Search

KR-102961772-B1 - DEVICE AND METHOD FOR PREDICTING PRODUCT SALES VOLUME USING DEEP LEARNING MODEL

KR102961772B1KR 102961772 B1KR102961772 B1KR 102961772B1KR-102961772-B1

Abstract

The present invention relates to an apparatus and method for predicting product sales volume using a deep learning model, and more specifically, to an apparatus and method for predicting product sales volume using a TED (TransformerEncoderDNN) model that connects an encoder of a Transformer model with a DNN deep learning model. A TED model according to one embodiment of the present invention learns the combination of characteristics of each item of input data and item-specific weights through an encoder including self-attention of a transformer, and can predict the sales volume of a product on a specific future date through an RQ loss function and a new hyperparameter tuning method, thereby enabling the calculation of expected sales revenue for each specific date.

Inventors

  • 이선호

Assignees

  • 하트솔루션즈 주식회사

Dates

Publication Date
20260513
Application Date
20231025

Claims (8)

  1. In a product sales volume prediction device using a deep learning model, Input section for inputting data for predicting product sales volume; A model unit that generates a TED model using the above-mentioned input data; A learning unit that learns the generated TED model above; and It includes an output unit that outputs a predicted value of product sales volume on a specific future date using the above-mentioned trained TED model, The above input unit Input at least one of the following data: product data including the price, manufacturing date, manufacturer, country of origin, and ingredients of the above product; sales volume data including sales volume by price range and region; and specific date data including New Year's Day, Lunar New Year, March 1st Independence Movement Day, Children's Day, Memorial Day, Chuseok, National Foundation Day, Hangeul Day, and Christmas. The above model part A transformer encoder unit that applies item-specific importance to the input data through the transformer's self-attention module; and A weight layer is applied to the output value of the above transformer encoder unit through the VectorLayerNorm model using the following Equation 1. (Mathematical Formula 1) It includes a weighting layer that assigns weights to the importance of at least one of the following data: product data including the price, manufacturing date, manufacturer, country of origin, and ingredients of the product; sales volume data including sales volume by price range and region of the product; and specific day data including New Year's Day, Lunar New Year, March 1st Independence Movement Day, Children's Day, Memorial Day, Chuseok, National Foundation Day, Hangeul Day, and Christmas, by adding the value obtained by multiplying the output value of the transformer encoder unit by the hyperparameter α with x using the value of x as input to the deep learning model through the above mathematical formula 1 and using it as the input value of the VectorLayerNorm model. The above learning unit Train the above TED model by applying the RQ loss function, The above RQ loss function is As a composite loss function that combines the RMSLE loss function and the Quantile loss function, RQ = α * RMSLE + β * QuantileLoss (Equation 3) Here, using the grid search technique, α is set to 0.9 and β to 1.2 in the RQ loss function, and Training the above TED model by selecting hyperparameter values of a learning rate of 0.04, 16 attention layers, 32 weight layers, a batch size of 512, Adagrad momentum, and 10 epochs through hyperparameter tuning. Product sales volume prediction device using a deep learning model.
  2. delete
  3. delete
  4. delete
  5. delete
  6. In a method for predicting product sales volume using a deep learning model performed by a product sales volume prediction device using a deep learning model, Step of inputting data for predicting product sales volume; A step of generating a TED model using the above-mentioned input data; A step of training the generated TED model above; and The method includes the step of outputting a predicted value of product sales volume on a specific future date using the above-mentioned trained TED model, wherein The step of inputting data for predicting the sales volume of the above-mentioned product is Input at least one of the following data: product data including product price, manufacturing date, manufacturer, country of origin, and ingredients; sales volume data including sales volume by price range and region; and specific date data including New Year's Day, Lunar New Year, March 1st Independence Movement Day, Children's Day, Memorial Day, Chuseok, National Foundation Day, Hangeul Day, and Christmas. The step of generating a TED model using the above-mentioned input data is: A step of applying item-specific importance to the input data through the Transformer's self-attention module; and Apply a weight layer using the VectorLayerNorm model to the output value in the step of applying item-specific importance to the above input data through the Transformer's self-attention module, (Mathematical Formula 1) The method includes a step of assigning weights to the importance of at least one of the following data: product data including the price, manufacturing date, manufacturer, country of origin, and ingredients of the product; sales volume data including sales volume by price range and region of the product; and specific day data including New Year's Day, Lunar New Year, March 1st Independence Movement Day, Children's Day, Memorial Day, Chuseok, National Foundation Day, Hangeul Day, and Christmas, by adding the value obtained by multiplying the output value x in the step of applying item-specific importance through the self-attention module of the Transformer via the above mathematical formula 1 and the value output using x as input to the deep learning model as the input value of the VectorLayerNorm model. The step of training the generated TED model above is Train the above TED model by applying the RQ loss function, The above RQ loss function is As a composite loss function that combines the RMSLE loss function and the Quantile loss function, RQ = α * RMSLE + β * QuantileLoss (Equation 3) Here, using the grid search technique, α is set to 0.9 and β to 1.2 in the RQ loss function, and Training the TED model by selecting the above hyperparameter values through hyperparameter tuning as a learning rate of 0.04, the number of attention layers of 16, the number of weight layers of 32, a batch size of 512, momentum of Adagrad, and epochs of 10. Method for predicting product sales volume using a deep learning model.
  7. delete
  8. In paragraph 6 A method for predicting product sales volume using a deep learning model, comprising executing the method using the deep learning model and including a computer-readable recording medium.

Description

Device and Method for Predicting Product Sales Volume Using Deep Learning Model The present invention relates to an apparatus and method for predicting product sales volume using a deep learning model, and more specifically, to an apparatus and method for predicting product sales volume using a TED (TransformerEncoderDNN) model that connects an encoder of a Transformer model with a DNN deep learning model. Recently, artificial intelligence systems capable of achieving human-level intelligence are being used in various fields. Among the fields where artificial intelligence is applied, inference and forecasting is a technology that evaluates information to logically reason and predict. There is a growing trend among shopping malls and companies selling various products to utilize accurate AI-based forecasting in business management to predict sales volumes at specific future points in time and to manage inventory levels. In corporate management, accurately forecasting product sales volume is paramount for determining the appropriate supply quantity and providing consumers with the goods they need at the desired time. Furthermore, the most critical aspect is determining when to achieve ROI (Return on Investment) through accurate sales forecasting. However, since product sales are influenced by various variables, such as surges on specific days like Christmas or Lunar New Year, there has been a problem where operators lacking specialized knowledge face excessive costs and effort when attempting to predict the sales volume of each individual product. The background technology of the present invention is disclosed in Korean Registered Patent No. 10-2025281. FIG. 1 is a drawing for explaining a product sales volume prediction system according to an embodiment of the present invention. FIGS. 2 and 3 are drawings relating to the structure of a product sales volume prediction device according to an embodiment of the present invention. FIG. 4 is a drawing illustrating the weighting layer section according to an embodiment of the present invention. FIGS. 5 to 9 are drawings illustrating a loss function according to an embodiment of the present invention. FIG. 10 is a diagram illustrating a hyperparameter tuning process according to an embodiment of the present invention. FIG. 11 is a drawing for explaining a method for predicting product sales volume according to an embodiment of the present invention. FIG. 12 is a diagram illustrating a method for generating a TED model in a product sales volume prediction method according to an embodiment of the present invention. FIGS. 13 to 17 are drawings of a process for preprocessing input data used in a product sales volume prediction device according to an embodiment of the present invention. FIGS. 18 to 28 are drawings illustrating the training of a TED model and the verification of performance in a product sales volume prediction device according to an embodiment of the present invention. FIGS. 29 to 31 are graphs relating to the accuracy and loss value of a TED model according to an embodiment of the present invention. FIGS. 32 and 33 are drawings of a return on investment (ROI) prediction simulator applying a product sales volume prediction model according to an embodiment of the present invention. The present invention is susceptible to various modifications and may have various embodiments. Specific embodiments are illustrated in the drawings and described in detail through the detailed description. However, this is not intended to limit the present invention to specific embodiments, and it should be understood that the invention includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the invention. In describing the present invention, detailed descriptions of related prior art are omitted if it is determined that such detailed descriptions may unnecessarily obscure the essence of the invention. The present invention will be described below with reference to the attached drawings. However, the present invention may be implemented in various different forms and is therefore not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification have been given similar reference numerals. FIG. 1 is a drawing for explaining a product sales volume prediction system according to one embodiment of the present invention. Referring to FIG. 1, a product sales volume prediction system according to one embodiment of the present invention includes a user terminal (100) associated with a shopping mall, a network (200), and a product sales volume prediction device (300). A shopping mall may operate based on a specific website on the internet. According to one embodiment of the present invention, a user terminal (100) associated with the shopping mall includes a terminal of a shopping ma