JP2020009157A - Sale prediction device and program - Google Patents

Abstract

To provide a sale prediction device and a program capable of performing sale prediction considering interest and concern of a customer to commodities.SOLUTION: A sale prediction device 1 comprises: a first prediction part 4 for calculating a first prediction value for each identity in a specific period in future from time series data for each identity relating to past customer's action with respect to commodities arranged at a specific place by using a first prediction model; and a second prediction part 6 for calculating a second prediction value of sales of the commodities in the specific period by using a second prediction model based on the first prediction value for each calculated identity.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a sales prediction device and a program.

  2. Description of the Related Art In recent years, a sales floor planning support system capable of predicting and presenting means having a high sales effect based on a relationship between a sales fluctuation factor and a customer's purchase behavior has been proposed (for example, see Patent Document 1). ).

  The sales floor planning support system described in Patent Literature 1 includes a hypothesis setting unit that sets a sales fluctuation factor, a flow line measurement unit that measures the movement of a customer in a sales floor, and flow line data measured by the flow line measurement unit. Data storage means for accumulating data and sales data, comparison / verification means for comparing and verifying the difference between the hypothesis set by the hypothesis setting means and the flow line data measured by the flow line measurement means, and storage in the data storage means Means for analyzing the relationship between the obtained sales data and flow line data, sales prediction means for predicting sales from the results analyzed by the analysis means, and factors for presenting factors corresponding to the contents predicted by the prediction means Presentation means.

JP 2004-118454 A

  An object of the present invention is to provide a sales prediction device and a program capable of performing sales prediction in consideration of a customer's interest or interest in a product.

[1] Using a first prediction model, a first prediction value for each feature in a future specific period is calculated from time-series data for each feature relating to past customer behavior with respect to a product placed at a specific location. A first predictor,
A second prediction unit that calculates a second prediction value of the sales of the product in the specific period using a second prediction model based on the calculated first prediction value for each of the features. Equipped sales forecasting device.
[2] The sales prediction device according to [1], wherein the second prediction unit presents relation information between the first prediction value and the second prediction value for each feature.
[3] The sales prediction device according to [2], wherein the relationship information is a contribution ratio of the first prediction value to the second prediction value for each feature.
[4] The sales prediction device according to any one of [1] to [3], further including an acquisition unit configured to acquire time-series data for each of the features from history data regarding the past customer behavior.
[5] The first prediction unit calculates a prediction value for each of the features using the first prediction model according to a change rate of a range of the history data for acquiring the time-series data. 4].
[6] The sales prediction device according to [5], wherein the first prediction unit uses a time-series model as the first prediction model when the change rate of the range of the history data is smaller than a threshold. .
[7] The sales prediction device according to [5], wherein the first prediction unit uses a machine learning model as the first prediction model when the change rate of the range of the history data is equal to or larger than the threshold. .
[8] The acquisition unit according to any one of [4] to [7], wherein the acquisition unit adds reliability to individual feature values of the time-series data for each feature and inputs the value to the first prediction unit. The sales forecasting device described.
[9] The sales prediction device according to [8], wherein the first prediction unit corrects the reliability based on a feature relating to customer behavior in a virtual store on the network of the product.
[10] The sales prediction device according to any one of [1] to [9], wherein the specific place is a shelf, and the feature relating to the customer's behavior includes the number of persons who have performed the customer's specific behavior on the shelf. .
[11] The number of people who have performed the specific action includes the number of people who passed in front of the shelf, the number of people who stayed in front of the shelf for a predetermined time, and the number of people who focused on the product for a predetermined time. The sales prediction device according to [10].
[12] The first prediction unit according to [11], wherein the first prediction unit calculates the first prediction value using the number of persons who have performed the specific action on another product whose relevance of the product is equal to or greater than a threshold. Sales forecasting device.
[13] The sales prediction device according to any one of [1] to [12], wherein the features related to the behavior of the customer include the number of persons for each attribute of the customer.
[14] The sales prediction device according to [13], wherein the attributes of the customer include gender and age.
[15] The device according to any one of [1] to [14], wherein the second prediction unit calculates the second prediction value from the first prediction value and a feature value of environmental information. Sales forecasting device.
[16] A computer calculates a first prediction value for each feature in a future specific period from a time-series data for each feature relating to past customer behavior with respect to a product placed at a specific location by using a first prediction model. A first prediction unit calculated using the first prediction unit;
A second prediction unit that calculates a second prediction value of the sales of the product in the specific period using a second prediction model, based on the calculated first prediction value for each of the features; Program to function as

According to the inventions according to claims 1, 4, and 16, it is possible to perform sales prediction in consideration of the customer's interest and interest in the product.
According to the second and third aspects of the invention, it is possible to present related information for improving sales prediction.
According to the invention according to claims 5 to 7, by using the prediction model according to the property of the history data, it is possible to perform the sales prediction with higher accuracy as compared with the case of using the same prediction model regardless of the property of the history data. Can be.
According to the eighth and ninth aspects of the present invention, it is possible to perform sales prediction with higher accuracy than when no reliability is used.
According to the tenth and eleventh aspects, by analyzing the behavior of the customer in front of the shelf, it is possible to know the customer's interest in the product and the degree of the interest.
According to the invention according to claim 12, when the customer who purchases food such as onigiri has a high relevance between products such as a tendency to purchase drinks such as tea, the number of data used for prediction is small. The sales forecast can be made with high accuracy.
According to the invention according to claims 13 and 14, for a product whose interest or interest in the product differs depending on the attribute of the customer, the sales prediction can be performed with higher accuracy as compared with the case where the attribute of the customer is not used for the feature. It can be carried out.
According to the invention according to claim 15, it is possible to perform sales prediction by integrating environmental information generally used for prediction.

FIG. 1 is a diagram illustrating a schematic configuration example of a sales prediction system according to an embodiment of the present invention. 2A and 2B show an example of a table stored in the second storage unit. FIG. 2A shows an example of an attribute-based number of visitors table, and FIG. 2B shows an example of a shelf front action data table. FIG. FIG. 3 is a diagram for explaining functions of the environment information acquisition unit and the behavior data acquisition unit. FIG. 4 is a diagram illustrating a configuration for recommending a measure for increasing sales after the second prediction model of the second prediction unit is completed. FIG. 5 is a diagram illustrating a first modification of the sales prediction device. 6A to 6D are diagrams for explaining the reliability. FIG. 7 is a diagram illustrating a result of simulating the prediction accuracy of the prediction model according to the related art and the present embodiment.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In each of the drawings, components having substantially the same function are denoted by the same reference numerals, and redundant description is omitted.

[Summary of Embodiment]
The sales prediction device according to the embodiment of the present invention calculates a first predicted value for each feature in a future specific period from time-series data for each feature regarding past customer behavior with respect to a product placed at a specific place. A first prediction unit that calculates using the first prediction model; and a second prediction model that calculates a second prediction value of product sales in a specific period based on the calculated first prediction value for each feature. And a second prediction unit that calculates using

  The “specific place” is, for example, a shelf on which merchandise in an actual store is displayed. The “product” may be a single product or a plurality of products of the same type displayed on the same shelf. The same kind of products include, for example, rice balls with different ingredients and green tea PET bottles with different capacities and product names. The “customer behavior” includes visits to stores, traffic in front of shelves, stays in front of shelves, attention to products, purchase of products, and the like. "Feature" is an element that can affect sales. The features related to the behavior of the customer include, for example, the number of visitors to the store for each attribute of the customer, and the behavior in front of the shelf. The action in front of the shelf includes the number of people passing in front of the shelf (number of people passing), the number of people staying at the front of the shelf for a certain period of time (number of people staying), the number of people watching the product (the number of people watching), and the like. The “time-series data” includes a plurality of feature values having different measured dates and times. The feature value is a specific value of the feature. For example, when the feature is a day of the week, Sunday, Monday, and the like are feature values. The “specific period” may be one day or a period extending over a plurality of days such as one week. The “prediction model” executes a prediction algorithm using parameters obtained as a result of learning.

[Embodiment]
FIG. 1 is a diagram illustrating a schematic configuration example of a sales prediction system according to an embodiment of the present invention. This sales prediction system 10 can be applied to a real store 100 such as a convenience store, a department store, and a shopping center, for example. In the actual store 100, for example, a plurality of shelves 11 on which products are displayed, a terminal device 12 for a clerk to perform accounting processing, and first to fourth cameras 13A, 13B, 13C, 13D for photographing the inside of the store (These are simply referred to as “camera 13” when collectively referred to.) And an analysis device 14 that analyzes the behavior of a customer in the actual store 100 from an image captured by the camera 13 are arranged. The shelf 11 is an example of a specific place. Note that the analysis device 14 may be provided in a cloud server.

  The sales prediction system 10 further includes a first storage unit 15 that stores environment information 150, a second storage unit 16 that stores customer behavior information 160 and purchase data 120 analyzed by the analysis device 14, A sales prediction device 1 that calculates a predicted value of sales in a specific future period based on the information 150 and the behavior information 160. The behavior information 160 is an example of history data regarding past customer behavior.

  The sales prediction device 1 includes an environment information acquisition unit 2, an action data acquisition unit 3, a first prediction unit 4 that calculates a feature prediction value 41 based on environment information 150 and action information 160, and data for integrating data. An integration unit 5 and a second prediction unit 6 that inputs data integrated by the data integration unit 5 and calculates a predicted sales value 60 are provided. The behavior data acquisition unit 3 is an example of an acquisition unit. The feature prediction value 41 is an example of a first prediction value. The sales forecast value 60 is an example of a second forecast value.

  The sales prediction device 1 includes, for example, a control unit including a CPU (Central Processing Unit) and an interface of a computer, and a storage unit including a ROM (Read Only Memory), a RAM (Random Access Memory), a hard disk, and the like. And can be realized by using Each unit of the sales prediction device 1 functions when, for example, the CPU operates according to a program stored in the storage unit.

Physical store 100, area _E 1 _{to E 12} that the customer can pass therethrough. The route R shown in the drawing shows, as an example, the travel route of the customer. The customer's travel route corresponds to the customer's purchase behavior process in the physical store 100 (for example, visiting a store, passing in front of a shelf, staying in front of a shelf, paying attention to a product, purchasing a product, and the like).

  The shelves 11 include shelves 11a to 11i that display, for example, rice balls, bento, tea, cup ramen, confectionery, daily necessities, breads, alcohols, and magazines as commodities.

  The terminal device 12 is a computer device called a POS (Point Of Sale) register arranged in a counter. The customer who has entered the physical store 100 brings the picked-up merchandise to the counter and conducts accounting. When a customer sells a product sold at the physical store 100, the terminal device 12 performs processing for accounting, issues a receipt recording the purchase of the product, and generates purchase data 120 indicating purchase for each product. And so on. The terminal device 12 transmits the purchase data 120 to the analysis device 14 wirelessly or by wire, and stores the purchase data 120 in the second storage unit 16.

The camera 13 may be a moving image or a still image. The camera 13 transmits the captured image to the analysis device 14 wirelessly or via a wire together with the date and time of shooting. The first camera 13A captures the store that contains the area _{E 1 ~E 7, E 11,} E 12, a second camera 13B captures the store that contains the area _E 7 _{to E 12.} Third camera 13C includes an area _{E 1,} E _{11, E 12,} taking a store containing a terminal device 12 and the inlet 101. The fourth camera 13D is an omnidirectional camera that is provided on the ceiling and detects the position of the customer in the store. The arrangement of the camera 13 is not limited to the above. Further, the position of the customer may be acquired by using a technology such as position sensing using a beacon.

  The analysis device 14 analyzes the behavior of the customer based on the image and the shooting date and time transmitted from the camera 13 and the purchase data 120 transmitted from the terminal device 12, collects the analysis result as the behavior information 160, Is stored in the storage unit 16. Specifically, the analysis device 14 determines the gender of the customer who has visited the actual store 100 based on the image of the face of the customer (hereinafter also referred to as “face image”) in the image transmitted from the camera 13. The age is analyzed, the number of visitors to the store is acquired for each gender and age, and the number of visitors to the store for each gender and age (hereinafter also referred to as “attribute visitor by attribute”) is taken together with the shooting date and time as action information 160 as a second storage unit. 16 is stored. Gender and age are examples of customer attributes. Note that only one of the gender and age of the customer may be analyzed.

  In addition, the analysis device 14, based on the image transmitted from the camera 13, for each of the shelves 11 a to 11 i, the number of people passing in front of the shelves 11 a to 11 i (hereinafter, also referred to as “passing people”) and the shelves 11 a to 11 i. The number of persons staying at predetermined intervals (for example, 5 seconds or more and less than 10 seconds, 10 seconds or more) (hereinafter also referred to as “number of staying persons”) and the products displayed on the shelf 11 are determined in advance. The number of people (hereinafter, also referred to as “number of people to watch”) at each time (for example, 1 second or more, less than 5 seconds, and 5 seconds or more) is acquired, and these are taken as action information 160 together with the shooting date and time and stored in the second storage unit 16. To memorize. The analysis device 14 can detect the direction of the face from the face image and specify the product of interest based on the direction of the face, thereby acquiring the number of passers, the number of stays, and the number of persons of interest for each of the shelves 11a to 11i. it can.

  The environment information 150 stored in the first storage unit 15 includes various features that can affect the sales of the product. Features include, for example, weather (eg, temperature, humidity, wind speed, precipitation, weather), calendar (eg, year, month, week, day, holiday, holiday, season, Golden Week (GW), Bon Festival, New Year's holiday) , Events (eg, sales, campaigns, etc.). The environment information 150 is information on past, present, and future plans or forecasts.

  The behavior information 160 stored in the second storage unit 16 includes an attribute-based number of visitors table 161 (see FIG. 2A) and a shelf front behavior data table 162 (see FIG. 2B).

  FIG. 2 is a diagram illustrating an example of the tables 161 and 162 stored in the second storage unit 16. FIG. 2A is a diagram illustrating an example of an attribute-based visitor number table 161, and FIG. 2B is a diagram illustrating an example of a shelf front behavior data table 162. As shown in FIG. 2 (a), the table of visitor numbers by attribute 161 includes a date of visit and a plurality of genders and ages such as male teens, male 20s,..., Female 60s,. Items. As shown in FIG. 2B, the shelf front behavior data table 162 includes the date of visit, the number of passers, the number of stays (5 seconds or more and less than 10 seconds), the number of stays (10 seconds or more), There are a plurality of items, such as seconds or more and less than 5 seconds, and the number of people to watch (5 seconds or more). For example, the number of staying persons (from 5 seconds to less than 10 seconds) and the number of attention persons (from 1 second to less than 5 seconds) are considered to be customers who have entered the actual store 100 in order to purchase the product. In addition, since the number of staying persons (10 seconds or more) and the number of attention persons (5 seconds or more) may be confused about the purchase of the product, if the product is not purchased, improvement other than the layout may be performed. It can help you make recommendations.

  The data integration unit 5 integrates the feature prediction value 41 predicted by the first prediction unit 4 and the binary strings 21 b, 22 b, and 23 b obtained by the environment information acquisition unit 2 and outputs the result to the second prediction unit 6.

  FIG. 2 is a diagram for explaining functions of the environment information acquisition unit 2 and the behavior data acquisition unit 3.

[Configuration of environmental information acquisition unit]
From the first storage unit 15, the environment information acquisition unit 2 retrieves, from the first storage unit 15, a feature value 21a of a feature (for example, temperature, humidity, wind speed, precipitation, and weather) and a feature (for example, year, month, week, and day of the week) related to a calendar. , Holidays, holidays, seasons, Golden Week (GW), Bon Festival, New Year's holidays) and feature values 23a of features related to events (eg, sale, campaign, etc.).

  The environment information acquisition unit 2 converts the feature value 21a relating to the weather into a binary string 21b with reference to the corresponding binary table 21, and converts the feature value 22a relating to the calendar into a binary string 22b with reference to the corresponding binary table 22. , And converts the feature value 23 a relating to the event into a binary string 23 b with reference to the corresponding binary table 23, and outputs them to the data integration unit 5. The binary string is composed of a plurality of binary values (0 or 1). The binary string is an example of a value obtained by digitizing the feature value.

  For example, when predicting the sales of ice cream on the last Sunday of July 2018, as the feature value 21a relating to the weather, for example, the temperature value and humidity predicted on the day are the feature values 21a, respectively. As the feature value 22a for the calendar, for example, July, 29, and Sunday of 2018 are the feature values 22a, respectively. Note that each of the feature values 21a, 22a, and 23a may be one or two or more.

[Configuration of behavior data acquisition unit]
The behavior data acquisition unit 3 obtains, from the behavior information 160 stored in the second storage unit 16, data on customer behavior required for sales prediction, that is, the feature value 31 of the number of store visitors by attribute and the feature value 32 of shelf front behavior data. It acquires and generates time-series data 310 including a plurality of feature values 31 and time-series data 320 including a plurality of feature values 32, and outputs them to the first prediction unit 4. Hereinafter, the feature value 31 and the time series data 310 related to the number of visitors to each store by attribute and the feature value 32 and the time series data 320 related to the shelf front action data are also referred to as action data. The feature value 32 and the time-series data 320 relating to the pre-shelf action data are also referred to as pre-shelf action data.

(Data in chronological order)
The time-series data 310 and 320 acquired by the behavior data acquisition unit 3 include a plurality of time-series feature values, and are, for example, data arranged in order from a new feature value at the top to an old feature value. The time-series data 310 and 320 can be represented by the following equation (1), for example.
{Y _t } ^T _{t = 1} = {y ₁ , y ₂ ,..., Y _T } (1)
Here, y: feature value, T: number of feature values, t = 1 to T: different time

  For example, when estimating the sales of ice cream on the last Sunday of July 2018, the time-series data 310 of the number of visitors to each attribute is, for example, the number of visitors to the store one week before the current day, the number of visitors to the store two weeks ago, A plurality of periodic feature values 31 such as the number of visitors to the store a week ago,... The time-series data 320 of the number of passers in the shelf front behavior data includes, for example, a plurality of periodic features such as the number of passers one week before, the number of passers two weeks ago, the number of passers three weeks ago,. The value may be 32.

(Length of data acquisition period)
The length of the range (also referred to as a period) of the behavior information 160 from which the behavior data acquisition unit 3 acquires the time-series data 310 and 320 may be determined as follows, for example. The period of the action information 160 is defined by a prediction model used by the first prediction unit 4.

(I) In the case of short term When the change rate of the range of the action information 160 for acquiring the time-series data 310 and 320 changes abruptly at or above the threshold value, for example, the data acquisition period is relatively short (for example, one month). The following may be used.

(Ii) In the case of long term When the change rate of the range of the action information 160 for acquiring the time-series data 310 and 320 is less than the threshold and changes slowly, for example, the data acquisition period is set to a relatively long term (for example, two months). Above).

(Whether the length of the data acquisition period is fixed or variable)
Whether the period of the action information 160 from which the action data obtaining unit 3 obtains the time-series data 310 and 320 may be fixed or variable may be determined as follows, for example.

(I) Fixed Case In the case of the time-series model, the period may be a fixed period (the last one week, the same day in the past four weeks, or a combination thereof). When the time-series data hardly changes over a long period (for example, one month or more), an average value may be used, or only the latest one day may be used. In the case of a machine learning model, the period may be a fixed period.

(Ii) In the case of fluctuation In the case of a time-series model, a period may be dynamically set for each factor that affects sales, such as setting a short period when the topic starts to be talked about in SNS. Further, when the value of the action data rapidly increases or decreases in the latest time, only the period may be used, or the weight may be increased as soon as possible. If the values differ greatly only for one day, the weight for that day may be set to zero. The period and the weight of the model may be changed according to the fluctuation in the fixed period. That is, if the latest fluctuation in the fixed period is large, the latest history data may be used, and if the latest fluctuation is small, long-term history data may be used. As for the model, when the fluctuation pattern during the fixed period has no regularity (such as randomly changing), it has a normal distribution, and when it has autoregression (correlates with its own past value), it has a time series. If the analysis model and the fluctuation pattern fluctuate according to a certain factor (such as being affected by weather or day of the week), a machine learning model may be used.

[Configuration of First Prediction Unit]
The first prediction unit 4 performs an operation using the first prediction model based on the time-series data 310 and 320 output from the behavior data acquisition unit 3, calculates a feature prediction value 41 for each feature, Is output to the data integration unit 5.

(Type of prediction model)
The first prediction unit 4 uses the first prediction model described below in accordance with the change rate of the range (for example, a period of a few months) of the behavior information 160 for acquiring the time-series data 310 and 320, and Is calculated for each feature.

(I) Time Series Model When the change rate of the range of the action information 160 is smaller than the threshold, the first prediction unit 4 uses the time series model as the first prediction model. The time series model can be realized using, for example, an autoregressive (AR) model, a moving average (MA) model, an autoregressive moving average (ARMA) model, or the like. As the change rate, for example, a variance or a difference between a maximum value and a minimum value of data may be used.

(Ii) Machine learning model The first prediction unit 4 uses a machine learning model as the first prediction model when the change rate of the range of the action information 160 is equal to or larger than a threshold. The machine learning model can be realized using, for example, a neural network, an SVM (Support Vector Machine), or the like. The machine learning model may be supervised learning or unsupervised learning. In the case of supervised learning, actual data corresponding to the time-series data 310 and 320 and actual data corresponding to the feature predicted value 41 are used.

(Modification of time series data)
The data output to the first prediction unit 4 is not limited to the above-mentioned behavior data such as the number of visitors to each store, the number of people passing, the number of people staying, and the number of people to watch, but is calculated by multiplying them by the conversion rate. It may be information. As the conversion rate, for example, the number of passers / number of visitors, the number of stays / number of passers can be considered.

(Prediction algorithm of the first prediction model)
When a time-series model is employed as the first prediction model, the first prediction unit 4 is, for example, one week before the current day (prediction date) of the time-series data 310 of the number of visitors to the store (for example, male teens) by attribute. , The number of visitors to the store two weeks ago, the number of visitors to the store three weeks ago,... May be used as the feature prediction value 41 of the number of visitors to each attribute, and the weighted average (weighted average) ) May be used. Note that the feature prediction value 41 may also be obtained by using an addition average or a weighted average with respect to shelf front data such as the number of passers, the number of stays, and the number of attendants.

[Configuration of Second Prediction Unit]
The second prediction unit 6 includes a binary string 21b for weather, a binary string 22b for calendar, and a binary string 23b for event output from the environment information acquisition unit 2 via the data integration unit 5, and a first prediction unit. 4, the number of visitors to each store, the number of passers, the number of stays, and the feature prediction value 41 for each number of persons of interest, which are output via the data integration unit 5, are input using the second prediction model for a specific period, for example, Calculate sales forecast value for a specific day.

  The second prediction model can be realized using a machine learning model such as a neural network or an SVM (Support Vector Machine). The machine learning model may be supervised learning or unsupervised learning. In the case of supervised learning, actual data corresponding to the feature predicted value 41 and actual purchase data corresponding to the predicted sales value 60 are used. In a prediction model based on general machine learning, information on the day's weather, calendar, events, etc. is used.However, the behavior data of the day, such as the number of store visitors by attribute and pre-shelf behavior data, cannot be obtained in advance. Is used.

  FIG. 4 is a diagram showing a configuration for recommending a measure for increasing sales after the second prediction model of the second prediction unit 6 is completed. The prediction model learns based on the purchase data 120 and the performance information 160. After the completion of the second prediction model, it is possible to set the behavior data to an arbitrary value and simulate sales.

  The second prediction unit 6 may present relationship information between the feature prediction value 41 and the sales prediction value 60 for each feature. The relationship information may be, for example, a contribution ratio of the feature prediction value 41 for each feature to the sales prediction value 60.

  It is also possible to extract behavior data that affects sales and recommend measures based on the interests and interests of customers. For example, for the number of visitors by attribute, improvement of the entrance guidance according to the attribute, for the number of passers-by, the layout of the store is improved, for the number of staying people, POP installation, product layout change, etc. In addition, it is possible to recommend a POP installation, a change in the number of faces (the number of identical products arranged on the display shelf in the horizontal direction), and the like. For example, if the number of visitors to the store increases and the number of passers in front of a certain shelf increases with this, but the number of staying or attention people on the shelf does not increase, the customer is interested in other products, Since there is no interest in the products arranged on the shelf, it can be suggested that the products hardly contribute to the increase in sales.

(Modification 1)
FIG. 5 is a diagram illustrating a first modification of the sales prediction device. In the first modification, the behavior data acquisition unit 3 adds the reliability 311 to the feature value 31 included in the time-series data 310 of the number of visitors by attribute shown in FIG.

  Since the behavior data is collected by camera image analysis, beacon sensing, or the like, as shown in FIG. 6A, the behavior data includes many outliers such as measurement errors and missing values. For this reason, as shown in FIG. 6B, when preprocessing such as outlier removal is performed, the number of data may be reduced and data necessary for prediction may be lost. Therefore, as shown in FIG. 6C, the distance between each value of the action data and the median of the distribution (for example, the average value) is evaluated as the reliability, and the reliability is added as a feature. Numbers may be kept. As shown in FIG. 6D, the reliability increases as the distance to the median of the distribution decreases, and decreases as the distance increases. The reliability may be added not only to the time-series data 310 of the number of visitors by attribute but also to the time-series data 320 of the behavior in front of the shelf.

(Modification 2)
The customer behavior of a virtual store on a network such as the Internet, for example, an EC (Electronic Commerce) site may be fed back to the first prediction model of the first prediction unit 4. That is, based on behavior data (number of visitors / time spent / number of clicks / conversion rate, etc.) on the EC site, the behavior data to be used may be selected, and the weight and reliability of the behavior data may be determined. In this case, the behavior data may be modified to match the tendency of the EC site behavior data, or the reliability may be calculated. When the number of clicks on the target product is low, the number of attention persons may not be used, or the weight or reliability may be reduced.

(Modification 3)
The customer behavior items may be changed depending on the type of the target product. The pre-shelf action data may be measured for each attribute. When the target is a single product, the number of contacts may be added to the shelf front behavior data. When the target is a single product, the number of persons of interest may be calculated from the gaze measurement result. If there is a strong relationship between the products, shelf-behavior behavior data on a highly relevant product shelf may be used. That is, the first prediction unit 4 may calculate the feature prediction value 41 using the number of persons who have performed a specific action (for example, a shelf front action) of another product whose relevance of the product is equal to or larger than the threshold.

(Modification 4)
The history data of the features of the environment information 150 may be combined with the behavior data. In this case, a unique day may be extracted from the features of the environment information 150, and the weight of the customer behavior data on the corresponding day may be reduced. The special day refers to a day having an environment different from a normal business day, such as an event held at the actual store 100 or a special sale day.

  FIG. 7 is a diagram illustrating a result of simulating the prediction accuracy of the prediction model according to the related art and the present embodiment. The baseline model is a model in which the weather, calendar, and event of the environment information 150 are used as features. + (A) the number of visitors to the store, + (b) behavior in front of the shelves, + (a) and (b) are prediction models corresponding to the present embodiment, and Tests 1, 2 and 3 are 3, 4 and 2016, respectively. The last week of May is divided into test data, and the data from October 2015 to one week before the test period is divided into learning data. Further, the chart of FIG. 7B shows the accuracy of each sales prediction model for Tests 1, 2, and 3, and is an index indicating that the closer to 100%, the higher the accuracy. Here, Sales SMA is the moving average of the number of units sold for the past four weeks on the same day, the base model is a machine learning model using weather / calendar / event information, + (a), + (b), + (a), (B) is a machine learning model obtained by adding both the number of visitors and the behavior before shelves to the features of the baseline model, the Customer model is a machine learning model using the number of visitors and the behavior before shelves as features, and the Ensemble model is the baseline. This is a model in which the average value of the prediction results of the model and the Customer model is used as the prediction value.

  As described above, the embodiments of the present invention have been described. However, the embodiments of the present invention are not limited to the above embodiments, and various modifications and implementations are possible without changing the gist of the present invention. is there.

  Each part of the control unit may be partially or entirely configured by a hardware circuit such as a reconfigurable circuit (FPGA: Field Programmable Gate array) or an application specific integrated circuit (ASIC).

  Further, some components of the above-described embodiment can be omitted or changed without departing from the scope of the present invention. In addition, steps can be added, deleted, changed, replaced, and the like in the flow of the above-described embodiment without changing the gist of the present invention. Further, the program used in the above embodiment can be provided by being recorded on a computer-readable recording medium such as a CD-ROM. In addition, the program used in the above embodiment can be stored in an external server such as a cloud server and used via a network.

DESCRIPTION OF SYMBOLS 1 ... Sales prediction apparatus, 2 ... Environmental information acquisition part, 3 ... Behavior data acquisition part, 4 ... First prediction part, 5 ... Data integration part, 6 ... Second prediction part, 10 ... Sales prediction system, 11, 11a-11i: shelf, 12: terminal device, 13, 13A-13D: camera, 14: analyzer, 15: first storage unit, 16: second storage unit, 21, 22, 23: binary table, 21a , 22a, 23a ... feature value, 21b, 22b, 23b ... binary string, 31, 32 ... feature value, 41 ... feature prediction value, 60 ... sales forecast value, 100 ... real store, 101 ... entrance, 120 ... purchase data, 150 ... environment information, 160 ... behavior information, 161 ... attribute-visit number table 162 ... rack before action table, 310, 320 ... time-series data, 311 ... reliability, 320 ... time-series data, _{E 1} -E ₁₂ ... Area, R… Le Door

Claims (16)

A first calculating unit that calculates a first predicted value for each feature in a future specified period from a time-series data for each feature related to past customer behavior with respect to a product placed at a specific location using a first prediction model; The prediction section of
A second prediction unit that calculates, using a second prediction model, a second prediction value of sales of the product in the specific period based on the calculated first prediction value for each of the features;
Sales forecasting device equipped with. The second prediction unit presents relationship information between the first prediction value and the second prediction value for each feature,
The sales prediction device according to claim 1. The relation information is a contribution rate of the first predicted value for each of the features to the second predicted value,
The sales prediction device according to claim 2. An acquisition unit that acquires time-series data for each of the features from the history data regarding the past customer behavior,
The sales forecasting device according to claim 1, further comprising: The first prediction unit calculates a prediction value for each feature using the first prediction model according to a change rate of a range of the history data for acquiring the time-series data.
The sales prediction device according to claim 4. The first prediction unit uses a time-series model as the first prediction model when the change rate in the range of the history data is smaller than a threshold.
The sales prediction device according to claim 5. The first prediction unit uses a machine learning model as the first prediction model when the change rate of the range of the history data is equal to or greater than the threshold.
The sales prediction device according to claim 5. The acquisition unit adds reliability to individual feature values of the time-series data for each feature and inputs the value to the first prediction unit.
The sales forecasting device according to any one of claims 4 to 7. The first prediction unit corrects the reliability based on a feature related to customer behavior in a virtual store on the network of the product,
The sales prediction device according to claim 8. The specific location is a shelf, and the features relating to the customer's behavior include the number of persons who have performed the customer's specific behavior on the shelf.
The sales forecasting device according to claim 1. The number of people who have performed the specific action includes the number of people who passed in front of the shelf, the number of people who stayed in front of the shelf for a predetermined time, and the number of people who focused on the product for a predetermined time,
The sales prediction device according to claim 10. The first prediction unit calculates the first prediction value using the number of persons who have performed the specific action on another product whose relevance of the product is equal to or greater than a threshold,
The sales prediction device according to claim 11. The features related to the behavior of the customer include the number of people for each attribute of the customer,
The sales forecasting device according to claim 1.   14. The sales prediction device according to claim 13, wherein the attributes of the customer include gender and age. The second prediction unit calculates the second prediction value from the first prediction value and the feature value of the environment information,
The sales prediction device according to any one of claims 1 to 14. Computer
A first calculating unit that calculates a first predicted value for each feature in a future specified period from a time-series data for each feature related to past customer behavior with respect to a product placed at a specific location using a first prediction model; The prediction section of
A second prediction unit that calculates a second prediction value of the sales of the product in the specific period using a second prediction model, based on the calculated first prediction value for each of the features;
Program to function as.