WO2023201510A1 - 销售预测方法、销售预测装置及销售预测模型的学习方法 - Google Patents

销售预测方法、销售预测装置及销售预测模型的学习方法 Download PDF

Info

Publication number
WO2023201510A1
WO2023201510A1 PCT/CN2022/087568 CN2022087568W WO2023201510A1 WO 2023201510 A1 WO2023201510 A1 WO 2023201510A1 CN 2022087568 W CN2022087568 W CN 2022087568W WO 2023201510 A1 WO2023201510 A1 WO 2023201510A1
Authority
WO
WIPO (PCT)
Prior art keywords
information
sales
product
vector
sequence information
Prior art date
Application number
PCT/CN2022/087568
Other languages
English (en)
French (fr)
Inventor
贾书军
刘征
刘晓涛
闫凤图
Original Assignee
烟台创迹软件有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 烟台创迹软件有限公司 filed Critical 烟台创迹软件有限公司
Priority to CN202280060322.8A priority Critical patent/CN117916758A/zh
Priority to PCT/CN2022/087568 priority patent/WO2023201510A1/zh
Publication of WO2023201510A1 publication Critical patent/WO2023201510A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis

Definitions

  • the present invention relates to a sales prediction method, a sales prediction device, and a learning method of a sales prediction model for predicting the sales status of merchandise in a store or the like.
  • Patent Document 1 proposes a shelf allocation support device that recognizes products from images of product shelves and generates displays that are effective for sales of the products based on analysis results based on the recognized products and sales data of the products. Multiple shelf allocation candidates for status.
  • Patent Document 1 Japanese Patent Application Publication No. 2021-108209
  • the present invention was made to solve the above-described problems, and an object thereof is to provide a sales prediction method, a sales prediction device, and a sales prediction model learning method that can improve the accuracy of product sales prediction.
  • the sales prediction method of the present invention includes: a step of obtaining a product information vector that expresses product information as a numerical vector; a step of obtaining status information of the product; a step of obtaining location information of the product; and generating a product information vector, status information, and location information.
  • the sales prediction device of the present invention includes: a status information acquisition unit that acquires status information of a product; a position information acquisition unit that acquires position information of a product; and a sequence information generation unit that generates a product information vector and status including information representing the product as a numerical vector. Sequence information of information and location information; and a sales forecasting department, which uses a sales forecast model to predict the sales quantity or sales of goods based on the sequence information.
  • the learning method of the sales prediction model of the present invention includes the steps of: generating sequence information including a product information vector representing product information as a numerical vector, product status information, and product position information; and using the sales prediction model to predict the product based on the sequence information. the steps of predicting the sales quantity or sales volume; and the steps of evaluating the predicted sales quantity or sales volume.
  • sequence information including product information vectors, status information, and location information as input data of the sales prediction model, sales predictions that also take into account the relationship between a plurality of pieces of information can be performed, and predictions can be improved. Accuracy.
  • FIG. 1 is a schematic structural diagram of the sales prediction device according to Embodiment 1.
  • FIG. 2 is a control block diagram of the sales prediction device according to Embodiment 1.
  • FIG. 2 is a control block diagram of the sales prediction device according to Embodiment 1.
  • FIG. 3 is a flowchart showing a learning method of the sales prediction model according to Embodiment 1.
  • FIG. 4 is a flowchart showing the flow of sales prediction processing according to Embodiment 1.
  • FIG. 4 is a flowchart showing the flow of sales prediction processing according to Embodiment 1.
  • Figure 5 is an image of the shelf at time t0.
  • Figure 6 is an image of the shelf at time t1.
  • FIG. 7 is a control block diagram of the sales prediction device according to Embodiment 3.
  • FIG. 8 is a flowchart showing the optimal position estimation method according to Embodiment 3.
  • FIG. 8 is a flowchart showing the optimal position estimation method according to Embodiment 3.
  • 1A Processing device; 2: Photographing device; 3: Shelf; 11: Image acquisition unit; 12: Object recognition unit; 13: Status information acquisition unit; 14: Position information acquisition unit; 15: Storage unit; 16: Sequence Information generation unit; 17: Sales forecast unit; 18: Output unit; 19: Determination unit; 100, 100A: Sales forecast device; 151: Product vector DB; 152: Sales forecast model.
  • FIG. 1 is a schematic structural diagram of the sales prediction device 100 according to Embodiment 1.
  • the sales prediction device 100 of this embodiment is a device used in a retail store such as a supermarket, and is a device that performs sales prediction of the product P stored on the shelf 3 in the store.
  • the sales prediction device 100 includes a processing device 1 and a photographing device 2 .
  • the processing device 1 is a PC or a server on the cloud equipped with a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit) and a memory.
  • the imaging device 2 is a camera installed on the ceiling or wall of the store and captures an image of the shelf 3 .
  • the processing device 1 and the imaging device 2 are communicably connected by wire or wirelessly. The image captured by the camera 2 is sent to the processing device 1 .
  • FIG. 2 is a control block diagram of the sales prediction device 100 according to Embodiment 1.
  • the processing device 1 has an image acquisition unit 11 , an object recognition unit 12 , a state information acquisition unit 13 , a position information acquisition unit 14 , a sequence information generation unit 16 , and a sales forecast unit as functional units implemented by executing a program on the CPU or the GPU. unit 17 and output unit 18.
  • the processing device 1 has a storage unit 15 including a volatile or non-volatile memory such as RAM, ROM, or flash memory.
  • the image acquisition unit 11 acquires the image of the shelf 3 captured by the imaging device 2 and sends it to the object recognition unit 12 .
  • the image acquisition unit 11 acquires images of the shelf 3 at preset time intervals.
  • the object recognition unit 12 uses known learned object detection models and object recognition models such as SSD (Single Shot Multibox Detector) using deep learning to recognize multiple products included in the image of the shelf 3 P.
  • SSD Single Shot Multibox Detector
  • the status information acquisition unit 13 acquires the status information of the commodity P based on the recognition result of the commodity P.
  • the status information of the product P includes three types: the product P exists on the shelf 3, the product P does not exist, or the status of the product P has changed.
  • the change in the state of the product P means that the product P exists and the position or angle of the product P has changed.
  • the status information of the product P is any one of three types: the product P exists on the shelf 3 (no change), the product P exists (there is a change), or the product P does not exist.
  • the status information of the product P may be two types: the presence of the product P on the shelf 3 or the absence of the product P.
  • the status information acquired by the status information acquisition unit 13 is sent to the sequence information generation unit 16 .
  • the position information acquisition unit 14 acquires the position information of the product P based on the recognition result of the product P.
  • the position information is a numerical value indicating the position of the product P on the shelf 3 .
  • the position information is, for example, a numerical value (0, 1, 2, 3, 4...) that is divided into three shelves and assigned sequentially.
  • the position information acquired by the position information acquisition unit 14 is sent to the sequence information generation unit 16 .
  • the storage unit 15 stores a product vector database 151 (hereinafter referred to as "product vector DB 151") and a sales prediction model 152.
  • the storage unit 15 also stores programs such as object detection models executed by the processing device 1, various parameters used in the programs, a plurality of images acquired by the image acquisition unit 11, status information acquisition unit 13 and position information acquisition unit 14. status information and location information, etc.
  • the product vector DB 151 includes product information such as product names of a plurality of products P, and product information vectors corresponding to the product information.
  • the product information vector is obtained by converting the product information into a vector.
  • the generation of product information vectors in this embodiment will be described. In the following example, four types of product A, product B, product C, and product D are explained.
  • the product information (product name) of each product is converted into One-Hot (one-hot code) representation as follows. By changing to One-Hot representation, the length of data can be unified and processing in the processor can be facilitated.
  • the product vector DB 151 includes product information vectors of all products P that are targets of sales prediction in the store.
  • the product vector DB 151 may be constructed using the processing device 1 , or the product vector DB 151 constructed using an external device may be stored in the storage unit 15 of the processing device 1 .
  • the sales prediction model 152 is a learned model that predicts the sales quantity of the product P based on the sequence information of the product P input from the sequence information generation unit 16 .
  • a long short-term memory (LSTM) model used in natural language processing (NLP) is used.
  • NLP natural language processing
  • the sales prediction model 152 not only the LSTM model but also an RNN model such as ConvLSTM or GRU, a series transformer model, and various other natural language processing models can be used.
  • the sequence information generation unit 16 integrates the product information vector included in the product vector DB 151, the status information acquired by the status information acquisition unit 13, and the position information acquired by the position information acquisition unit 14, to generate sequence information. The following explains the generation of sequence information.
  • the sequence information generation unit 16 converts the status information acquired by the status information acquisition unit 13 into a One-Hot representation as follows.
  • sequence information generation unit 16 converts the One-Hot representation of each product as a word into a numerical vector using Word2Vec.
  • Word2Vec The following is an example of a transformed state information vector.
  • sequence information generation unit 16 adds the status information vector of the corresponding product P to the product information vector included in the product vector DB 151 .
  • a status information vector is added to the product information vector as follows.
  • sequence information generation unit 16 adds the position information of each product to the product information vector and the status information vector to generate sequence information.
  • the sequence information generation unit 16 converts the position information into a value PE calculated using the following equations (1) and (2). pos in formulas (1) and (2) is the position, i is the dimension, and dmodel is the number of dimensions.
  • PE(pos,2i) sin(pos/10000 2i/dmodel )...(1)
  • PE(pos, 2i+1) cos(pos/10000 2i/dmodel )...(2)
  • the value PE indicating the position information of each of the products A to D is as follows.
  • PE(0,0) sin(0/10000 2 ⁇ 0/4 )
  • PE(0,1) cos(0/10000 2 ⁇ 0/4 )
  • PE(0,2) sin(0/10000 2 ⁇ 1/4 )
  • PE(0,3) cos(0/10000 2 ⁇ 1/4 )
  • PE(1,0) sin(1/10000 2 ⁇ 0/4 )
  • PE(1,1) cos(1/10000 2 ⁇ 0/4 )
  • PE(1,2) sin(1/10000 2 ⁇ 1/4 )
  • PE(1,3) cos(1/10000 2 ⁇ 1/4 )
  • PE(2,0) sin(2/10000 2 ⁇ 0/4 )
  • PE(2,1) cos(2/10000 2 ⁇ 0/4 )
  • PE(2,2) sin(2/10000 2 ⁇ 1/4 )
  • PE(2,3) cos(2/10000 2 ⁇ 1/4 )
  • PE(3,0) sin(3/10000 2 ⁇ 0/4 )
  • PE(3,1) cos(3/10000 2 ⁇ 0/4 )
  • PE(3,2) sin(3/10000 2 ⁇ 1/4 )
  • the position information By expressing the position information in the above method, the length of the sequence information can be ensured, and the sequence information can be used as a position code unique to various positions. In addition, the relationship between two positions can be modeled through affine transformation between these position information. In addition, the above description of the location information is an example, and the location information can also be expressed in other ways.
  • the sequence information generation unit 16 adds the value PE indicating the position information to the product information vector and the status information vector to generate sequence information.
  • the location information may be appended to the numerical values of the product information vector and the status information vector respectively, or may be added after the product information vector and the status information vector.
  • the sequence information of each of the products A to D when position information is added to the numerical values of the product information vector and the status information vector is as follows.
  • the sequence information of each of the products A to D when position information is added to the product information vector and the status information vector is as follows.
  • A [0.56, 0.44, 0.20, PE(0,0), PE(0,1), PE(0,2), PE(0,3)]
  • the sales prediction unit 17 uses the sequence information generated by the sequence information generation unit 16 and the sales prediction model 152 stored in the storage unit 15 to predict the sales of the product P on the shelf 3 .
  • the sequence information generated by the sequence information generation unit 16 is input to the sales prediction model 152 as one sentence including a plurality of words. That is, the sequence information is a sentence (language) indicating the state of the shelf 3 .
  • the sales prediction model 152 processes the input sequence information into natural language, and outputs a sales prediction as a result of translation, for example. Since the sales quantity varies depending on the product, the sales forecasting unit 17 of this embodiment outputs a value obtained by normalizing the sales quantity.
  • the output unit 18 outputs the sales forecast output from the sales forecast unit 17 to store managers and the like.
  • the output unit 18 may display the sales forecast on the display unit provided in the processing device 1 or may transmit the sales forecast to an external device.
  • the external device is, for example, a management device of a store in which the sales prediction device 100 is installed, or an information communication terminal such as a smartphone or tablet owned by the manager.
  • FIG. 3 is a flowchart showing a learning method of the sales prediction model 152 according to Embodiment 1.
  • the sales forecast model 152 may be learned by the processing device 1 of the sales forecast device 100 , or may be learned by an external computer and then stored in the storage unit 15 of the processing device 1 .
  • description will be given taking the case where learning is performed using the processing device 1 as an example.
  • the sequence information generation unit 16 acquires product information vectors of a plurality of products to be learned from the product vector DB 151 (S1). Then, the image of the shelf 3 including a plurality of products is captured by the imaging device 2 and acquired by the image acquisition unit 11 of the processing device 1 (S2). Next, the object recognition unit 12 detects and recognizes the merchandise included in the acquired image, and the state information acquisition unit 13 acquires the state information of each merchandise (S3). In addition, the location information acquisition unit 14 acquires the location information of each product (S4). The status information acquired by the status information acquisition unit 13 and the position information acquired by the position information acquisition unit 14 are sent to the sequence information generation unit 16 .
  • the sequence information generation unit 16 converts the status information received from the status information acquisition unit 13 into a status information vector (S5). Then, the sequence information generation unit 16 adds the state information vector and the position information to the product information vector of the product acquired from the product vector DB 151, and generates sequence information for each product (S6). The sequence information generated by the sequence information generation unit 16 is input to the sales prediction model 152 .
  • the sales prediction unit 17 uses the sales prediction model 152 to predict the sales quantity of each product on the shelf 3 (S7). Then, the sales forecasting unit 17 uses the Loss function to evaluate the predicted sales quantity.
  • the sales prediction unit 17 uses the actual sales quantity of the product in the same time period as the time period in which the image was acquired as training data, and calculates the Loss function (S8).
  • the training data is not limited to the sales quantity in the same time period, but may be the sales quantity in the time period when the display is judged to be the same as the time period when the image was acquired.
  • the Loss function use the least squares method or other methods. Then, the sales forecasting unit 17 determines whether the loss (Loss) calculated by the Loss function is greater than the threshold value (S9).
  • the sales prediction unit 17 adjusts the parameters of the sales prediction model 152 so as to reduce the loss (S10). After that, the process returns to step S2 and the subsequent processing is repeated. On the other hand, when the loss is below the threshold (S9: “Yes"), learning of the sales prediction model 152 is completed.
  • the sales prediction model 152 can output the predicted sales quantity of the product.
  • FIG. 4 is a flowchart showing the flow of sales prediction processing according to Embodiment 1.
  • FIG. The method is executed by the processing device 1 .
  • an image of the shelf 3 including a plurality of products is captured by the imaging device 2 and acquired by the image acquisition unit 11 of the processing device 1 (S21).
  • the object recognition unit 12 detects and recognizes the merchandise included in the acquired image
  • the state information acquisition unit 13 acquires the state information of each merchandise (S22).
  • the location information acquisition unit 14 acquires the location information of each product (S23).
  • the status information acquired by the status information acquisition unit 13 and the position information acquired by the position information acquisition unit 14 are sent to the sequence information generation unit 16 .
  • the sequence information generation unit 16 converts the status information received from the status information acquisition unit 13 into a status information vector (S24). In addition, the sequence information generation unit 16 acquires the product information vector of the product included in the acquired image from the product vector DB 151 (S25). Then, the state information vector and the position information are added to the product information vector to generate sequence information for each product (S26). The sequence information generated by the sequence information generation unit 16 is input to the sales prediction unit 17 .
  • the sales prediction unit 17 inputs the sequence information as input data to the sales prediction model 152, and uses the sales prediction model 152 to predict the sales quantity of each product on the shelf 3 (S27). Output the sales quantity as a normalized value.
  • the predicted sales quantity output from the sales prediction unit 17 is sent to the output unit 18 and output (S28).
  • FIG. 5 is an image of the shelf 3 at time t0
  • FIG. 6 is an image of the shelf 3 at time t1.
  • Time t1 and time t0 are different times.
  • the sequence information generating unit 16 At time t0 when the image in FIG. 5 is acquired, the sequence information generating unit 16 generates the following sequence information.
  • D 0 is the sequence information of the product D on the left side of the lowest level of the shelf 3
  • D 1 is the sequence information of the product D on the right side of the bottom level of the shelf 3 .
  • A [0.56, 0.44, 0.20, PE(0,0), PE(0,1), PE(0,2), PE(0,3)]
  • D 0 [0.36, 0.64, 0.20, PE(3,0), PE(3,1), PE(3,2), PE(3,3)]
  • D 1 [0.36, 0.64, 0.20, PE(4,0), PE(4,1), PE(4,2), PE(4,3)]
  • the predicted sales quantity output from the sales prediction model 152 using the above sequence information as input is, for example, as follows.
  • the sequence information generating unit 16 At time t1 when the image in FIG. 6 is acquired, the sequence information generating unit 16 generates the following sequence information.
  • the state information vector of D 1 is the non-existent vector 0.41, which is different from the case of FIG. 5 .
  • A [0.56, 0.44, 0.20, PE(0,0), PE(0,1), PE(0,2), PE(0,3)]
  • D 0 [0.36, 0.64, 0.20, PE(3,0), PE(3,1), PE(3,2), PE(3,3)]
  • D 1 [0.36, 0.64, 0.41, PE(4,0), PE(4,1), PE(4,2), PE(4,3)]
  • the predicted sales quantity output from the sales prediction model 152 using the above sequence information as input is, for example, as follows.
  • real-time information on site is acquired through images, information used for sales prediction is effectively integrated to generate sequence information, and sales prediction is performed, thereby improving prediction accuracy. That is, in this embodiment, natural language processing is applied to product management, each product and status information are processed into independent words, and the status of each product on the shelf 3 is expressed in sentences, so that the corresponding knowledge can be used for processing to improve prediction accuracy.
  • Embodiment 2 will be described.
  • the sales prediction device 100 performs sales prediction based on product information, status information, and location information.
  • Embodiment 2 differs from Embodiment 1 in that attribute information of the product is used in addition to these information.
  • the structure of the sales prediction device 100 is the same as that of Embodiment 1.
  • the attribute information of the product is, for example, price information, sales unit, price reduction information, or sales period. An example of adding price information as attribute information will be described below.
  • the product vector DB 151 of this embodiment includes a price information vector in addition to the product information vector.
  • the price information vector is obtained by converting the price information of the product into a vector.
  • the generation of the price information vector is explained. In the following example, four types of product A, product B, product C, and product D are explained. First, the price information of each product is converted into One-Hot representation.
  • the One-Hot expression of the price information of each product is used as a word and converted into a numerical vector using Word2Vec.
  • the converted price information vector is appended to the product information vector as described below, and then stored in the product vector DB 151.
  • the sequence information generation unit 16 adds the status information vector of the corresponding product to the product information vector and the price information vector included in the product vector DB 151 . For example, when all products A to D exist, a status information vector is added to the product information vector as follows.
  • position information is added to generate sequence information for each product A to D.
  • the sequence information in this case becomes any of the following.
  • A [0.56+PE(0,0), 0.44+PE(0,1), 0.05+PE(0,2), 0.20+PE(0,3)]
  • A [0.56, 0.44, 0.05, 0.20, PE(0,0), PE(0,1), PE(0,2), PE(0,3)]
  • the sequence information generated by the sequence information generation unit 16 is input to the sales forecast unit 17, and the sales forecast unit 17 uses the sales forecast model 152 to predict and output the sales quantity.
  • Embodiment 3 will be described.
  • the processing device 1A of Embodiment 3 uses the sales prediction model 152 to estimate and recommend the optimal product placement.
  • FIG. 7 is a control block diagram of the sales prediction device 100A according to the third embodiment.
  • the processing device 1A is different from Embodiment 1 in that it also has a determination unit 19 as a functional unit realized by executing a program on the CPU or the GPU.
  • the other structures of the sales prediction device 100A are the same as those in Embodiment 1.
  • the determination unit 19 determines whether the sales quantity predicted by the sales prediction unit 17 reaches a predetermined quantity. Then, when the predicted sales quantity does not reach the predetermined quantity, the determination unit 19 instructs the sequence information generation unit 16 to change the position information written in the sequence information. On the other hand, when the predicted sales quantity reaches the predetermined quantity, the determination unit 19 outputs the position information at that time to the output unit 18 as the optimal position information. The output unit 18 outputs the optimal position information to the manager of the store and the like.
  • the sequence information generation unit 16 of this embodiment receives an instruction to change the position information from the determination unit 19 , the sequence information generation unit 16 sequentially changes the position information to position information corresponding to a conceivable arrangement in the shelf 3 .
  • the sequence information generation unit 16 changes the position information so as to correspond to the arrangement in which the positions of item A and item B are exchanged, and generates a written The sequence information of the changed position information.
  • the sequence information generation unit 16 sequentially changes the position information based on the instruction from the determination unit 19 and generates sequence information.
  • FIG. 8 is a flowchart showing the optimal position estimation method according to Embodiment 3.
  • FIG. The method is carried out by the processing device 1 .
  • Steps S21 to S27 of the optimal position estimation method are the same as the sales prediction method of Embodiment 1.
  • the determination unit 19 determines whether the predicted sales quantity is more than a predetermined quantity (S30).
  • the sequence information generation unit 16 changes the position information (S31) and generates sequence information (S26). Then, the sales forecasting unit 17 predicts the sales quantity under the changed location information (S27). Then, when the sales quantity predicted by the sales prediction unit 17 becomes more than a predetermined quantity (S30: Yes), the determination unit 19 outputs the current position information as the optimal position information via the output unit 18.
  • the output position information is information indicating the shelf allocation table of the product.
  • the shelf allocation table includes not only the arrangement of the goods on the shelf 3, but also the arrangement in the refrigerator or the arrangement on the plane.
  • the optimal position information can be estimated by repeatedly changing the position information until the sales quantity predicted by the sales prediction unit 17 becomes a predetermined number or more.
  • the sales volume can be increased by having the store manager or the like adjust the shelf allocation of the products based on the estimated location information.
  • the sales prediction device 100 in the above-described embodiment predicts the sales quantity of each product, but the sales volume of each product may also be predicted.
  • the sales prediction model 152 learns the actual sales volume as training data, and outputs the predicted sales volume after inputting the sequence information.
  • the sales forecasting unit 17 may output the sales quantity or sales volume as it is without standardization.
  • the state information when generating sequence information, the state information is converted into a numerical vector, but the present invention is not limited to this.
  • the numerical value may be included in the sequence information as it is without converting the information represented by numerical values into vectors.
  • the status information and position information of the product when the sales prediction model 152 is learned is not limited to the information obtained from the image.
  • the sequence information may be created using the status information and position information included in the shelf allocation table of the merchandise stored in advance in the storage unit 15 .

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

销售预测方法具备：取得用数值向量表现商品的信息的商品信息向量的步骤；取得商品的状态信息的步骤；取得商品的位置信息的步骤；生成包括商品信息向量、状态信息及位置信息的序列信息的步骤；以及使用销售预测模型，根据序列信息预测商品的销售数量或者销售额的步骤。

Description

销售预测方法、销售预测装置及销售预测模型的学习方法 技术领域
本发明涉及预测店铺等中的商品的销售状况的销售预测方法、销售预测装置及销售预测模型的学习方法。
背景技术
在超级市场等零售店中，商品的配置影响销售额，所以已知事先进行销售额或者销售数量等的销售预测来变更配置。例如，在专利文献1中，提出一种货架分配支援装置，从商品货架的图像辨识商品，根据基于辨识出的商品和该商品的销售数据的解析结果，生成表示对商品的销售额有效的陈列状态的多个货架分配候补。
现有技术文献
专利文献
专利文献1：日本特开2021-108209号公报
发明内容
在专利文献1的货架分配支援装置等中的以往的销售预测中，个别地考虑预测所需的商品信息、现场的状况以及位置信息等各种信息来进行预测，所以有未考虑各信息之间的关系而预测精度不充分的情况。
本发明是为了解决如上述那样的课题而完成的，其目的在于提供一种能够提高商品的销售预测的精度的销售预测方法、销售预测装置及销售预测模型的学习方法。
本发明的销售预测方法具备：取得用数值向量表现商品的信息的商品信息向量的步骤；取得商品的状态信息的步骤；取得商品的位置信息的步骤；生成包括商品信息向量、状态信息及位置信息的序列信息的步骤；以及使用销售预测模型，根据序列信息预测商品的销售数量或者销售额的步骤。
本发明的销售预测装置具备：状态信息取得部，取得商品的状态信息；位置信息取得部，取得商品的位置信息；序列信息生成部，生成包括用数值向量表现商品的信息的商品信息向量、状态信息及位置信息的序列信息；以及销售预测部，使用销售预测模型，根据序列信息预测商品的销售数量或者销售额。
本发明的销售预测模型的学习方法具备：生成包括用数值向量表现商品的信息 的商品信息向量、商品的状态信息及商品的位置信息的序列信息的步骤；使用销售预测模型，根据序列信息预测商品的销售数量或者销售额的步骤；以及评价预测出的销售数量或者销售额的步骤。
根据本发明的方法以及装置，通过将包括商品信息向量、状态信息以及位置信息的序列信息作为销售预测模型的输入数据，能够进行还考虑了多个信息之间的关系的销售预测，能够提高预测精度。
附图说明
图1是实施方式1所涉及的销售预测装置的概略结构图。
图2是实施方式1所涉及的销售预测装置的控制框图。
图3是示出实施方式1所涉及的销售预测模型的学习方法的流程图。
图4是示出实施方式1所涉及的销售预测处理的流程的流程图。
图5是时间t0时的货架的图像。
图6是时间t1时的货架的图像。
图7是实施方式3所涉及的销售预测装置的控制框图。
图8是示出实施方式3所涉及的最佳位置推测方法的流程图。
(符号说明)
1、1A：处理装置；2：拍摄装置；3：货架；11：图像取得部；12：物体识别部；13：状态信息取得部；14：位置信息取得部；15：存储部；16：序列信息生成部；17：销售预测部；18：输出部；19：判定部；100、100A：销售预测装置；151：商品向量DB；152：销售预测模型。
具体实施方式
以下，参照附图，说明实施本发明的实施方式的销售预测方法的销售预测装置100。此外，在各图中，对同一或者相当的部分附加同一符号，适宜地省略或者简化其说明。另外，关于各图记载的结构，其形状、大小以及配置等能够在本发明的范围内适宜地变更。
实施方式1.
图1是实施方式1所涉及的销售预测装置100的概略结构图。本实施方式的销售预测装置100是在超级市场等零售店铺中使用的装置，且是进行收容于店铺内的货 架3的商品P的销售预测的装置。销售预测装置100包括处理装置1和拍摄装置2。处理装置1是具备CPU(Central Processing Unit，中央处理单元)或者GPU(Graphics Processing Unit，图形处理单元)以及存储器的PC或者云上的服务器等。拍摄装置2是设置于店铺的顶棚或者墙壁并拍摄货架3的图像的照相机。处理装置1和拍摄装置2以有线或者无线方式可通信地连接。由拍摄装置2拍摄的图像被发送给处理装置1。
图2是实施方式1所涉及的销售预测装置100的控制框图。在处理装置1中，作为通过CPU或者GPU执行程序而实现的功能部，具有图像取得部11、物体识别部12、状态信息取得部13、位置信息取得部14、序列信息生成部16、销售预测部17以及输出部18。另外，处理装置1具有包括RAM、ROM或者闪存存储器等易失性或者非易失性的存储器的存储部15。
图像取得部11取得由拍摄装置2拍摄的货架3的图像，发送给物体识别部12。图像取得部11以预先设定的时间间隔取得货架3的图像。
物体识别部12利用使用了深度学习的SSD(Single Shot Multibox Detector，单发多框检测器)等已知的已学习的物体探测模型以及物体识别模型，识别包含于货架3的图像的多个商品P。
状态信息取得部13根据商品P的识别结果，取得商品P的状态信息。商品P的状态信息是在货架3上存在商品P、不存在商品P、或者商品P的状态发生了变化这3种。商品P的状态发生了变化是指，存在商品P、并且商品P的位置或者角度发生了变化。换言之，商品P的状态信息是在货架3上存在商品P(无变化)、存在商品P(有变化)或者不存在商品P这3种内的任意一种。此外，在其他实施方式中，商品P的状态信息也可以是在货架3上存在商品P、或者不存在商品P这2种。由状态信息取得部13取得的状态信息被发送给序列信息生成部16。
位置信息取得部14根据商品P的识别结果，取得商品P的位置信息。位置信息是表示货架3中的商品P的位置的数值。位置信息例如是将货架3分割而依次分配的数值(0，1，2，3，4…)。由位置信息取得部14取得的位置信息被发送给序列信息生成部16。
存储部15存储有商品向量数据库151(以下称为“商品向量DB151”)和销售预测模型152。存储部15还存储由处理装置1执行的物体探测模型等程序、在程序中使用的各种参数、由图像取得部11取得的多个图像、由状态信息取得部13以及位置信息取得部14取得的状态信息以及位置信息等。
商品向量DB151包括多个商品P的商品名等商品信息、和与商品信息对应的商品信息向量。商品信息向量是将商品信息变换为向量而得到的。以下，说明本实施方式中的商品信息向量的生成。在以下的例子中，说明商品A、商品B、商品C以及商品D这4种。首先，将各商品的商品信息(商品名)如下所述变换为One-Hot(独热码)表现。通过变更为One-Hot表现，能够使数据的长度统一，并且处理器中的处理也容易。
A：{1，0，0，0}
B：{0，1，0，0}
C：{0，0，1，0}
D：{0，0，0，1}
然后，将各商品的One-Hot表现作为单词，使用Word2Vec变换为数值向量。以下是变换后的商品信息向量的一个例子。在商品向量DB151中，包括在店铺中成为销售预测的对象的所有商品P的商品信息向量。商品向量DB151的构筑既可以用处理装置1进行，也可以将用外部设备构筑的商品向量DB151存储到处理装置1的存储部15。
A：{0.56，0.44}
B：{0.82，0.18}
C：{0.18，0.82}
D：{0.36，0.64}
销售预测模型152是根据从序列信息生成部16输入的商品P的序列信息来预测商品P的销售数量的已学习模型。作为销售预测模型152，使用了在自然语言处理(NLP)中使用的长短期记忆(LSTM)模型。此外，作为销售预测模型152，不仅能够使用LSTM模型，而且还能够使用ConvLSTM或者GRU等RNN模型、系列变换(Transformer)模型、以及其他各种自然语言处理模型。
序列信息生成部16将包含于商品向量DB151的商品信息向量、由状态信息取得部13取得的状态信息、以及由位置信息取得部14取得的位置信息进行整合，生成序列信息。以下说明序列信息的生成。
序列信息生成部16将由状态信息取得部13取得的状态信息如下所述变换为One-Hot表现。
有存在：{1，0，0}
无存在：{0，1，0}
有变化：{0，0，1}
然后，序列信息生成部16将各商品的One-Hot表现作为单词，使用Word2Vec变换为数值向量。以下是变换后的状态信息向量的一个例子。
有存在：{0.20}
无存在：{0.41}
有变化：{0.90}
而且，序列信息生成部16对包含于商品向量DB151的商品信息向量追加对应的商品P的状态信息向量。例如，在商品A～D全部存在的情况下，对商品信息向量如下所述追加状态信息向量。
A：{0.56，0.44，0.20}
B：{0.82，0.18，0.20}
C：{0.18，0.82，0.20}
D：{0.36，0.64，0.20}
而且，序列信息生成部16对商品信息向量以及状态信息向量追加各商品的位置信息而生成序列信息。在本实施方式中，序列信息生成部16将位置信息变换为使用下述的式(1)以及(2)求出的值PE。式(1)以及(2)中的pos是位置，i是维度，dmodel是维度的数量。
PE(pos，2i)＝sin(pos/10000 2i/dmodel)…(1)
PE(pos，2i+1)＝cos(pos/10000 2i/dmodel)…(2)
具体而言，表示各商品A～D的位置信息的值PE如下述所示。
A的位置(位置信息＝0)
PE(0，0)＝sin(0/10000 2×0/4)
PE(0，1)＝cos(0/10000 2×0/4)
PE(0，2)＝sin(0/10000 2×1/4)
PE(0，3)＝cos(0/10000 2×1/4)
B的位置(位置信息＝1)
PE(1，0)＝sin(1/10000 2×0/4)
PE(1，1)＝cos(1/10000 2×0/4)
PE(1，2)＝sin(1/10000 2×1/4)
PE(1，3)＝cos(1/10000 2×1/4)
C的位置(位置信息＝2)
PE(2，0)＝sin(2/10000 2×0/4)
PE(2，1)＝cos(2/10000 2×0/4)
PE(2，2)＝sin(2/10000 2×1/4)
PE(2，3)＝cos(2/10000 2×1/4)
D的位置(位置信息＝3)
PE(3，0)＝sin(3/10000 2×0/4)
PE(3，1)＝cos(3/10000 2×0/4)
PE(3，2)＝sin(3/10000 2×1/4)
PE(3，3)＝cos(3/10000 2×1/4)
通过用上述方法表现位置信息，能够确保序列信息的长度，能够作为各种位置所固有的位置编码。另外，能够通过这些位置信息之间的仿射变换，对2个位置之间的关系进行模型化。此外，上述位置信息的记载是一个例子，也可以用其他方法表现位置信息。
序列信息生成部16将表示位置信息的值PE追加到商品信息向量以及状态信息向量而生成序列信息。位置信息既可以分别追加到商品信息向量以及状态信息向量的数值，也可以在商品信息向量以及状态信息向量之后追加。
将位置信息分别追加到商品信息向量以及状态信息向量的数值的情况的各商品A～D的序列信息如下所述。
A＝[0.56+PE(0，0)，0.44+PE(0，1)，0.20+PE(0，2)，PE(0，3)]
B＝[0.82+PE(1，0)，0.18+PE(1，1)，0.20+PE(1，2)，PE(1，3)]
C＝[0.18+PE(2，0)，0.82+PE(2，1)，0.20+PE(2，2)，PE(2，3)]
D＝[0.36+PE(3，0)，0.64+PE(3，1)，0.20+PE(3，2)，PE(3，3)]
将位置信息追加到商品信息向量以及状态信息向量之后的情况的各商品A～D的序列信息如下所述。
A＝[0.56，0.44，0.20，PE(0，0)，PE(0，1)，PE(0，2)，PE(0，3)]
B＝[0.82，0.18，0.20，PE(1，0)，PE(1，1)，PE(1，2)，PE(1，3)]
C＝[0.18，0.82，0.20，PE(2，0)，PE(2，1)，PE(2，2)，PE(2，3)]
D＝[0.36，0.64，0.20，PE(3，0)，PE(3，1)，PE(3，2)，PE(3，3)]
销售预测部17使用由序列信息生成部16生成的序列信息、和存储于存储部15的销售预测模型152，进行货架3的商品P的销售预测。将由序列信息生成部16生 成的序列信息作为包括多个单词的一个句子，输入到销售预测模型152。即，序列信息是表示货架3的状态的句子(语言)。销售预测模型152将输入的序列信息处理为自然语言，例如作为翻译的结果输出销售预测。取决于不同商品，销售数量有差异，所以本实施方式的销售预测部17输出对销售数量进行标准化而得到的值。
输出部18针对店铺的管理者等输出从销售预测部17输出的销售预测。输出部18既可以将销售预测显示于处理装置1具备的显示部，也可以将销售预测发送给外部设备。外部设备是例如设置有销售预测装置100的店铺的管理装置、或者管理者所拥有的智能手机或者平板等信息通信终端。
图3是示出实施方式1所涉及的销售预测模型152的学习方法的流程图。销售预测模型152的学习既可以由销售预测装置100的处理装置1进行，也可以在通过外部的计算机进行学习之后，存储到处理装置1的存储部15。以下，以用处理装置1进行学习的情况为例子进行说明。
首先，序列信息生成部16从商品向量DB151取得成为学习对象的多个商品的商品信息向量(S1)。然后，包括多个商品的货架3的图像由拍摄装置2拍摄，并通过处理装置1的图像取得部11取得(S2)。接下来，通过物体识别部12探测以及识别包含于取得的图像的商品，通过状态信息取得部13取得各商品的状态信息(S3)。另外，通过位置信息取得部14取得各商品的位置信息(S4)。由状态信息取得部13取得的状态信息以及由位置信息取得部14取得的位置信息被发送给序列信息生成部16。
序列信息生成部16将从状态信息取得部13接收到的状态信息变换为状态信息向量(S5)。然后，序列信息生成部16对从商品向量DB151取得的商品的商品信息向量追加状态信息向量以及位置信息，生成每个商品的序列信息(S6)。由序列信息生成部16生成的序列信息被输入到销售预测模型152。
然后，销售预测部17使用销售预测模型152，预测货架3的各商品的销售数量(S7)。然后，销售预测部17使用Loss函数，进行预测出的销售数量的评价。在此，销售预测部17将与取得图像的时间段相同的时间段中的商品的实际的销售数量作为训练数据，计算Loss函数(S8)。此外，训练数据不限定于相同的时间段中的销售数量，是判断为陈列与取得图像的时间段等同的时间中的销售数量即可。另外，作为Loss函数，使用最小二乘法或者其他方法。然后，销售预测部17判断通过Loss函数的计算求出的损失(Loss)是否大于阈值(S9)。
在损失大于阈值的情况下(S9：“否”)、即销售预测模型152的学习不充分的情况下，销售预测部17以使损失变小的方式调整销售预测模型152的参数(S10)。之后，返回到步骤S2，重复进行以后的处理。另一方面，在损失成为阈值以下的情况下(S9：“是”)，结束销售预测模型152的学习。由此，销售预测模型152在被输入商品的序列信息作为输入数据时，能够输出该商品的预测的销售数量。
图4是示出实施方式1所涉及的销售预测处理的流程的流程图。本方法由处理装置1执行。首先，包括多个商品的货架3的图像由拍摄装置2拍摄，并通过处理装置1的图像取得部11取得(S21)。然后，通过物体识别部12探测以及识别包含于取得的图像的商品，通过状态信息取得部13取得各商品的状态信息(S22)。另外，通过位置信息取得部14取得各商品的位置信息(S23)。由状态信息取得部13取得的状态信息以及由位置信息取得部14取得的位置信息被发送给序列信息生成部16。
序列信息生成部16将从状态信息取得部13接收到的状态信息变换为状态信息向量(S24)。另外，序列信息生成部16从商品向量DB151取得包含于取得的图像的商品的商品信息向量(S25)。然后，对商品信息向量追加状态信息向量以及位置信息，生成每个商品的序列信息(S26)。由序列信息生成部16生成的序列信息被输入给销售预测部17。
然后，销售预测部17将序列信息作为输入数据输入给销售预测模型152，使用销售预测模型152预测货架3的各商品的销售数量(S27)。将销售数量输出为标准化后的值。从销售预测部17输出的预测销售数量被发送给输出部18而输出(S28)。
参照图5以及图6，说明具体例子。图5是时间t0时的货架3的图像，图6是时间t1时的货架3的图像。时间t1和时间t0是不同的时间。在取得图5的图像的时间t0，通过序列信息生成部16生成下述的序列信息。此外，D 0是货架3的最下层的左侧的商品D的序列信息，D 1是货架3的最下层的右侧的商品D的序列信息。
A＝[0.56，0.44，0.20，PE(0，0)，PE(0，1)，PE(0，2)，PE(0，3)]
B＝[0.82，0.18，0.20，PE(1，0)，PE(1，1)，PE(1，2)，PE(1，3)]
C＝[0.18，0.82，0.20，PE(2，0)，PE(2，1)，PE(2，2)，PE(2，3)]
D 0＝[0.36，0.64，0.20，PE(3，0)，PE(3，1)，PE(3，2)，PE(3，3)]
D 1＝[0.36，0.64，0.20，PE(4，0)，PE(4，1)，PE(4，2)，PE(4，3)]
将上述序列信息作为输入而从销售预测模型152输出的预测销售数量例如如下所述。
A：0.3
B：0.2
C：0.4
D：0.1
在取得图6的图像的时间t1，通过序列信息生成部16生成下述的序列信息。在下述的序列信息中，在D 1的状态信息向量为无存在的向量0.41这一点，与图5的情况不同。
A＝[0.56，0.44，0.20，PE(0，0)，PE(0，1)，PE(0，2)，PE(0，3)]
B＝[0.82，0.18，0.20，PE(1，0)，PE(1，1)，PE(1，2)，PE(1，3)]
C＝[0.18，0.82，0.20，PE(2，0)，PE(2，1)，PE(2，2)，PE(2，3)]
D 0＝[0.36，0.64，0.20，PE(3，0)，PE(3，1)，PE(3，2)，PE(3，3)]
D 1＝[0.36，0.64，0.41，PE(4，0)，PE(4，1)，PE(4，2)，PE(4，3)]
将上述序列信息作为输入而从销售预测模型152输出的预测销售数量例如如下所述。
A：0.4
B：0.3
C：0.3
D：0.0
如以上所述，在本实施方式中，经由图像取得现场的实时的信息，有效地整合在销售预测中使用的信息来生成序列信息，进行销售预测，从而能够提高预测精度。即，在本实施方式中，将自然语言处理应用于商品管理，将各商品以及状态信息等处理为独立的单词，用句子表现货架3中的各商品的状态，从而能够将对应的知识使用于处理，提高预测精度。
实施方式2.
说明实施方式2。在实施方式1中，销售预测装置100根据商品信息、状态信息以及位置信息进行销售预测，但在实施方式2中，除了这些信息以外还使用商品的属性信息，这一点与实施方式1不同。销售预测装置100的结构与实施方式1相同。商品的属性信息例如是价格信息、销售单位、降价信息或者销售时期等。以下，说明作为属性信息加上价格信息的情况的例子。
在本实施方式的商品向量DB151中，除了商品信息向量以外，还包括价格信息向量。价格信息向量是将商品的价格信息变换为向量而得到的。以下，说明价格信息向量的生成。在以下的例子中，说明商品A、商品B、商品C以及商品D这4种。首 先，将各商品的价格信息变换为One-Hot表现。
然后，将各商品的价格信息的One-Hot表现作为单词，使用Word2Vec变换为数值向量。变换后的价格信息向量如下所述被追加到商品信息向量之后，存储到商品向量DB151。
A：{0.56，0.44，0.05}
B：{0.82，0.18，0.10}
C：{0.18，0.82，0.82}
D：{0.36，0.64，0.45}
序列信息生成部16对包含于商品向量DB151的商品信息向量以及价格信息向量追加对应的商品的状态信息向量。例如，在商品A～D全部存在的情况下，对商品信息向量如下所述追加状态信息向量。
A：{0.56，0.44，0.05，0.20}
B：{0.82，0.18，0.10，0.20}
C：{0.18，0.82，0.82，0.20}
D：{0.36，0.64，0.45，0.20}
而且，还追加位置信息，生成各商品A～D的序列信息。该情况的序列信息成为下述中的任意一个。
A＝[0.56+PE(0，0)，0.44+PE(0，1)，0.05+PE(0，2)，0.20+PE(0，3)]
B＝[0.82+PE(1，0)，0.18+PE(1，1)，0.10+PE(1，2)，0.20+PE(1，3)]
C＝[0.18+PE(2，0)，0.82+PE(2，1)，0.82+PE(2，2)，0.20+PE(2，3)]
D＝[0.36+PE(3，0)，0.64+PE(3，1)，0.45+PE(3，2)，0.20+PE(3，3)]
或者
A＝[0.56，0.44，0.05，0.20，PE(0，0)，PE(0，1)，PE(0，2)，PE(0，3)]
B＝[0.82，0.18，0.10，0.20，PE(1，0)，PE(1，1)，PE(1，2)，PE(1，3)]
C＝[0.18，0.82，0.82，0.20，PE(2，0)，PE(2，1)，PE(2，2)，PE(2，3)]
D＝[0.36，0.64，0.45，0.20，PE(3，0)，PE(3，1)，PE(3，2)，PE(3，3)]
由序列信息生成部16生成的序列信息被输入给销售预测部17，通过销售预测部17使用销售预测模型152预测并输出销售数量。
考虑了对商品的销售额还关联商品的价格以及销售单位等其他属性信息。因此，在本实施方式中，通过将价格信息等属性信息追加作为销售预测模型152的输入数据，能够进一步提高预测结果的精度。
实施方式3.
说明实施方式3。实施方式3的处理装置1A使用销售预测模型152推测并推荐最佳的商品的配置。图7是实施方式3所涉及的销售预测装置100A的控制框图。在处理装置1A中，作为通过CPU或者GPU执行程序而实现的功能部，还具有判定部19，这一点与实施方式1不同。其他销售预测装置100A的结构与实施方式1相同。
判定部19判定由销售预测部17预测出的销售数量是否达到规定数量。然后，判定部19在预测销售数量未达到规定数量的情况下，指示序列信息生成部16变更写入到序列信息的位置信息。另一方面，在预测销售数量达到了规定数量的情况下，判定部19将此时的位置信息作为最佳的位置信息输出给输出部18。输出部18将最佳的位置信息输出给店铺的管理者等。
本实施方式的序列信息生成部16在从判定部19接收到位置信息的变更指示的情况下，将位置信息依次变更为与在货架3中可设想的配置对应的位置信息。例如，在当前商品的位置信息与图5所示的配置对应的情况下，序列信息生成部16以使得与将商品A和商品B的位置交换后的配置对应的方式变更位置信息，生成写入了变更后的位置信息的序列信息。序列信息生成部16根据来自判定部19的指示，依次变更位置信息，生成序列信息。
图8是示出实施方式3所涉及的最佳位置推测方法的流程图。本方法通过处理装置1执行。最佳位置推测方法的步骤S21～S27与实施方式1的销售预测方法相同。然后，在预测货架3的各商品的销售数量后(S27)，判定部19判断预测出的销售数量是否为规定数量以上(S30)。
然后，在预测出的销售数量小于规定数量的情况下(S30：“否”)，序列信息生成部16变更位置信息(S31)，生成序列信息(S26)。然后，通过销售预测部17预测变更后的位置信息下的销售数量(S27)。然后，在由销售预测部17预测出的销售数量成为规定数量以上的情况下(S30：“是”)，判定部19将当前的位置信息作为最佳的位置信息经由输出部18输出。输出的位置信息是表示商品的货架分配表的信息。此外，货架分配表不仅是货架3中的商品的配置，而且还包括电冰箱内的配置或者平面上的配置。
如以上所述，在本实施方式中，通过反复变更位置信息直至由销售预测部17预测的销售数量成为规定数量以上为止，能够推测最佳的位置信息。由此，通过由店铺的管理者等根据推测出的位置信息调整商品的货架分配，能够提高销售数量。
以上为实施方式的说明，但上述实施方式可变形以及组合。例如，上述实施方 式中的销售预测装置100预测各商品的销售数量，但也可以预测各商品的销售额。在该情况下，将实际的销售额作为训练数据使销售预测模型152进行学习，在被输入序列信息后输出预测销售额。另外，销售预测部17也可以不进行标准化而原样地输出销售数量或者销售额。
另外，在上述实施方式中，在生成序列信息时，将状态信息变换为数值向量，但不限定于此。状态信息仅有3个样式，所以也可以不将表示3个样式的数值“0：无存在”、“1：有存在”、“2：有变化”变换为向量而包含到序列信息。进而，关于位置信息以及属性信息，也可以不将用数值表现的信息变换为向量，而将数值原样地包含到序列信息。
另外，使销售预测模型152进行学习时的商品的状态信息以及位置信息不限定于从图像取得的信息。例如，也可以使用预先存储于存储部15的商品的货架分配表中所包含的状态信息以及位置信息，制作序列信息。

Claims (8)

  1. 一种销售预测方法，具备：
    取得用数值向量表现商品的信息的商品信息向量的步骤；
    取得所述商品的状态信息的步骤；
    取得所述商品的位置信息的步骤；
    生成包括所述商品信息向量、所述状态信息及所述位置信息的序列信息的步骤；以及
    使用销售预测模型，根据所述序列信息预测所述商品的销售数量或者销售额的步骤。
  2. 根据权利要求1所述的销售预测方法，其中，
    所述销售预测模型是自然语言处理模型。
  3. 根据权利要求1或者2所述的销售预测方法，其中，
    所述销售预测方法还具备取得所述商品的图像的步骤，
    所述状态信息以及所述位置信息从所述图像取得。
  4. 根据权利要求1～3中的任意一项所述的销售预测方法，其中，
    所述销售预测方法还具备将所述状态信息变换为用所述数值向量表现的状态信息向量的步骤，
    所述序列信息包括所述商品信息向量、所述状态信息向量以及所述位置信息。
  5. 根据权利要求1～4中的任意一项所述的销售预测方法，其中，
    所述销售预测方法还具备取得所述商品的价格信息的步骤，
    所述序列信息还包括所述商品的所述价格信息。
  6. 根据权利要求1～5中的任意一项所述的销售预测方法，还具备：
    判定通过所述销售预测模型预测出的所述商品的销售数量或者销售额是否为规定数量以上的步骤；
    在通过所述销售预测模型预测出的所述商品的销售数量或者销售额小于规定数量的情况下，变更所述位置信息来生成所述序列信息的步骤；以及
    在通过所述销售预测模型预测出的所述商品的销售数量或者销售额是规定数量以上的情况下，输出包含于所述序列信息的所述位置信息的步骤。
  7. 一种销售预测装置，具备：
    状态信息取得部，取得商品的状态信息；
    位置信息取得部，取得所述商品的位置信息；
    序列信息生成部，生成包括用数值向量表现所述商品的信息的商品信息向量、所述状态信息及所述位置信息的序列信息；以及
    销售预测部，使用销售预测模型，根据所述序列信息预测所述商品的销售数量或者销售额。
  8. 一种销售预测模型的学习方法，包括：
    生成包括用数值向量表现商品的信息的商品信息向量、所述商品的状态信息及所述商品的位置信息的序列信息的步骤；
    使用销售预测模型，根据所述序列信息预测所述商品的销售数量或者销售额的步骤；以及
    评价预测出的所述销售数量或者销售额的步骤。
PCT/CN2022/087568 2022-04-19 2022-04-19 销售预测方法、销售预测装置及销售预测模型的学习方法 WO2023201510A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202280060322.8A CN117916758A (zh) 2022-04-19 2022-04-19 销售预测方法、销售预测装置及销售预测模型的学习方法
PCT/CN2022/087568 WO2023201510A1 (zh) 2022-04-19 2022-04-19 销售预测方法、销售预测装置及销售预测模型的学习方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/087568 WO2023201510A1 (zh) 2022-04-19 2022-04-19 销售预测方法、销售预测装置及销售预测模型的学习方法

Publications (1)

Publication Number Publication Date
WO2023201510A1 true WO2023201510A1 (zh) 2023-10-26

Family

ID=88418952

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/087568 WO2023201510A1 (zh) 2022-04-19 2022-04-19 销售预测方法、销售预测装置及销售预测模型的学习方法

Country Status (2)

Country Link
CN (1) CN117916758A (zh)
WO (1) WO2023201510A1 (zh)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140351011A1 (en) * 2013-05-23 2014-11-27 Oracle International Corporation Retail sales forecast system with promotional cross-item effects prediction
CN112241904A (zh) * 2020-10-23 2021-01-19 浙江集享电子商务有限公司 商品销售量预测方法、装置、计算机设备和存储介质
CN113553540A (zh) * 2020-04-24 2021-10-26 株式会社日立制作所 一种商品销量的预测方法
WO2022044924A1 (ja) * 2020-08-27 2022-03-03 株式会社Ｎｔｔドコモ 棚割情報生成装置および予測モデル

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140351011A1 (en) * 2013-05-23 2014-11-27 Oracle International Corporation Retail sales forecast system with promotional cross-item effects prediction
CN113553540A (zh) * 2020-04-24 2021-10-26 株式会社日立制作所 一种商品销量的预测方法
WO2022044924A1 (ja) * 2020-08-27 2022-03-03 株式会社Ｎｔｔドコモ 棚割情報生成装置および予測モデル
CN112241904A (zh) * 2020-10-23 2021-01-19 浙江集享电子商务有限公司 商品销售量预测方法、装置、计算机设备和存储介质

Also Published As

Publication number Publication date
CN117916758A (zh) 2024-04-19

Similar Documents

Publication Publication Date Title
JP7061536B2 (ja) 最適化装置、シミュレーションシステム及び最適化方法
JP7140410B2 (ja) 予測システム、予測方法および予測プログラム
KR102420715B1 (ko) 시스템 강화 학습 방법 및 장치, 전자 기기, 컴퓨터 저장 매체
JP7263463B2 (ja) 推奨モデルを決定し、物品価格を決定する方法、装置、電子機器、記憶媒体およびコンピュータプログラム
US20180314978A1 (en) Learning apparatus and method for learning a model corresponding to a function changing in time series
US20230316720A1 (en) Anomaly detection apparatus, anomaly detection method, and program
KR101635283B1 (ko) 행렬 분해 모델 기반 데이터 분석 방법 및 장치
JP2021043477A (ja) 需要予測装置、需要予測方法、及びプログラム
WO2023201510A1 (zh) 销售预测方法、销售预测装置及销售预测模型的学习方法
US11847389B2 (en) Device and method for optimizing an input parameter in a processing of a semiconductor
US11568264B2 (en) Using shape information and loss functions for predictive modelling
US11210566B2 (en) Training apparatus, training method and recording medium
JP7196933B2 (ja) 学習装置および学習方法
JP7489275B2 (ja) 情報処理装置、情報処理システムおよび情報処理方法
US11783215B2 (en) Information processing apparatus and recommendation control method
CN108629062A (zh) 用于定价优化的方法、装置和系统
WO2018002967A1 (ja) 情報処理システム、情報処理方法、及び、記録媒体
CN114155422A (zh) 一种视觉问题回答的实现方法、装置、设备及存储介质
JP2011081451A (ja) 表示データ生成装置及びプログラム
CN114596120B (zh) 一种商品销量预测方法、系统、设备及存储介质
US20190236473A1 (en) Autonomous Hybrid Analytics Modeling Platform
CN111694945A (zh) 基于神经网络的法条关联推荐方法及装置
JP7198474B2 (ja) モデリングシステム
JP6753442B2 (ja) モデル生成装置、モデル生成方法、及びプログラム
KR102594173B1 (ko) 타임 시리즈 데이터 예측을 위한 학습 이미지를 생성하는 방법 및 장치

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22937748

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 202280060322.8

Country of ref document: CN