized sales data may be referred to as "sales data".

FIG. 13 is an example of the cleansing process flow 1300.
Since there is a possibility that the sales data stored in the company DB 221 is not in a format suitable for input errors, design errors, demand forecasts, etc., the cleansing module 212 cleanses the data in order to improve the accuracy of sales forecasts. conduct.
The cleansing module 212 acquires the stylized sales data from the stylized data DB 222 (step 1310).
The cleansing module 212 analyzes the stylized sales data and selects a cleansing process suitable for the content (step 1320).
The cleansing module 212 executes a selected cleansing process on the stylized sales data (step 1330).

The cleansing module 212 cooperates with the visualization module 216 to display the sales data after cleansing (step 1340).
Specifically, the visualization module 216 acquires the sales data processed by the cleansing module 212, and the visualization module 216 generates and uses screen information in which the sales data before cleansing and the sales data after cleansing are associated with each other. It is transmitted to the person terminal 102 or the like.
The cleansing result display module 311 of the user terminal 102 displays the screen information in which the received sales data and the sales data after cleansing are associated with each other on the output device 305.

The visualization module 216 does not generate screen information in which the sales data before cleansing and the sales data after cleansing are associated with each other, but transmits the corresponding data and the information necessary for display to the user terminal 102, and the user terminal. The cleansing result display module 311 of 102 may be configured to generate a screen in which the received sales data and the sales data after cleansing are associated with each other.
Further, the sales data before cleansing may not be displayed, and only the sales data after cleansing may be displayed.
Further, the cleansing module 212 may be configured to display instead of the visualization module 216.

The cleansing module 212 confirms whether or not there is a correction to the sales data after cleansing, and accepts the correction of the data if there is a correction, via the cleansing result display module 311 of the user terminal 102 (step 1350). The cleansing module 212 determines the modified data after the modified cleansing as the sales data after the cleansing.
The cleansing module 212 stores the sales data after cleansing whose correction is confirmed in the post-cleansing data DB 223 (step 1360).

FIG. 19 is an example of the post-cleansing data display screen 1900.
The cleansing result display module 311 of the user terminal 102 cooperates with the cleansing module 212 and the visualization module 216 of the sales forecast server 101, and displays the sales data before cleansing and the sales data after cleansing on the output device 305 such as a monitor. ..
The cleansing module 212 can detect the part of the data to be corrected by rule base, machine learning, etc., determine the cleansing process to be applied, and execute the correction. Such uncertainties can occur.
・ Is the part that should be corrected really the part that should be corrected? ・ Is the part that should be corrected correctly corrected? Therefore, by giving the user a function to confirm the above uncertainty. , The accuracy of cleansing is improved, and the accuracy of sales forecast is improved.
The sales data after cleansing may be displayed on the user terminal 102 and the sales data after cleansing may be stored directly in the cleansing data DB 223 without being confirmed by the user.

On the post-cleansing data display screen 1900, the price abnormality detection process 1901 and the mass customer order process 1902, which are the performed cleansing processes, are displayed. In the example of the figure, the sales data in which the price abnormality detection process 1901 is applied to the product A1903 is displayed.
In the display part of the sales information after cleansing 1930, the sales data 1931 before the cleansing process and the sales data 1933 after the cleansing process are displayed in a comparable form, and the price of 100 yen on 3/23/2020 has been revised. It has been revised to 1000 yen as the revised part 1934 of the later sales data 1933.

Regarding the value of the sales data 1933 after cleansing, there is a correction confirmation column 1932 for receiving a correction confirmation from the user, and the user accepts the confirmation of the correction value or the correction of the value. For example, when "OK" is entered in the correction confirmation column 1932, it indicates that the corrected value has been confirmed to be correct. Alternatively, if nothing is entered, processing may be performed assuming that the corrected value is correct. When correcting the value, the user enters the correct value in the correction confirmation column 1932.
The sales data 1931 before the cleansing process and the sales data 1933 after the cleansing process are displayed on the cleansing process graph display 1910 in a form in which the respective data can be compared. In the example of the figure, the actual price of 100 yen on March 23, 2020 is an erroneous numerical value, and is corrected to 1000 yen by the correction by the cleansing process.

The modified and confirmed post-cleansing sales data is transmitted to the sales forecast server 101 by the cleansing result display module 311 and stored in the post-cleansing data DB 223 by the cleansing module 212.
In the example of the figure, sales data and post-cleansing data are displayed in a form that can be compared with respect to a plurality of cleansing processes 1901 and 1902.
For each of the displayed multiple cleansing processes, it is possible to select whether or not to actually use it for sales forecast, that is, to confirm whether or not it can be used, and to accept correction of the value if it is not used. The data may be discarded without modifying the value. In addition, the cleansing process settings may also be output, and corrections and confirmations of the settings may be accepted.
The cleansing module 212 is based on the modified post-cleansing data, the confirmed post-cleansing data, or the modification or confirmation of the cleansing process settings, and the rules and settings (that is, the conditions and conditions) for executing the cleansing process. Processing details and settings) can also be updated.

In addition, the user's cleansing result display module 311 can return the following feedback to the sales forecast server 101 through the Web App of the visualization module 216.
-Select whether to execute the displayed cleansing process.
-Check if the part targeted for cleansing by the system really needs to be corrected.
-Corrected-Check if the content you plan to do is correct.
-Change the cleansing process to be used and the settings of each cleansing process.

Here, individual cleansing processes will be described.
The demand forecast system itself is a general-purpose system that does not depend on the industry, but in this embodiment, an example of cleansing processing assuming the retail industry will be described.
Although these processes are performed by the cleansing module 212, the sub-modules of the cleansing module 212 may be configured to perform the respective cleansing processes.

1. 1. Missing value complement The missing value complementing process is a process that complements a part of the sales data that is missing.
When the cleansing module 212 finds a partial loss of data, the loss is complemented by selecting an appropriate one for each column from the following methods.
-Complement with the values of the date and time before and after.
・ Complement with a fixed value. (For example, sales can be complemented with 0 (zero) for missing parts, and the missing item flag can be supplemented as if there is no missing item.)
・ Calculate the average value or median value, and supplement with that value.

2. 2. Price anomaly detection Price anomaly detection is a process of detecting and correcting a part where the sales data may be incorrect as compared with other corresponding sales data.
The cleansing module 212 detects and corrects potential price data errors. For example, there are the following methods.
-Detect and correct areas where the price is higher than the corresponding other data. (For example, it detects that the price is higher than the list price and corrects it so that the upper limit is the list price.)
-Detect and correct areas where the price is lower than the corresponding other data. (For example, it detects those that are 50% or less of the list price, and corrects them so that the lower limit is 50% of the list price.)
-Calculate the average value of prices for a predetermined number of days before and after, detect an abnormality if there is a difference of a predetermined value (for example, 10%) or more from this average value, and correct the abnormal value to this average value, for example. do.

3. 3. Anomaly detection of sales data (especially for large numbers of customers)
The detection of a large number of customers is a process of excluding the sales when it is detected that the sales exceeding a predetermined value are generated within a predetermined period.
In retail stores, in addition to normal sales at stores and EC, there is a possibility of receiving large customer orders from companies, etc. Such large customer orders are excluded and only normal sales are excluded for the following reasons. It is better for machine learning to make predictions with.
・ Mass customer orders often come on a large scale irregularly for products in a specific category, making it extremely difficult to predict.
-Since the scale is large, it is often the case that production starts after receiving an order, so it is not appropriate to include it in the forecast target.

There are the following methods for determining whether the cleansing module 212 is a mass customer order.
-When it is detected that the sales at a certain store exceed the threshold value set for each product, it is determined that there is a large amount of customer orders at that store.
-For each product, calculate the average μ and standard deviation σ of daily sales nationwide, and determine that there was a large number of customer orders on the day of sales of μ + nσ or more.
-For each product, calculate the average μ and standard deviation σ of daily sales for each store, and determine that there was a large number of customer orders on the day of sales of μ + nσ or more.
・ For each product, the average μ and standard deviation σ of the daily ratio of sales at a certain store to the sales nationwide are calculated, and on days with a ratio of μ + nσ or more, a large number of customer orders are sent to that store. Judged as

When it is determined that there is a large number of customer orders, the cleansing module 212 may exclude the data of the location for the large number of customer orders as follows.
-Complement with the sales values on the dates before and after.
-Take the average or median of sales for n days before and after, and supplement with that value.

4. JAN linking JAN linking is a process of associating product identification information before and after switching with respect to a product whose product identification information has been switched.
If the product standard changes slightly, the JAN code may change even if the actual product is the same. Therefore, the cleansing module 212 associates the products by the following method.
-The JAN code is universal product identification information that indicates "which company and which product". Products with switched JAN codes are treated as the same product by associating the JAN codes before and after the switching.
Specifically, JAN linking (correspondence) is performed by a machine learning algorithm or a rule base by inputting waveforms of product name data and sales data.

For example, in the example of the sales total information 700 in FIG. 7, the sales of the “plastic ruler size” with the product ID 701 of 5 decreased from around 4/29 and became zero around 6/3, while being replaced with this. Sales of the "BIG ruler" with product ID 701 of 12 increased from around April 29, and reached about the same level as the "plastic ruler size" around June 3.
By analyzing these sales, the cleansing module 212 determines that the name of "plastic ruler size" has been changed to "BIG ruler", and performs a process of associating (corresponding) the JAN code of these products.

5. Processing of set products There are cases where multiple products are sold as a set. For example, it is assumed that product A and product B are sold as a new life support set. In this case, the actual sales of the product A are the sum of the sales of the product A alone and the sales of the product A included in the set. However, as data, since the sales of the set product and the sales of the product A remain as separate data, there is a possibility that the data does not reflect the demand of each individual product.

Therefore, in order to consider the demand of each individual product, there is the following method as data processing performed by the cleansing module 212 for the set product.
-Add and record the sales of set products to the sales of each individual product included in it.
-After that, add a column indicating how many set products each product is included in.
By this processing by the cleansing module 212, it is possible to partially compensate for the information that the product is sold as a set product, which has been lost by simply adding the sales.

6. Anomaly detection of variables that represent sales status Some companies manage information such as whether each product is currently on sale, discontinued, or discontinued. Such information is useful for learning the demand forecast model, and data after production is discontinued can be noise in the learning of sales forecast, such as lower than usual sales, so it should be removed appropriately. If possible, the accuracy will be improved.
For example, the cleansing module 212 confirms the integrity of data by the following processing.
-If it is determined that the sales of the product has been discontinued, check whether the sales are 0 (zero), and if not, correct the data.
・ It remembers the change of the state of normal sales → production discontinuation → end of sales, analyzes the sales data, confirms whether the sales correspond to this change of state, and if not, data Correct.
・ Detect time-series errors such as end of sales → end of production and correct the data.

Although there are multiple cleansing processes as described above, the visualization module 216 cooperates with the cleansing module 212 to compare the sales data with the multiple post-cleansing data generated by executing the multiple cleansing processes. It can be displayed with and accept the selection of a specific cleansing process to be used for sales forecast from multiple cleansing processes.
By confirming which cleansing process to use from multiple cleansing processes from the user, confirming the availability of post-cleansing data, and accepting corrections to each data, each cleansing process is executed based on these data. The rules and settings to be performed (that is, conditions, processing contents and settings) can be updated, and the accuracy of each cleansing process is improved.

FIG. 14 is an example of a cleansing model generation processing flow 1400.
The cleansing module 212 analyzes the stylized sales data and executes the cleansing process by rule-based or machine learning. The cleansing model generation process flow 1400 is a process flow for generating and learning a model for executing this cleansing process by machine learning.
The cleansing module 212 acquires standardized sales data from the standard data DB 222 (step 1410).

The cleansing module 212 performs cleansing processing on the standardized sales data, accepts corrections and confirmations from the user for the data, and acquires the sales data after cleansing in which the corrections are confirmed (step 1420).
The cleansing module 212 uses the sales data before the cleansing process and the sales data after the cleansing process determined to be usable after the correction is confirmed as the teacher data, the input is the sales data, and the output is the output after the cleansing. A cleansing process determination model to be used as sales data is generated by machine learning (step 1430).
The cleansing module 212 stores and outputs the generated cleansing processing model (step 1440).

The cleansing module 212 calculates and outputs the cleansed sales data estimated from the input stylized sales data by using the stored and output cleansing processing determination model.
For the post-cleansing data, a display is output that allows the user to select whether or not to use the data, but the post-cleansing data that is determined to be usable by the user or the correction of the data is accepted from the user. Later, by using the post-cleansing data determined to be usable as the teacher data for generating the cleansing model, the accuracy of the cleansing process using the model can be improved.

By accepting the feedback that the user confirms and corrects the sales data after cleansing as described above, the following effects are obtained.
-The prediction accuracy is improved by correcting the data.
-The feedback loop continuously improves the accuracy of the cleansing process. More specifically, the user gives feedback on the result of the cleansing process of the cleansing module 212, so that the annotated data of the cleansing process can be obtained. By updating the rules and machine learning model of each cleansing process based on the data, the accuracy of the cleansing process itself is improved. By rotating this human-in-the-loop, the accuracy of cleansing processing is continuously improved during operation.

FIG. 15 is an example of the feature amount generation processing flow 1500.
In the feature amount generation processing flow 1500, the feature amount generation module 213 generates and adds a feature amount for improving the accuracy of the machine learning model to the sales data after cleansing while making the form suitable for inputting the model. It is a process.
The feature amount generation module 213 acquires the cleansed sales data from the cleansed data DB 223 (step 1510).
The feature amount generation module 213 generates and adds a feature amount for improving the accuracy of the machine learning model (step 1520). The cleansing module 212 can be used by customizing various feature quantity generation modules, and external data can also be added to the feature quantity.
The feature amount generation module 213 outputs the sales data after cleansing to which the feature amount is added (step 1530).

FIG. 16 is an example of the sales forecast model learning processing flow 1600.
The sales forecast model learning process flow 1600 is a process in which the sales forecast model learning module 214 generates and learns a machine learning model for forecasting sales.
The sales forecast model learning module 214 acquires the sales data after cleansing with the added features (step 1610).
The sales forecast model learning module 214 inputs a plurality of feature quantities calculated from the cleansing data, and generates or learns a sales forecast model that outputs the sales forecast data by machine learning (step 1620). Various hyperparameters are adjusted for the input features.

The sales forecast model learning module 214 calculates the accuracy of a plurality of generated sales forecast models, and selects a sales forecast model whose accuracy is equal to or higher than a predetermined value as a champion model (step 1640).
For example, the predetermined value is the most accurate of all the calculated sales forecast models, and the most accurate sales forecast model is selected as the champion model. In addition, a plurality of sales forecasting models whose accuracy is equal to or higher than the threshold value may be selected.
The sales forecast model learning module 214 stores the sales forecast model of the selected champion model and outputs it for sales forecast (step 1650).
Hyperparameters and sales forecast models can be changed and combined, and multiple sales forecast models are generated and the one with the highest accuracy is selected as the champion model.

The high accuracy of the sales forecast model means that the error when comparing the result of performing sales forecast processing on the input data of the verification data and the correct answer data that is the prediction target of the verification data is the minimum. Or it is less than or equal to a predetermined threshold. The accuracy evaluation can be performed by a holdout method, a cross validation method, or the like.
The verification data is, for example, data having the same structure as the learning data but different from the data used as the learning data. For example, the sales data up to August is used as learning data, and the sales data after September at different times are used as verification data.
For example, in the example of FIG. 8, the input data of the verification data is a product ID 801 other than the number of units sold 806, a category 802, a selling price 805, a month 807, a day 808, and a discount rate 809. It is the value of the number of units sold.

FIG. 17 is an example of the sales forecast processing flow 1700.
The sales forecast processing flow 1700 is a process in which the sales forecast module 215 performs sales forecast using a champion model.
The sales forecast module 215 acquires the sales data after cleansing to which the feature amount is added (step 1710).
The sales forecast module 215 executes the sales forecast process using the champion model (step 1720).
The sales forecast module 215 stores the forecast result in the forecast result DB 224 (step 1730).

FIG. 20 is an example of the sales forecast display screen 2000.
The sales forecast display module 312 of the user terminal 102 cooperates with the visualization module 216 of the sales forecast server 101, and displays the sales forecast result on the output device 305 such as a monitor. Specifically, the Query API of the visualization module 216 extracts the prediction result information of the prediction result DB 224, and the Web Application generates a display screen and sends it to the user terminal 102.
It should be noted that the configuration may be such that only the information for screen generation is transmitted to the user terminal 102 and the sales forecast display module 312 of the user terminal 102 generates the screen without generating the screen by Web App.
Further, the configuration may be such that the administrator terminal 103 displays instead of the user terminal 102.

The sales forecast display screen 2000 can display sales forecasts for a plurality of products in a plurality of categories 2001 and 2002, and in the example of the figure, the sales forecast data of the product A2003 in the category X2001 selected by the user is displayed. Has been done.
The sales forecast display screen 2000 displays the forecast result information 2030 read from the forecast result DB 224.
Further, on the graph display 2010, the measured value and the predicted value of sales are displayed in a state in which they can be distinguished from each other with the boundary 2011 as a boundary.
In the example of the figure, the number of units sold is displayed as the value of sales, but it can also be displayed by the sales amount.
The forecast result is displayed in a form that can be edited by the user, and the edited information can be used by the sales forecast display module 312 for ordering.

According to the first embodiment described above, future demand can be predicted by using machine learning based on the past actual data.
By forecasting demand using machine learning, it is possible to forecast future demand with higher accuracy than humans make forecasts based on empirical rules, and by placing orders based on demand forecasts, opportunity loss It also has the effect of greatly reducing surplus inventory and disposal.
In addition, by formulating production management, order management, production plan formulation, purchase volume adjustment, discount strategy, etc. in conjunction with demand forecasting using machine learning, work efficiency and accuracy related to these management and plan formulation can be improved. Can be improved.

In this embodiment, an example of predicting future sales of each product based on past sales data has been described for the retail industry. The sales data may be the number of products sold or the sales amount.
Further, the target of the demand forecast of this embodiment is not limited to the retail industry. For example, in the advertising industry, it can be applied to the order amount of advertising expenses, the number of orders, etc., and in the logistics industry, it can be applied to logistics. It can also be applied to the order amount and the number of distribution orders.

In the first embodiment, a mechanism for improving the accuracy of the cleansing process by receiving feedback from the user to confirm and correct the sales data after cleansing has been described.
In the second embodiment, the sales prediction system 1 evaluates the result of the cleansing process and provides a mechanism for further improving the accuracy of the cleansing process.
In the example of the figure illustrating the data flow of the sales forecast system 1 of FIG. 11, the sales forecast model learning module 214 trains the sales forecast model using the data generated by the feature amount, evaluates the accuracy, and evaluates the accuracy. The evaluation result is fed back to the cleansing module 212.
The cleansing module 212 determines the selection and combination of cleansing processes to be actually used according to the feedback result, and sets the values of the parameters used in each cleansing process.

FIG. 18 is an example of a cleansing process modification / change process flow 1800.
The cleansing module 212 calculates a plurality of post-cleansing data by changing some settings such as parameters of each cleansing process to execute the cleansing process, or by executing a plurality of different types of cleansing processes. (Step 1810).
The feature amount generation module 213 executes the feature amount generation processing flow 1500 of FIG. 15, and generates and adds a feature amount to a plurality of cleansed sales data (step 1820).
The sales forecast model learning module 214 generates a plurality of sales forecast models by training the sales forecast model for the plurality of cleansed sales data to which the feature quantity is added (step 1830).

The sales forecast model learning module 214 evaluates the accuracy of a plurality of generated sales forecast models (step 1840).
The sales forecast model learning module 214 selects a cleansing process setting or combination whose accuracy is equal to or higher than a predetermined value (step 1850), and notifies the cleansing module 212 (step 1860). For example, the predetermined value is the most accurate of all the calculated sales forecast models, and the most accurate sales forecast model is selected. In addition, a plurality of sales forecasting models whose accuracy is equal to or higher than the threshold value may be selected.
The notified cleansing module 212 executes the cleansing process by using the notified high-precision cleansing process settings and combinations.

In step 1810, it has been described that a plurality of cleansing processes and a plurality of cleansing processes are executed in parallel while changing the settings thereof. First, for each cleansing process, accuracy evaluation is performed while changing the settings. After selecting the optimum setting for each cleansing process, accuracy evaluation may be performed on a combination of a plurality of cleansing processes to determine the optimum combination.
The accuracy evaluation can be performed by a holdout method, a cross validation method, or the like. Further, the search for the optimum setting (parameter) can be performed by a random search for randomly searching for a value, a Bayesian optimization using probabilistic prediction, or the like.

FIG. 9 is an example of the accuracy evaluation information 900.
The accuracy evaluation information 900 indicates the accuracy of the sales forecast when one cleansing process is performed and the parameters are changed.
For example, if the target cleansing process 901 is A, it is a price abnormality detection process. The price abnormality detection process is executed as a cleansing process, and the accuracy (prediction error) for a plurality of sales forecast models generated when various parameters are changed is calculated.
The first parameter 902 indicates the lower limit of the ratio of the normal selling price to the list price (abnormal if it falls below), and the second parameter 903 indicates the upper limit of the ratio of the selling price that is abnormal to the list price (abnormal if it exceeds).

The prediction error 904 indicates an error when the result of performing the sales prediction process using the input data of the verification data as an input and the correct answer data which is the prediction target of the verification data are compared.
In the example of FIG. 9, the minimum prediction error 905 is 0.13, and the most accurate sales prediction is performed. The sales prediction model learning module 214 notifies the cleansing module 212 of the combination of the setting parameters. Upon receiving the notification, the cleansing module 212 executes the process of the cleansing process A using these parameters. As an example of another cleansing process, for example, when the cleansing process is the missing value complement, the missing value is stored at zero. A sales forecast model is generated and evaluated for each case where the missing value is stored as the value of the previous day.

FIG. 10 is an example of another accuracy evaluation information 1000.
The accuracy evaluation information 1000 indicates the accuracy of sales forecast when a plurality of cleansing processes are combined and executed.
For example, regarding the target cleansing process 1001, A is a price abnormality detection process, B is a set product process, C is a missing value complement process, and D is a JAN link process.
The prediction error 1002 indicates an error when the result of performing the sales prediction process on the input data of the verification data for evaluation and the correct answer data of the data for verification are compared.
In the example of FIG. 10, the one with the minimum prediction error of 1003 in which the combination of A, C, and D is performed as the target cleansing process has an error of 0.09, and the most accurate sales prediction is performed.
The sales forecast model learning module 214 notifies the cleansing module 212 of the combination of the cleansing processes A, B, and C. Upon receiving the notification, the cleansing module 212 executes these cleansing processes A, B, and C.

In the second embodiment, it is possible to carry out a plurality of cleansing processes, and the sales forecast model learning module 214 analyzes the sales data to select a cleansing process suitable for the sales data, and the cleansing module 212 selects the cleansing process. I gave feedback.
The selection of this cleansing process was executed by the sales forecast model learning module 214, but the result of the sales forecast process and the evaluation result are displayed on the user terminal 102 and the administrator terminal 103 by the visualization module 216, and the user and management. Can also make choices through these terminals. In addition, it can be set based on a predetermined rule based on the type of business of the target company of the sales data. In addition, as data accumulates, statistical analysis can be used to make selections.

In the third embodiment, the cleansing module 212 generates a cleansing process determination model using machine learning using the execution results of the first and / or the second embodiment, and the cleansing process determination model is used to generate sales data. Select the cleansing process to apply to.

FIG. 21 is an example of a cleansing process determination model generation process flow 2100.
The cleansing module 212 acquires standardized sales data from the standard data DB 222 (step 2110).
The cleansing module 212 acquires information regarding the cleansing process applied to the sales data used as a result of executing the processes shown in the first and / or second embodiments (step 2120).
The cleansing module 212 acquires accuracy information of the sales forecast model generated and learned by the sales forecast model learning module 214 used as a result of executing the processes shown in the first and / or second embodiments (step 2130). ..

The cleansing module 212 uses the acquired sales data and information identifying one or more cleansing processes used in the sales forecast model whose accuracy is equal to or higher than a predetermined value as teacher data, and inputs the sales data. Then, a cleansing process determination model in which the output is information for specifying the cleansing process is generated by machine learning (step 2140). The acquired sales data may be subjected to preprocessing such as feature quantity generation.
The cleansing module 212 inputs the acquired stylized sales data into the generated cleansing processing determination model, outputs information for specifying one or a plurality of cleansing processes suitable for this data, and identifies one. Alternatively, the cleansing process is executed by applying a combination of a plurality of cleansing processes to the sales data.

In the example of FIG. 9, a machine learning model that optimizes the parameters of individual cleansing processes is generated.
In the example of FIG. 10, a machine learning model that optimizes a combination of a plurality of cleansing processes is generated.
These processes can be performed in combination, and it is possible to optimize the combination of a plurality of cleansing processes while optimizing the parameters of the individual cleansing processes.

According to Example 3 described above, by accumulating correct answer data for the selection of the optimum cleansing process, statistics are obtained in addition to the configuration of Example 1 or 2 or in combination with the configuration of Example 1 or 2. Objective analysis, rule-based, and machine learning make it possible to set up multiple cleansing processes and optimize combinations.
In addition, by generating and using a cleansing process judgment model that selects a combination of multiple cleansing processes by machine learning, the accuracy of sales forecast is improved compared to the manual selection of cleansing processes, and the search time for selection is improved. There is an effect that the speed of calculating the sales forecast is improved by shortening.

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. Further, 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. Further, it is possible to add / delete / replace a part of the configuration of each embodiment with another configuration.

Further, each of the above configurations, functions, processing units, processing means and the like may be realized by hardware by designing a part or all of them by, for example, an integrated circuit. Further, 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 memory, a recording device such as a hard disk or SSD (Solid State Drive), or a recording medium such as an IC card, SD card, or DVD.

In addition, 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.
It should be noted that the above-mentioned embodiment discloses at least the configuration described in the claims.

1 ... Sales forecast system 1, 101 ... Sales forecast server, 102 ... User terminal, 103 ... Administrator terminal, 211 ... Standardization module, 212 ... Cleansing module, 213 ... Feature quantity generation module, 214 ... Sales forecast model learning module , 215 ... Sales forecast module, 216 ... Visualization module, 221 ... Corporate DB, 222 ... Standard data DB, 223 ... Post-cleansing data DB, 224 ... Forecast result DB, 311 ... Cleansing result display module, 312 ... Sales forecast display module, 321 ... User terminal information, 411 ... Management module, 421 ... Administrator terminal information

Claims (26)

The sales data acquisition department that acquires product sales data,
A cleansing processing unit that calculates post-cleansing data by executing a combination of multiple cleansing processes that modify the data for forecasting sales for the acquired sales data.
A sales forecast model that inputs the cleansing data and outputs the sales forecast data is generated by machine learning, the accuracy of the generated sales forecast model is calculated, and the sales forecast model whose accuracy is equal to or higher than a predetermined value is obtained. The model learning processing unit to select and
A sales forecast processing unit that outputs sales forecast data based on the selected sales forecast model, and
Demand forecast server with. The cleansing processing unit is
The sales data and the post-cleansing data generated by executing the cleansing process are output in a comparable form.
The demand forecast server according to claim 1, which outputs a display that accepts at least one of the modification of the post-cleansing data, the confirmation of the availability of the post-cleansing data, and the modification of the setting of the cleansing process. The display is output to the user terminal and is output to the user terminal.
The demand forecast server according to claim 2, which receives at least one of modification of the post-cleansing data, confirmation of availability of the post-cleansing data, and modification of the setting of the cleansing process from the user terminal. The cleansing processing unit is
Conditions and processes for executing the cleansing process based on at least one of the modified post-cleansing data, the confirmed post-cleansing data, or the modified post-cleansing process settings. The demand forecast server according to claim 2 or 3, which updates at least one of the contents or settings. The cleansing processing unit is
Using the sales data and the post-cleansing data confirmed to be usable as teacher data, a cleansing processing model in which the input is the sales data and the output is the post-cleansing data is generated by machine learning.
The demand forecast server according to any one of claims 1 to 4, which executes the cleansing process using the cleansing process model and outputs post-cleansing data estimated from the input sales data. The cleansing processing unit is
The sales data and the post-cleansing data generated by executing the plurality of cleansing processes are output in a comparable form.
The demand forecast server according to any one of claims 1 to 5, which outputs a display accepting selection of the cleansing process used for sales forecast from the plurality of cleansing processes. The demand forecast server according to any one of claims 1 to 6, wherein the cleansing process includes a missing value complementing process for complementing a part of the sales data. The demand according to any one of claims 1 to 7, wherein the cleansing process includes an abnormality detection process for detecting and correcting a portion where the sales data may be erroneous as compared with other corresponding sales data. Forecast server. The method according to any one of claims 1 to 8, wherein the cleansing process includes a mass customer order detection process for excluding sales when it is detected that sales exceeding a predetermined value are generated within a predetermined period. Demand forecast server. The demand forecast server according to any one of claims 1 to 9, wherein the cleansing process includes a product identification information linking process for associating product identification information before and after switching with respect to a product whose product identification information has been switched. The demand forecast server according to any one of claims 1 to 10, wherein the cleansing process includes a set product process for adding the sales of the set product to the sales of each individual product included in the set product. The demand forecast server according to any one of claims 1 to 11, wherein the cleansing process includes an abnormality detection process for a variable indicating a sales state. The cleansing processing unit calculates a plurality of post-cleansing data in which the settings of the cleansing processing are changed, and the cleansing processing unit calculates.
The model learning processing unit generates a plurality of the sales forecast models based on the plurality of post-cleansing data, executes evaluation of the plurality of sales forecast models, and performs evaluation of the plurality of sales forecast models.
The demand forecast according to any one of claims 1 to 12, wherein the cleansing processing unit executes the cleansing processing by using the setting of the cleansing processing corresponding to the sales forecast model whose evaluation is equal to or higher than a predetermined value. server. The cleansing processing unit calculates a plurality of post-cleansing data by combining a plurality of cleansing processes.
The model learning processing unit generates a plurality of the sales forecast models based on the plurality of post-cleansing data, executes evaluation of the plurality of sales forecast models, and performs evaluation of the plurality of sales forecast models.
The method according to any one of claims 1 to 12, wherein the cleansing processing unit executes a plurality of the cleansing processes based on a combination of the cleansing processes corresponding to the sales forecast model whose evaluation is equal to or higher than a predetermined value. Demand forecast server. The cleansing processing unit is
The sales data and the information specifying one or more of the cleansing processes used in the sales forecast model whose accuracy is equal to or higher than a predetermined value are used as teacher data, the input is the sales data, and the output is the cleansing process. A cleansing process judgment model that is used as information to specify is generated by machine learning,
The demand forecast server according to any one of claims 1 to 14, which outputs information specifying the cleansing process to be executed for the input sales data by using the cleansing process determination model. The accuracy of the predetermined value or more is the accuracy of the most accurate value among all the calculated sales forecast models, and the sales forecast model whose selected accuracy is the predetermined value or more is all calculated. The demand forecast server according to any one of claims 1 to 15, which is the most accurate of the sales forecast models. Get product sales data and
The post-cleansing data is calculated by executing a combination of a plurality of cleansing processes that modify the data for forecasting the acquired sales data.
A sales forecast model that inputs the cleansing data and outputs the sales forecast data is generated by machine learning.
The accuracy of the plurality of generated sales forecast models is calculated, and the sales forecast model whose accuracy is equal to or higher than a predetermined value is selected.
A demand forecasting method that outputs sales forecast data based on the selected sales forecast model. The sales data and the post-cleansing data generated by executing the cleansing process are output in a comparable form.
The demand forecasting method according to claim 17, wherein a display that accepts at least one of modification of the post-cleansing data, confirmation of availability of the post-cleansing data, or modification of the setting of the cleansing process is output. The display is output to the user terminal and is output to the user terminal.
The demand forecasting method according to claim 18, which receives at least one of the modification of the post-cleansing data, the confirmation of the availability of the post-cleansing data, and the modification of the setting of the cleansing process from the user terminal. Conditions and processes for executing the cleansing process based on at least one of the modified post-cleansing data, the confirmed post-cleansing data, or the modified post-cleansing process settings. The demand forecast server according to claim 18 or 19, which updates at least one of the contents or settings. Using the sales data and the post-cleansing data confirmed to be usable as teacher data, a cleansing processing model in which the input is the sales data and the output is the post-cleansing data is generated by machine learning.
The demand forecasting method according to any one of claims 17 to 20, wherein the cleansing process is executed using the cleansing process model, and post-cleansing data estimated from the input sales data is output. The sales data and the post-cleansing data generated by executing the plurality of cleansing processes are output in a comparable form.
The demand forecasting method according to any one of claims 17 to 21, which outputs a display accepting selection of the cleansing process used for sales forecasting from a plurality of the cleansing processes. Calculate multiple post-cleansing data with changed cleansing processing settings,
Generate a plurality of the sales forecast models based on the plurality of post-cleansing data, perform evaluation of the plurality of sales forecast models, and perform evaluation of the plurality of sales forecast models.
The demand forecasting method according to any one of claims 17 to 22, wherein the cleansing process is executed using the cleansing process setting corresponding to the sales forecast data whose evaluation is equal to or higher than a predetermined value. Combine multiple cleansing processes to calculate multiple post-cleansing data
Generate a plurality of the sales forecast models based on the plurality of post-cleansing data, perform evaluation of the plurality of sales forecast models, and perform evaluation of the plurality of sales forecast models.
The demand forecasting method according to any one of claims 17 to 22, wherein a plurality of the cleansing processes are executed based on a combination of the cleansing processes corresponding to the sales forecast data whose evaluation is equal to or higher than a predetermined value. The accuracy of the predetermined value or more is the accuracy of the most accurate value among all the calculated sales forecast models, and the sales forecast model whose selected accuracy is the predetermined value or more is all calculated. The demand forecasting method according to any one of claims 17 to 24, which is the most accurate of the sales forecasting models. A program that causes a demand forecast server to execute each step of the demand forecast method according to any one of claims 17 to 25.

Images