CN113554455B

CN113554455B - Store commodity analysis method, device and storage medium based on artificial intelligence

Info

Publication number: CN113554455B
Application number: CN202110748297.XA
Authority: CN
Inventors: 徐意; 吴勇民
Original assignee: Hangzhou Pinjie Network Technology Co Ltd
Current assignee: Hangzhou Pinjie Network Technology Co Ltd
Priority date: 2021-06-30
Filing date: 2021-06-30
Publication date: 2024-09-24
Anticipated expiration: 2041-06-30
Also published as: CN113554455A

Abstract

The application discloses a store commodity analysis method and device based on artificial intelligence and a storage medium, wherein the method comprises the following steps: acquiring historical purchase order data of a store; generating three-dimensional attribute coordinates of the store according to historical purchase order data of the store; carrying out K-Means clustering operation by using coordinate values of the three-dimensional attribute coordinates of the store; dividing the K-Means clustering operation result into a plurality of store attribute sets; constructing order prediction models corresponding to different store attribute sets; and inputting the historical purchase order data of the store to an order prediction model corresponding to the store attribute set to which the historical purchase order data belongs. The application has the advantages of providing the store commodity analysis method, the store commodity analysis device and the storage medium based on artificial intelligence, wherein the store commodity analysis method, the store commodity analysis device and the storage medium are used for comprehensively predicting the future commodity order condition of the store according to the historical data of the store and the type positioning of the store.

Description

Store commodity analysis method, device and storage medium based on artificial intelligence

Technical Field

The application relates to the field of electronic commerce data management, in particular to a store commodity analysis method and device based on artificial intelligence and a storage medium.

Background

The existing small and medium-sized convenience stores often carry out online commodity purchasing at suppliers through own channels, and cannot purchase goods in large quantities due to the characteristics of the small and medium-sized convenience stores, so that effective bargained price cannot be carried out with the suppliers, and meanwhile, due to the requirement of wholesale of the suppliers on purchasing quantity, the small and medium-sized convenience stores need to guarantee a certain scale every time of purchasing, so that inventory problems are caused.

From the perspective of suppliers, the scattered purchasing mode of small and medium-sized convenience stores leads to the increase of the warehouse cost of the suppliers, thereby leading to the high supply price.

In the related art, as disclosed in the chinese patent document with publication number CN112801759a, the technical solution is to integrate the buyers, sellers and carriers together through the e-commerce platform, so that the orders of multiple buyers can be "spelled" into one logistics order, thereby implementing a centralized purchasing and centralized distribution manner, and further reducing the purchasing cost of small and medium-sized convenience stores.

However, due to the "spelling" mode, the store orders are more random, the system needs to analyze the orders possibly generated in the future of the store to configure the goods sources and storages in order to meet the discrete demands of the store, the analysis schemes of the existing system are more considered from the individual view of the store, and in fact, the actual orders of different stores are quite different due to the difference of location and operation idea, if only one model is constructed to predict, accurate results cannot be obtained, and the data analysis is invalid.

Disclosure of Invention

In order to solve the defects of the prior art, the application provides an artificial intelligence-based store commodity analysis method, which comprises the following steps: acquiring historical purchase order data of a store; generating three-dimensional attribute coordinates of the store according to historical purchase order data of the store; carrying out K-Means clustering operation by using coordinate values of the three-dimensional attribute coordinates of the store; dividing the K-Means clustering operation result into a plurality of store attribute sets; constructing order prediction models corresponding to different store attribute sets; and inputting the historical purchase order data of the store to an order prediction model corresponding to the store attribute set to which the historical purchase order data belongs.

Further, the generating three-dimensional attribute coordinates of the store according to the historical purchase order data of the store comprises the following steps: setting a commodity classification table to divide commodities into quick-elimination classes, living classes and stationery classes; classifying commodities in the historical purchase order data of one store into classifications of the commodity classification table according to the commodity classification table; calculating the total classification price of the commodities in the classifications of each commodity classification table of the shops; and establishing a coordinate system of three-dimensional attribute coordinates by taking three classifications of the commodity classification table as coordinate axes respectively, and taking the total classification price of the store in the three classifications as coordinate values.

Further, the input data of the order prediction model is historical order data, and the output data of the order prediction model is store order prediction data and corresponding confidence.

Further, the order prediction model is a BP neural network model.

Further, the order prediction model is a logistic regression analysis model.

Further, the store commodity analysis method based on artificial intelligence further comprises the following steps: judging whether the confidence coefficient of the store order forecast data output by the order forecast model is larger than a preset threshold value, and if so, adopting the store order forecast data as analysis data.

Further, if the confidence coefficient of the store order forecast data is smaller than or equal to a preset threshold value, the historical purchase order data of the store is returned to be input into the corresponding order forecast model.

Further, the store commodity analysis method based on artificial intelligence further comprises the following steps: judging whether the confidence coefficient of the store order forecast data is smaller than or equal to a preset threshold value times and larger than or equal to 3, and if so, taking the store order data with the largest confidence coefficient output by the current order forecast model as analysis data.

As another aspect of the present application, the present application also provides an artificial intelligence-based store commodity analysis apparatus, comprising: a memory for storing a computer program; and the processor is used for realizing the store commodity analysis method based on artificial intelligence when executing the computer program.

As another aspect of the present application, there is also provided a computer client storage medium having stored therein a computer program which, when executed by a processor, implements the artificial intelligence based store commodity analysis method as described above.

The application has the advantages that: provided are a store commodity analysis method, device and storage medium based on artificial intelligence, wherein the store commodity analysis method, device and storage medium are used for comprehensively predicting future commodity order conditions of stores according to historical data of the stores and types of the stores.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application, are incorporated in and constitute a part of this specification. The drawings and their description are illustrative of the application and are not to be construed as unduly limiting the application. In the drawings:

FIG. 1 is a schematic diagram of the steps of an automated order generation method according to one embodiment of the application;

FIG. 2 is a schematic diagram of an order prediction model according to one embodiment of the application;

FIG. 3 is a schematic diagram of a sales prediction model according to an embodiment of the present application;

FIG. 4 is a schematic diagram of a price prediction model according to one embodiment of the application;

FIG. 5 is a schematic diagram of steps of a store order trend prediction method according to one embodiment of the present application;

FIG. 6 is a rectangular schematic of observed characteristic data according to one embodiment of the application;

FIG. 7 is a schematic diagram after clustering three-dimensional coordinate sets by K-Means clustering;

FIG. 8 is a schematic diagram of a business turn prediction model according to one embodiment of the application;

FIG. 9 is a schematic diagram of steps of a store commodity analysis method according to one embodiment of the present application;

FIG. 10 is a schematic diagram of an order prediction model according to one embodiment of the application;

FIG. 11 is a schematic diagram of a matrix of input data for an order prediction model according to one embodiment of the application;

Fig. 12 is a schematic block diagram of an apparatus for carrying out an embodiment of the method of the present application.

Detailed Description

In order that those skilled in the art will better understand the present application, a technical solution in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in which it is apparent that the described embodiments are only some embodiments of the present application, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the present application without making any inventive effort, shall fall within the scope of the present application.

It should be noted that, without conflict, the embodiments of the present application and features of the embodiments may be combined with each other. The application will be described in detail below with reference to the drawings in connection with embodiments.

From an overall point of view, the application builds a system that uses the following methods, which essentially fulfil three functions:

1. According to the corresponding data of the store, a suggested purchase order is automatically provided for the store, so that the management cost of a store operator is reduced, and the system is convenient for efficiently configuring suppliers, storages and logistics resources.

2. Analysis is performed from a range of similar stores to analyze the trend of orders as a whole, and based thereon, the orders that may occur in the stores are analyzed.

3. The business model of the store is analyzed so that similar historical data of the store assists in analyzing future order trends for that type of store.

Specifically, referring to fig. 1 to 4, as a first aspect of the present application, an automated order generation method of the present application includes: collecting inventory data and sales data of a store; store state characteristic data are generated according to store inventory data and sales data; inputting store status feature data into an order prediction model to cause the order prediction model to output a suggested order data and corresponding suggested confidence levels; judging whether the suggested confidence coefficient is larger than a preset suggested confidence coefficient threshold value, and generating a purchase order at least according to the suggested order data if the suggested confidence coefficient is larger than the suggested confidence coefficient threshold value.

Specifically, the stock data and sales data of the store are collected as the stock data and sales data of the store on the same day.

In order to be able to timely provide suggestions of purchase orders for shops, as a preferred solution, the system triggers the execution of the automatic order generation method according to the application daily, i.e. the system runs the program daily according to user settings or system settings.

In view of convenience of business habits of a store, the trigger time may be set at 22 points, or as another preferred scheme, a trigger condition may be set, and when the trigger condition is satisfied, a program for implementing the automated order generation method is automatically executed. For example, when the store uploads inventory data and sales data after daily inventory, the program is automatically triggered; for another example, the store user uses the user terminal to trigger a program when a request for an automated order is made to the server (the user may click a recommendation button in the user terminal APP interface).

The inventory data and sales data may be managed by a program of the system platform, or may be provided by a system to develop a data port to interface with other cargo statistics programs owned by the store.

The stock data and sales data referred to herein refer to SKU codes and corresponding quantity data, respectively, of items in stock or sold on the same day.

According to a general analysis scheme, a machine learning module is generally trained by adopting historical inventory data and sales data respectively, then prediction of the inventory data and the sales data is carried out respectively, then the situation of possible picking is judged according to the prediction result, and then corresponding order suggestions and the like are generated according to the out-of-stock situation.

However, because of uncertainty in sales data, unless all commodities are included in the input data, the input data is huge and the invalid data is large, and finally the trained machine learning model cannot be converged.

In addition, although inventory data is relatively stable with respect to sales data, the machine learning model may not converge due to the problem of the inventory data having a large number of product categories.

For the above reasons, there are major technical barriers to machine learning model training and prediction using inventory data and sales data directly.

Based on the above, the technical scheme of the application adopts a new technical conception, specifically, as one scheme, after collecting inventory data and sales data; and acquiring commodity SKU data and sales quantity of the first five digits of sales quantity and the existing stock quantity corresponding to the first five digits of commodities from sales data (the current day) to form a five-row three-column matrix, wherein the matrix is used as store state characteristic data.

As an extension, the number of rows of the matrix may be set according to the size of the store scale, and for a store with a large transaction amount, the number of rows of the matrix may be ten or more rows, for example, for a store with a large transaction amount, which has a large number of goods sold daily.

As a preferred solution, the total number of commodity categories in the store inventory is S, and the header value m=s×k is calculated, where K is the header percentage, which is determined according to the numerical interval in which S is located.

When the S is more than or equal to 300 and is more than 0, the value range of the head percentage is 7%; when the value of the head percentage is more than or equal to 1000 and is more than or equal to 300, the value range of the head percentage is 15 percent; when the value of the head percentage is more than or equal to 300 and is more than 0, the value range of the head percentage is 20 percent. The rectangular number of rows N is equal to the rounded value of the header value M.

The training order prediction model may employ two schemes, with inputs being the store state characteristic data described above. The difference is the source of the output data in the training set.

The first way is: the method comprises the steps of processing historical data of stores, taking actual historical purchase orders as output data, specifically outputting a matrix formed by SKU data and purchase quantity of commodities in the purchase orders, determining two columns, specifically, how many rows are generated according to actual conditions of the purchase orders, and taking store state characteristic data of a day before the actual historical purchase orders occur as input data, so that a group of training data is formed.

With such a scheme, the suggested order data output by the order prediction model is output as a matrix of SKU data and purchase quantity. The scheme has the advantages that the method can directly generate the purchase order, but due to the uncertainty of an output matrix, the order prediction model is equivalent to an empirical model, the sequence of inputting a training set and the setting of model parameters during training have great influence on final output, the accuracy fluctuation is high, the output has reference significance only by setting a high confidence threshold, and thus the problem of program circulation is caused during running. In this way, as a preferred scheme, the order prediction model may be a CNN neural network model.

The second way is: the order prediction model is built into a prediction type machine learning model, namely, the shop state characteristic data before the current day is used as input data, the shop state characteristic data on the current day is used as output data, and model training is carried out on the order prediction model, so that the output data and the output data are regular data and a determined matrix, and training and convergence are easy. In this way, as a preferred scheme, the order prediction model may be a BP neural network model.

Preferably, the order prediction model is constructed in a second mode, and in this case, store status feature data (real matrix) of tomorrow, which is described above, is output, although not as direct purchase order data, and SKU data, sales quantity, and inventory quantity of the first-ranked commodity, which may occur on the second day, are predicted in the matrix.

The system traverses sales quantity and inventory quantity data of all commodities in the prediction matrix and judges whether the sales quantity and inventory quantity data meet a preset relative relation or not. Specifically, the correlation is specifically sales number > inventory number x balance coefficient; wherein the balance coefficient is in the range of 0.27 to 0.7. Preferably, the balance coefficient takes a value of 0.5, and when the correlation judges that the sales number is more than 50% of the stock number, the commodity is selected into the purchase order.

Preferably, an order prediction model is trained for each store, i.e. one order prediction model corresponds to one of the stores. The data for each store is input into a corresponding order prediction model.

As a preferred aspect, the automated order generation method based on artificial intelligence further includes: if the suggested confidence level is less than or equal to the suggested confidence level threshold, returning to the step of collecting inventory data and sales data, namely processing by using the order prediction model again, and outputting new output data and confidence level.

As a further solution, if the confidence level of the suggestion cannot meet the confidence threshold value of the suggestion all the time, for example, the confidence level cannot be met after exceeding the preset times (the preset times may be 0 of course, that is, the next step is directly performed, and the order prediction model is not returned to perform data processing), the store state feature data of the current day is used as the predicted store state feature data, that is, the purchase order is directly generated by the data of the current day.

The generated purchase order can be sent to store users for reference, and if the store users determine the purchase order, automatic ordering can be realized, and of course, under the condition of authorization of the store users, unmanned management can be realized by the system directly according to the purchase order.

Preferably, considering the price floating condition of the e-commerce platform, generating the purchase order according to the suggested order data comprises the following steps: generating a purchased commodity list according to the suggested order data; acquiring the instant purchase price of the commodity in the purchased commodity list; acquiring future predicted prices of commodities in a purchased commodity list; calculating total price difference values of instant purchase total prices and future forecast total prices of all commodities in the purchased commodity list; and judging whether the total price difference is larger than a difference threshold value, and if the total price difference is smaller than the difference threshold value, generating the total price of the purchase order at the instant purchase price. If the total price difference is greater than the difference threshold, prompting the store user of the difference between the present order and the tomorrow order (total price difference) and suggesting the commodity category of the order data which may be out of stock (through the difference screening of the stock quantity and the sales quantity in the prediction matrix), and selecting the present order or the tomorrow automatic order by the user.

Alternatively, if the total price difference is greater than the difference threshold, inputting sales data to the sales prediction model to cause the sales prediction model to output a future predicted sales; calculating to obtain future prediction inventory according to the current inventory data and the future prediction sales; if the future predicted inventory is less than or equal to 0, the current instant purchase price generates a total price for the purchase order, and if the future predicted inventory is greater than 0, the total price for the purchase order is generated at the future predicted price. When the total price of the purchase order is generated at the future predicted price, the system places the order for the store user the next day or prompts the store user to place the order the next day.

The sales prediction model is adopted independently, and has the advantage that sales prediction of single commodities is influenced by other variables compared with the order prediction model in a plurality of feature dimensions.

Whereas machine learning models for individual commodity SKUs trained with historical sales data for individual commodities are more regular in data input and input, which is easier to converge. Preferably, the sales prediction model is a BP neural network model, the input data of which is the historical sales data of a commodity before the current day, and the output data of which is the predicted sales (i.e. future predicted sales) of the commodity on the next day.

This is mainly used to compare the cost variation of the current day order with the next day order, since in most cases the generation of the purchase order is based on future (next day) predictions, avoiding immediate losses to be incurred for future demand.

As a further preferable scheme, the future price prediction can be performed by the following method, and the instant purchase price is input into a price prediction model; obtaining a predicted price and price confidence coefficient output by a price prediction model; if the price confidence is greater than the price confidence threshold, outputting the predicted price as a future predicted price; and if the price confidence coefficient is smaller than or equal to the price confidence coefficient threshold value, outputting the average price in the specified period as a future predicted price.

Similar to sales prediction models, machine learning models for individual goods are easier to train. The price prediction model is also a BP neural network model, the input data of the price prediction model is the historical price data of a commodity before the current day, and the output data of the price prediction model is the predicted price of the commodity in the next day. I.e., one price prediction neural network corresponds to one commodity SKU.

The price confidence threshold may actually be set to a smaller value because the price of the commodity does not vary significantly due to cost considerations and the like. In addition, commodity price herein refers to the "collage" price of the e-commerce platform, rather than the vendor price or retail price of the store.

Further alternatively, the specified period may be one month or one week when the price confidence is equal to or less than the price confidence threshold.

Through the scheme, automatic order generation and management can be realized from the angle of a store, so that a background data base which is convenient for store users to realize functions such as automatic ordering, price comparison and the like through a store terminal is provided.

As another aspect of the present application, in order to achieve analysis of store order trends within a business district, as shown in fig. 5, the store order trend prediction method of the present application includes the steps of: dividing stores into a plurality of associated business district sets according to preset rules; generating observation feature data according to all historical orders of all shops in the associated business district set in a set observation period; inputting the observation characteristic data into a business district prediction model so as to enable the business district prediction model to output business district order prediction data and corresponding business district confidence coefficient; judging whether the business turn confidence coefficient is larger than a preset business turn confidence coefficient threshold value, and if so, acquiring store order prediction data of one store in the associated business turn set at least according to the business turn order prediction data.

Wherein the set of associated business circles is essentially a set of stores, which may be expressed as IDs of the stores from a data embodiment perspective; of course, the name of the store may be expressed, but the text itself is not suitable for data processing, and preferably is also a unique store ID code of the store, or a unique account ID of the store user may be employed, and the functions of both are the same.

As a specific scheme, the method for dividing the associated business district sets is adopted, and comprises the following steps: establishing a two-dimensional coordinate system according to the geographic position; acquiring coordinate values of a shop in a two-dimensional coordinate system; K-Means clustering operation is carried out by using coordinate values of shops in a two-dimensional coordinate system; and dividing the associated business circle set according to the result of the K-Means clustering operation.

By such a method, a result of clustering stores in a geographical location can be obtained, but the commercial properties of the stores cannot be reflected only from the geographical location. Clustering cannot be achieved commercially.

As a preferred scheme based on the above, the stores can be subjected to three-dimensional clustering operation by adding one business attribute dimension, so that the business association between stores can be reflected by the association business circle set.

As a specific scheme, the method comprises the following steps: acquiring historical order data of a store; calculating the average order value of the store according to the historical order data of the store, wherein the average order value is the average value of the order values of all purchase orders of the store in an observation period; establishing a three-dimensional coordinate system by taking the average order value as a third dimension; acquiring coordinate values of a shop in a three-dimensional coordinate system; K-Means clustering operation is carried out by coordinate values of shops in a three-dimensional coordinate system; and dividing the associated business circle set according to the K-Means clustering operation result.

The scheme is ingenious in that the size of the store and the purchase intention of the store on the e-commerce platform and the viscosity of the client can be simultaneously reflected by one datum, namely the average value of the purchase order. This is not possessed by other item data as a dimension, and the dimension can better divide the association business district sets according to the actual application effect.

Preferably, the observation period is quarterly or annual. Such a longer period of time may be a relatively stable characteristic of the reaction store.

FIG. 7 shows a three-dimensional clustering result, similar to the state represented by the changed graph, in which the above three-dimensional clustering operation is performed to divide the associated business district sets, so that business districts formed by stores that affect each other are more accurately reflected.

In this way, the plurality of business turn prediction models trained according to the division results of the associated business turn sets can more accurately output required analysis data.

Specifically, the store order trend prediction method further includes the following steps: all historical orders of all shops in the associated business district set in a set observation period are used as training set data.

The observation period may be set to one week according to the data analysis requirements, such as week analysis or day analysis, as a preferable scheme.

As shown in fig. 6, assuming that there are 6 stores in a certain related business district set, the observation period is one week, historical order data of one week is collected at this time, and the following data are acquired for each store ID: order data, total amount of order (sum of all order amounts), quantity SKU, and value SKU. Then, they are formed into a matrix as input data of a business district prediction model in the manner of fig. 6, that is, the observation feature data is a matrix expressing stores, commodities and their correspondence relations, wherein the correspondence relations are the correspondence relations of the attributes after the commodities are summarized.

More specifically, the observation feature data is constructed in a matrix including five columns of store IDs, order data, total amounts of orders, numbers SKUs, and value SKUs of the same store, respectively, and rows of the matrix corresponding to different stores in the associated business district set, respectively.

The number SKU is the SKU code of the commodity with the largest number after all orders are summarized, and the value SKU is the SKU code of the commodity with the highest summarized price after all orders are summarized.

At the same time, preferably, the output data of the business turn prediction model (business turn order prediction data) is also a same matrix, that is, the business turn order prediction data is also a matrix expressing stores, commodities and their corresponding relations, that is, the matrix described above. The information such as the possible order number, the total order amount and the like of a certain store in the next observation period can be obtained through the matrix.

The business circle prediction model can be a BP neural network model or a logistic regression analysis model.

By adopting the scheme of associating the business district sets and the business district prediction models, the shops with strong business association can be classified together for analysis, so that the order trend of the shops is analyzed and predicted from the perspective of the whole business district.

As another aspect of the present application, as shown in fig. 9, the present application also provides a store commodity analysis method, which includes the steps of: acquiring historical purchase order data of a store; generating three-dimensional attribute coordinates of the store according to historical purchase order data of the store; K-Means clustering operation is carried out by coordinate values of three-dimensional attribute coordinates of shops; dividing the K-Means clustering operation result into a plurality of store attribute sets; constructing order prediction models corresponding to different store attribute sets; the historical purchase order data of the store is input into an order prediction model corresponding to the attribute set of the store to which the store belongs.

Specifically, the three-dimensional attribute coordinate construction includes the steps of: setting a commodity classification table to divide commodities into quick-elimination classes, living classes and stationery classes; classifying commodities in the historical purchase order data of a store into classifications of commodity classification tables according to the commodity classification tables; calculating the total classified price of the commodities in the classification of each commodity classification table of the store; three classifications of the commodity classification table are respectively used as coordinate axes to establish a coordinate system of three-dimensional attribute coordinates, and the total price of the classifications of the shops is used as a coordinate value.

Preferably, the system can set the length of the time period for collecting the historical purchase orders, such as quarterly or annually, as needed for analysis. Preferably, if the acquisition time period length is set to be annual in order to obtain a more stable order prediction model and store attribute set.

When the time period length is annual, the above method is specifically: the annual purchase orders of a store are summarized, the commodities in the summary are classified into three classifications according to the three classifications in the commodity classification table, the classification total prices of all the commodities in the three classifications are respectively counted, the classification total price of the store under the three classifications is the coordinate value of the store in the three-dimensional attribute coordinates, and the counting unit of the classification total price is hundred yuan in view of the time period length, so that the coordinate value is not excessively large, and the coordinate points of a representative circuit are excessively scattered due to the calculating unit in clustering operation.

By the three-dimensional attribute coordinate establishment and clustering operation, stores can be divided into different store attribute sets. According to ideal state, designing conception per se according to quick-elimination type, living type and stationery type is to divide shops into corresponding business district type, district type and school type, wherein the quick-elimination type in the business district type shop purchase order is main commodity purchase type; the living type in the shop purchase order of the district is the main purchase commodity type; stationery in the store purchase order of the school type is the main purchase commodity type. Alternatively, the quick-vanishing class may include: beverages, snack foods, instant noodles, and the like; the living classes may include: seasoning, cleaning agent and articles for daily use; stationery may include: stationery, toys, etc.

And when the actual data are sorted and analyzed, the attributes of a plurality of shops are found to be complex, if the shops are only classified into business district types, community types and school types, the classified highest total price of classification can be adopted to belong to the classification, namely the classification, but the model training is more difficult due to the fact that the simple classification is found through later model construction and verification. For example, even if a store is located as a school, the quick-release purchasing amount is larger than the stationery purchasing amount. Simple classification does not bring practical value to later analysis and model construction.

By adopting the scheme, stores can be divided into store attribute sets according to actual conditions through dimension division and three-dimensional clustering, and the store attribute sets reflect actual attributes instead of manually dividing into classification attributes.

After the above division of the store attribute sets, the stores in the store attribute sets can be considered to have similar attribute characteristics, and based on this, an order prediction model is constructed for each store attribute set.

The order prediction model is a prediction model, the input of which is historical data, and the output of which is prediction data and corresponding confidence.

From a training perspective, all historical purchase orders for all stores in the store attribute set are consolidated into a matrix as shown in FIG. 11. Only one row of the matrix corresponds to a purchase order, which is divided into order amounts, quantity SKUs, and value SKUs. The historical purchase order of a store is used as corresponding training data to train the order prediction model.

The quantity SKU is the SKU value of the commodity with the largest quantity in the purchase order, and the value SKU is the SKU value of the commodity with the largest payment amount in the purchase order.

As a preferred embodiment, an observation period may be set, and the order prediction model is trained without being limited by the observation period, and when the order prediction model is used, the input data may be historical data in the observation period, such as one week or one month.

The order prediction model may be a BP neural network model or a logistic regression analysis model.

Thus, the order prediction model belonging to a certain store attribute set can predict the amount of the next purchase order, the commodity with the largest purchase amount, and the commodity with the highest total purchase price of the store by inputting the history data of the certain store in the store attribute set.

Preferably, the method further comprises the steps of: judging whether the confidence coefficient of the store order forecast data is smaller than or equal to a preset threshold value times and larger than or equal to 3, and if so, taking the store order data with the largest confidence coefficient output by the current order forecast model as analysis data. Because of adopting a store attribute set classification mode, in general, no confidence problem exists, and the data with highest confidence output by the model can also be used as analysis data.

By the three methods, the recommended order data is obtained from the perspective of store management, the business circle order prediction data is obtained from the perspective of business circles, and the store order prediction data is obtained from the perspective of store management types.

The system compares and analyzes the suggested order data, the business turn order forecast data, and the store order forecast data to obtain further application functions and analysis functions.

As one of the preferable schemes, the suggested order data can be compared with the corresponding business circle order prediction data and store order prediction data after being generated, and the differences between the business circle order prediction data and the store order prediction data are analyzed, so that the accuracy of the suggested order data is determined.

As another scheme, the business district order prediction data and the store order prediction data are compared, so that deviation of the business district from the business direction is analyzed, and a suggestion is given to a user to make the business direction match the business district.

As another aspect of the present application, as shown in fig. 12, the present application also provides a server 100, i.e., a device executing a program, which includes a memory 101 and a processor 102. Wherein the memory 101 is for storing a computer program and the processor 102 is for implementing the steps of the method as provided above when the computer program is executed.

The application also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor performs the steps of the method as provided above.

The above description is only of the preferred embodiments of the present application and is not intended to limit the present application, but various modifications and variations can be made to the present application by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. The store commodity analysis method based on artificial intelligence is characterized in that: the store commodity analysis method based on artificial intelligence comprises the following steps:

acquiring historical purchase order data of a store;

Generating three-dimensional attribute coordinates of the store according to historical purchase order data of the store;

Carrying out K-Means clustering operation by using coordinate values of the three-dimensional attribute coordinates of the store;

dividing the K-Means clustering operation result into a plurality of store attribute sets;

constructing order prediction models corresponding to different store attribute sets;

Inputting historical purchase order data of the store to an order prediction model corresponding to the store attribute set to which the historical purchase order data belongs;

The generating the three-dimensional attribute coordinates of the store according to the historical purchase order data of the store comprises the following steps:

setting a commodity classification table to divide commodities into quick-elimination classes, living classes and stationery classes;

classifying commodities in the historical purchase order data of one store into classifications of the commodity classification table according to the commodity classification table;

calculating the total classification price of the commodities in the classifications of each commodity classification table of the shops;

Establishing a coordinate system of three-dimensional attribute coordinates by taking three classifications of the commodity classification table as coordinate axes respectively, and taking the total classification price of the store in the three classifications as coordinate values;

the input data of the order prediction model is historical order data, and the output data of the order prediction model is store order prediction data and corresponding confidence;

the order prediction model is a BP neural network model;

The order prediction model is a logistic regression analysis model;

The store commodity analysis method based on artificial intelligence further comprises the following steps:

Judging whether the confidence coefficient of the store order forecast data output by the order forecast model is larger than a preset threshold value, and if so, adopting the store order forecast data as analysis data;

if the confidence coefficient of the store order forecast data is smaller than or equal to a preset threshold value, returning to input the historical purchase order data of the store to a corresponding order forecast model;

Judging whether the confidence coefficient of the store order forecast data is smaller than or equal to a preset threshold value times and larger than or equal to 3, and if so, taking the store order data with the largest confidence coefficient output by the current order forecast model as analysis data.