KR102050855B1 - Apparatus and method for forecasting demand - Google Patents

Abstract

The present invention relates to an apparatus and method for demand forecasting and, more specifically, to an apparatus and method capable of demographic demand forecasting of a product based on deep learning. According to one embodiment of the present invention, the apparatus may provide demand forecasting information for product by performing deep learning on information about at least one of product information, purchaser information, weather, and holiday. The apparatus comprises: an information management unit; a product information unit; a user information unit; a meta information unit; and a demand forecasting unit.

Description

Demand Forecasting Device and Method {APPARATUS AND METHOD FOR FORECASTING DEMAND}

The present invention relates to the demand forecasting of goods, and more particularly, to an apparatus and method for predicting demographic goods demand for analyzing various information affecting the purchase of goods based on deep learning.

In general, the method of estimating the demand of goods uses a statistical method. Usually, moving average, exponential smoothing, or linear regression using time series data including time and demand data is used. However, since the user who wants to purchase a product makes a purchase decision based on various factors, the forecast of product demand, which is simply performed using time series data, is not accurate. In addition, due to various purchasing factors, it is difficult to predict future demand simply by historical demand data. Due to incorrect demand forecast, sellers or manufacturers may have difficulty in distribution, manufacturing, or inventory management, and purchase customers may have low purchase satisfaction due to delay in delivery.

Background of the present invention is disclosed in Republic of Korea Patent No. 10-1504292.

The present invention provides an apparatus and method for predicting demographic goods demand that can be categorized by gender, age, and region by analyzing attributes affecting product purchase using deep learning.

According to an aspect of the present invention, a demand forecasting apparatus is provided.

The demand prediction apparatus according to an embodiment of the present invention includes an information management unit for acquiring and managing information about a buyer and a product, a product information unit for extracting product feature values based on deep learning from product attribute information, a gender, age and The user information unit extracts user feature values based on deep learning from age information, the meta information unit extracts meta feature values based on deep learning from meta information affecting purchases, and the demand predictor unit that predicts demographic demand of products. can do.

According to another aspect of the present invention, there is provided a computer-readable recording medium on which a demand forecasting method and a computer program for executing the same are recorded.

The method and the recording medium according to an embodiment of the present invention comprises the steps of calculating the product information as a product embedding vector value using deep learning, calculating the purchased user information as a user embedding vector value using deep learning, purchase Calculating meta information that affects the meta feature value using deep learning, calculating the probability of purchase by mapping the product embedding vector value to the user embedding vector value using an auto-encoder conversion function, and demand prediction Learning a transfer model from the product embedding vector value and the meta feature value of a desired product to extract an appropriate user embedding vector value for the product for which the demand is predicted, based on the appropriate user embedding vector value Extracting proximity user information and demographics from the proximity user information; By calculating the expression frequency information about enemy may include the step of predicting product demand.

According to one embodiment of the present invention, it is possible to more accurately predict the demand of goods by analyzing the properties that affect the purchase of goods using deep learning. Through the learned deep learning model, product purchasing probability by gender, age, and region can be calculated to predict the demographic demand of products to enable product production and proper inventory management, and close to the consumption area for efficient delivery of goods. Can arrange goods for efficient delivery.

1 to 5 are diagrams for explaining the demand prediction apparatus according to an embodiment of the present invention.
6 and 7 are diagrams for explaining a demand forecasting method according to an embodiment of the present invention.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the written description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. In describing the present invention, when it is determined that the detailed description of the related known technology may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. Also, the singular forms used in the specification and claims are to be interpreted as generally meaning "one or more", unless stated otherwise.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, and in the following description with reference to the accompanying drawings, the same or corresponding components are given the same reference numerals and redundant description thereof will be omitted. Shall be.

1 to 5 are diagrams for explaining a demand forecasting apparatus according to an embodiment of the present invention.

Referring to FIG. 1, the demand prediction device 10 collects product purchase information to predict product demand prediction information.

Referring to FIG. 2, the demand prediction apparatus 10 includes an information management unit 100, a product information unit 200, a user information unit 300, a meta information unit 400, and a demand prediction unit 500.

The information manager 100 collects and manages meta information that affects a buyer, a purchased product, and a purchase.

Referring to FIG. 3, the product information unit 200 may extract features of a product based on deep learning and group similar products through an embedding model. The product information unit 200 includes at least one of a product name, a product category, a price, and a sales amount as a product attribute. The product information unit 200 preprocesses and uses the product attributes. The product information unit 200 calculates a product name as a 128-dimensional word embedding vector trained by Word2Vec. The product category is a value obtained by concatenating a three-level category by converting a binary value. The commodity price is a real value that is converted from the selling price of a commodity using a log. Sales volume is the sales index value normalized between 0 and 1. The commodity information unit 200 concatenates the preprocessed commodity attributes, passes through a fully connected layer (FC-Layer), and normalizes them to Euclidean norm (L2-norm). The product information unit 200 extracts the feature from the attribute of the product, and converts the feature into a 128-dimensional feature vector by multiplying the weights. The converted 128-dimensional feature vector is a real value and is used for product embedding learning. Product embedding learning uses Word2Vec learning which converts words into vector values. Word2Vec learning method is a method that maps to a similar space by reducing the internal value of the words that are expressed together. For example, it may be a skip-gram model that predicts surrounding words as a center word of a Word2Vec learning model. Product embedding learning uses relationship information between products. The relationship information includes purchase information and may further include click information. The purchase information is information of products purchased by the buyer at the same time, and the click information is a product group searched by the buyer during the unit session (session) to view detailed information or put in a shopping cart.

Referring to FIG. 4, the user information unit 300 may classify and group users by gender, age, and region for product demand prediction. The user information unit 300 sets user's gender, age and region information as user attributes. The user information unit 300 concatenates user attribute information, normalizes to Euclid Norm (L2-norm) after passing through a 1-depth Fully Connected-Layer (FC-Layer). . The user information unit 300 extracts a user feature using the user attribute information and generates a 128-dimensional feature vector. The user information unit 300 uses the user 128-dimensional feature vector for user embedding learning. The user information unit 300 performs embedding learning through the dot product with relation information of users who purchased the same product in a session. For example, the demand prediction apparatus 10 may map (map) users who have purchased the same product by using the embedding learning result user embedding vector value.

Referring back to FIG. 1, the meta information unit 400 extracts features of meta information affecting a product purchase. The meta information unit 400 includes at least one of a day, a holiday, a sale period, a sale type, a weather, fine dust, and ultra fine dust as a meta information attribute. The meta information unit 400 concatenates meta information attributes, passes through a fully connected layer, and normalizes them to Euclidean norm or L2-norm. The meta information unit 400 extracts a meta feature using meta information and generates a 128-dimensional feature vector. The meta information unit 400 uses a user embedding learning result obtained by purchasing a product as a ground truth.

The demand prediction unit 500 calculates a purchase probability of each product by gender / age / region. The demand predictor 500 calculates a correlation between the product and the purchased user using the product feature vector and the user feature vector. The demand prediction unit 500 maps (maps) the product embedding vector to the user embedding vector by using an auto-encoder format conversion method. Equation 1 is a conversion function of a product embedding vector and a user embedding vector.

prod = g (x)

x: product attribute vector

g: product embedding extraction model

f: transform from product to user space

U: Buyer Characteristics

Referring to FIG. 5, the demand predictor 500 may extract the most appropriate user embedding information using a product embedding learning result and a meta embedding learning result in order to predict demand of an arbitrary product. The demand predictor 500 acquires purchased user information top-k that is close to the extracted user embedding information. For example, if k is 100, information of 100 nearby users may be acquired. The demand prediction unit 500 uses Euclidean and cosine similarity as functions for similarity metrics. The demand prediction unit 500 utilizes an Approximate Nearest Neighbors (ANN) algorithm to retrieve purchased user information. The demand prediction unit 500 calculates a purchase probability of a product based on gender, age, and region information of the retrieved user information. The demand prediction unit 500 calculates a frequency of expression for a user's gender, age, and region, and assumes a multinomial distribution to calculate a probability value.

6 and 7 are diagrams for describing a demand forecasting method according to an embodiment of the present invention.

Referring to FIG. 6, in operation S610, the demand prediction apparatus 10 collects and manages product attribute information, demographic characteristics (gender, age, region) information of the user, and purchase product list information of the user.

In operation S620, the demand prediction apparatus 10 calculates product information as a 128-dimensional product embedding vector value using deep learning. The demand prediction apparatus 10 includes a product name, a product category, a price, and a sales amount as attributes of a product. The demand prediction apparatus 10 passes through a one-depth full connection layer after concatenating the product attributes. The demand forecasting device 10 converts the product attributes, which have passed through the one-depth fully connected layer, into Euclidean norm or L2-norm and converts them into 128-dimensional product feature values. The demand prediction apparatus 10 calculates a product image feature value using the ResNet50 model that the product image is pre-trained with ImageNet. The demand prediction apparatus 10 calculates a product image feature value obtained by compressing the last layer of ResNet50 in 128 dimensions by using an autoencoder model. The demand prediction apparatus 10 concatenates the 128-dimensional commodity image feature value and the 128-dimensional commodity feature value and calculates the 128-dimensional final product feature value through the FC layer. The demand prediction apparatus 10 performs product embedding learning using the 128-dimensional final product feature value as an input vector. The demand prediction apparatus 10 embeds learning using a skip-gram model as relation information of purchase information and click information between products. The purchase information used as the related information is the goods purchased by the buyer in the same order, and the click information is the goods that a user searches for detailed information or puts them in a shopping cart during a session. The demand prediction apparatus 10 may group similar products by using the product embedding vector value calculated after the product embedding learning.

In operation S630, the demand prediction apparatus 10 calculates the purchased user information as a 128-dimensional user embedding vector value using deep learning. The demand prediction apparatus 10 calculates a 128-dimensional user feature value through deep learning using at least one user attribute information among gender, age, and region information of the user as an input vector. The demand prediction apparatus 10 concatenates all user attribute information to normalize after passing through a complete connection layer. The demand prediction apparatus 10 determines the users who purchased the same product in a session as a relation group and performs embedding learning through the inner product. The demand prediction apparatus 10 groups users by using a user embedding vector value calculated through embedding learning. In detail, users who purchase the same product may have similar vector values, and may be mapped to similar spaces and grouped.

In operation S640, the demand prediction apparatus 10 calculates meta information affecting a purchase as a 128-dimensional meta feature value using deep learning. The demand prediction apparatus 10 calculates the meta information characteristic value using the meta information of the product as an input vector. The meta information includes at least one of the day of the week, holiday status, sale period, sale type, weather, fine dust, and ultrafine dust that affect the purchase of the product.

In operation S650, the demand prediction apparatus 10 calculates a purchase probability by mapping the product embedding vector value to the user embedding vector value using an auto encoder conversion function.

In operation S660, the demand prediction apparatus 10 calculates user information that can be purchased using product information and meta information of a corresponding time point for predicting purchase demand. In other words. The transfer model is trained on the product embedding vector value and the meta feature value of the product for which the demand is predicted, and an appropriate user embedding vector value is extracted for the product for which the demand is predicted.

In operation S670, the demand prediction apparatus 10 extracts close user top-k information based on the extracted appropriate user embedding vector value. The demand prediction apparatus 10 uses Euclidean and cosine similarity as a function for similarity measurement to extract proximity user information. The demand prediction apparatus 10 calculates gender, age, and region information based on the top-k information, and calculates a purchase probability value. The demand prediction apparatus 10 may generate a user information index database in advance in order to calculate gender, age, and region-specific information using the buyer top-K information. The user information index database is not a traditional index based database. The user information index database retrieves user information using a user feature value as a search key. A query uses features, retrieves top-k information closest to the feature value, and provides metadata together. The demand prediction apparatus 10 may use an Approximate Nearest Neighbors (ANN) search method to process a user information search in real time.

In operation S680, the demand prediction apparatus 10 calculates an expression frequency of demographic information from the adjacent user top-K information to predict product demand. The demand prediction apparatus 10 calculates an expression frequency for gender, age, and region using the searched user metadata of top-k, and assumes a multinomial distribution to calculate the probability. The demand prediction apparatus 10 uses conditional probabilities such as P (buyer attribute information | goods) and P (goodness | buyer attribute information) to associate the buyer information with the goods. The buyer attribute information includes all genders, ages, and regions. For example, demand is forecasted by P ((gender, age, region) | product) rather than P (gender | product), P (age | good) and P (region | good).

Referring to FIG. 7, when the product is searched for, the demand prediction apparatus 10 provides a demographic demand forecasting result for the product through a model using deep learning. For example, the information predicted in step S680 may be displayed in a position on the map such as “30s / man / 100 Seoul” for the demand forecast for the product.

 The demand forecasting method may be embodied in the form of program instructions that can be executed by various computer means and recorded on a computer readable medium. Computer-readable media may include, alone or in combination with the program instructions, data files, data structures, and the like. The program instructions recorded on the computer readable medium may be those specially designed and constructed for the present invention, or may be known and available to those skilled in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Hardware devices specially configured to store and execute optical instructions and program instructions such as ROM, RAM, flash memory and the like. In addition, the above-described medium may be a transmission medium such as an optical or metal wire, a waveguide, or the like including a carrier wave for transmitting a signal specifying a program command, a data structure, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

So far, the present invention has been described with reference to the embodiments. Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

10: demand forecast device
100: Information Management Department
200: product information department
300: user information
400: meta information
500: demand forecasting unit

Claims (17)

In the demand forecasting device,
An information management unit for acquiring and managing information on buyers and products;
A product information unit to extract a product feature value based on deep learning from the product attribute information;
A user information unit extracting a buyer feature value based on deep learning from the gender, age and age information of the buyer;
A meta information unit for extracting meta feature values based on deep learning from meta information affecting a purchase; And
Includes a demand forecasting unit that predicts the demographic needs of a product,
In the product information unit, the product attribute information includes a product name, a product category, a product price, and a sales amount.
The trade name is a word embedding vector learned with Word2Vec,
The product category is a value obtained by concatenating a third-level category by converting a binary value.
The product price is a real value converted to a log,
The sales volume is a sales volume index value normalized to 0-1,
The product information unit,
Concatenate product attribute information, pass it through a Fully Connected Layer, and normalize it to Euclid Gambling,
The normalized product attribute information is calculated as a product feature value, and the product feature value is a real value.
Grouping similar products by performing embedding learning with the product feature values
The meta information unit,
At least one of the days of the week, whether it is a holiday, sales period, type of sale, weather, fine dust, and ultrafine dust that affects the purchase of the product is determined as meta information, and the meta feature value is calculated through deep learning.
The demand prediction unit calculates an appropriate user embedding vector value by mapping a product embedding vector value into a user embedding vector value using a conversion function of an auto-encoder format in order to calculate a correlation between a product and a purchased user.
Extracting close user top-k information based on the appropriate user embedding vector value,
Demand prediction apparatus for retrieving demographic information of the user top-k from the user information index database using the user top-k information.
delete delete delete delete delete delete delete The method of claim 1,
The user information unit,
The buyer's gender, age and age are the buyer attribute information, and the buyer attribute information is connected and normalized to Euclid gambling through a fully connected layer.
And calculating the normalized user attribute information as a user feature value.
The method of claim 9,
The user information unit
Demand prediction apparatus characterized in that the grouping of buyers by performing embedding learning by the user feature value.
delete delete In the demand prediction method of the demand prediction apparatus
Calculating product information as a product embedding vector value using deep learning;
Calculating purchased user information as a user embedding vector value using deep learning;
Calculating meta information affecting a purchase as a meta feature value using deep learning;
Calculating a purchase probability by mapping the product embedding vector value to the user embedding vector value using a conversion function of an auto encoder format;
Learning a transfer model of the product embedding vector value and the meta feature value of the product for which demand is predicted and extracting an appropriate user embedding vector value for the product for which the demand is predicted;
Extracting close user information based on the appropriate user embedding vector value; And
Calculating a frequency of expression for demographic information in the adjacent user information to predict product demand,
The step of calculating the product information as a product embedding vector value using deep learning
The product information is calculated based on deep learning by connecting the product information with a product name, a product category, a product price, and a sales volume.
The product image feature value is calculated based on the deep learning of the product image.
The deep learning-based 128-dimensional product feature value calculated using the product feature value and the product image feature value is used as a product embedding vector value.
Computing the purchased user information as a user embedding vector value using deep learning
Calculating a user characteristic value based on deep learning from the user attribute information connected by using the user information as a gender, an age, and a region;
Embedding learning information related to users who have purchased the same product within the unit time and the user feature value;
Computing a meta feature value based on deep learning from the meta information affecting the purchase
Day, holiday or not, a period of three days, three days kind of weather, fine and ultra-fine dust, wherein the meta information at least one of the dust affects the purchase and demand forecasting methods.
delete delete delete A computer program which executes the demand forecasting method of claim 13 and is recorded on a computer-readable recording medium.

Images