Artificial intelligence-based demand prediction model learning and method, device, and computer program for predicting demand of product using same

Abstract

Provided are artificial intelligence-based demand prediction model learning and a method, device, and computer program for predicting a demand of a product using same. In artificial intelligence-based demand prediction model learning and a method for predicting a demand of a product using same according to various embodiments of the present disclosure, the method performed by a computing device comprises the steps of: training a demand prediction model by using a plurality of pieces of first product data for first products as training data; and analyzing second product data for a second product by using the trained demand prediction model to predict a demand for the second product.

Description

AI-based demand forecasting model training and method, device, and computer program for forecasting product demand using the same

[0001] Various embodiments of the present disclosure relate to artificial intelligence-based demand forecasting model learning and a method, apparatus, and computer program for forecasting the demand of a product using the same.

[0002] Success in the commercial industry depends heavily on the accuracy of product demand forecasting. In particular, forecasting demand for new products serves as a critical factor in terms of corporate inventory management and cost efficiency. Accurate demand forecasting contributes to meeting consumer needs while minimizing unnecessary inventory costs by producing and distributing the appropriate amount of products at the right time. However, if such forecasts are inaccurate, it can lead to increased inventory management costs due to overproduction, or conversely, the loss of sales opportunities resulting from an inability to meet demand.

[0003] In particular, new products face a significant challenge in demand forecasting due to the lack of historical sales data. This is known as the "cold start" problem, which refers to the difficulty in predicting sales volume when a new product is first launched in the market. This issue is particularly pronounced for trend-sensitive products such as books, fashion, and home appliances. Since initial sales performance significantly impacts the entire product life cycle of trend-sensitive products, early demand forecasting is crucial. Inaccurate forecasts can lead to excess inventory or stockouts, resulting in financial losses for the company.

[0004] Conventional demand forecasting methods are primarily based on time-series analysis, and it is common practice to predict future demand using historical sales data. This method analyzes market demand patterns based on data accumulated over a certain period and performs forecasts under the assumption that similar patterns will repeat. However, for new products, historical sales data is scarce or non-existent, making these traditional methods less effective. Consequently, existing demand forecasting methods have limitations in accurately predicting market response to new products.

[0005] To address this problem, methods based on similarity to existing products have traditionally been widely used to forecast demand for new products. This approach predicts demand by comparing the attributes of existing and new products and analyzing the sales history of similar products. For example, it involves matching the new product with existing ones by considering various attributes such as color, size, material, and price range, and then estimating the sales volume of the new product using sales data from that product group. However, this approach has several significant limitations.

[0006] First, there is the issue of simplification that arises during the attribute matching process between existing and new products. Traditional matching methods tend to rely on superficial criteria when judging the similarity between new and existing products. This may fail to accurately reflect detailed product characteristics or consumer preferences, and, particularly in the case of trend-sensitive products, it overlooks subtle differences that influence consumer purchasing decisions. For example, in the case of fashion products, even small changes in color and design can lead to significant differences in consumer response, yet these factors may not be sufficiently considered.

[0007] Second, there is the limited performance of the techniques used for attribute matching. Techniques such as K-Nearest Neighbors (KNN) are commonly used, but they perform matching by transforming various attributes into a simple vector space. However, these simplified embedding techniques fail to adequately reflect the complex interactions between product attributes. In particular, even when attempting to reflect various attributes using multimodal data, there are limitations in capturing the non-linear relationships between these attributes within a simple vector space. Consequently, this approach has limitations in accurately predicting the demand for new products.

[0008] The aforementioned background technology is one that the inventor possessed or acquired in the process of deriving the content of the present disclosure, and it cannot be considered as prior art disclosed to the general public prior to the filing of this application.

[0009] The problem that the present disclosure aims to solve is to provide an artificial intelligence-based demand forecasting model training, a method, an apparatus, and a computer program for forecasting product demand using the same, which can accurately predict demand even for new products for which past sales data does not exist, by training a demand forecasting model using multiple existing product data corresponding to existing products as training data for the purpose of resolving the aforementioned conventional problems, and analyzing new product data corresponding to new products using the same.

[0010] The problems that this disclosure aims to solve are not limited to those mentioned above, and other unmentioned problems will be clearly understood by a person skilled in the art from the description below.

[0011] A method for learning an artificial intelligence-based demand forecasting model and predicting the demand for a product using the same, according to one embodiment of the present disclosure for solving the above-described problem, may include a step of learning a demand forecasting model using a plurality of first product data for first products as learning data, and a step of predicting the demand for a second product by analyzing second product data for a second product using the learned demand forecasting model.

[0012] In various embodiments, the step of training the demand forecasting model may include the step of determining a loss function and the step of training the demand forecasting model such that the determined loss function is minimized.

[0013] In various embodiments, the determined loss function can be expressed as shown in Equation 1 below.

[0014] <Mathematical Formula 1>

[0015]

[0016] Here, the above is the above-determined loss function, above is a first loss function corresponding to a demand forecasting module included in the above demand forecasting model, the above and above is a second loss function and a third loss function corresponding to a similarity judgment module included in the above demand forecasting model, the above is a fourth loss function corresponding to data corruption of the above similarity judgment module, the above , above and above can be a pre-set weight.

[0017] In various embodiments, the first loss function may be expressed as shown in Equation 2 below.

[0018] <Mathematical Formula 2>

[0019]

[0020] Here, the above is the above first loss function, above is the actual demand value of the first products included in the plurality of first product data and the may be the expected demand value of the first products derived by analyzing the plurality of first product data through the above demand forecasting model.

[0021] In various embodiments, the second loss function may be expressed as shown in Equation 3 below.

[0022] <Mathematical Formula 3>

[0023]

[0024] Here, the above is the above second loss function, above and above is a size embedding vector generated from the plurality of first product data, the is the actual demand value of the first products included in the plurality of first product data, the is the result of row-wise shuffling the actual demand values ​​of the above-mentioned first products, the above The above As a result of shuffling row by row, the above is a function that calculates the similarity between vectors and the above can be a size similarity function that calculates the size similarity between time series data.

[0025] In various embodiments, the third loss function may be expressed as shown in Equation 4 below.

[0026] <Mathematical Formula 4>

[0027]

[0028] Here, the above is the above third loss function, above and above is a shape embedding vector generated from the plurality of first product data, the is the actual demand value of the first products included in the plurality of first product data, the is the result of row-wise shuffling the actual demand values ​​of the above-mentioned first products, the above The above As a result of shuffling row by row, the above is a function that calculates the similarity between vectors and the above can be a shape similarity function that calculates shape similarity between time series data.

[0029] In various embodiments, the fourth loss function may be expressed as shown in Equation 5 below.

[0030] <Mathematical Formula 5>

[0031]

[0032] Here, the above is the above-mentioned fourth loss function, above , above , above and above is an embedding vector generated from the plurality of first product data above, the , above , above and above The above , above , above and above The result of row-wise shuffling and the above can be a function that calculates the similarity between vectors.

[0033] In various embodiments, the step of training a demand forecasting model such that the determined loss function is minimized may include: generating a plurality of first embedding vectors by encoding the plurality of first product data; generating a plurality of training data including a plurality of positive samples containing mutually similar embedding vector pairs and a plurality of negative samples containing mutually dissimilar embedding vector pairs using the generated plurality of first embedding vectors; and performing contrastive learning using the generated plurality of training data so that a similarity judgment module included in the demand forecasting model learns the similarity between embedding vectors.

[0034] In various embodiments, the generated plurality of first embedding vectors include a plurality of first image embedding vectors generated by encoding a plurality of image data included in the plurality of first product data and a plurality of first text embedding vectors generated by encoding a plurality of text data included in the plurality of first product data, and the similarity determination module includes a size similarity determination module for determining size similarity between embedding vectors and a shape similarity determination module for determining shape similarity between embedding vectors, and the step of generating the plurality of training data comprises: a step of generating a plurality of first image attention vectors by attention pooling the plurality of first image embedding vectors; a step of generating a plurality of first text attention vectors by attention pooling the plurality of first text embedding vectors; a step of selecting a first size embedding vector that reflects characteristics regarding the size of the embedding vector among the generated plurality of first image attention vectors and the generated plurality of first text attention vectors, and generating first training data using the selected first size embedding vector; and the generated plurality of first image attention The method includes the step of selecting a vector and a first shape embedding vector among the plurality of first text attention vectors generated above that reflects characteristics regarding the shape of the embedding vector, and generating second training data using the selected first shape embedding vector, wherein the step of performing contrast learning may include the step of performing contrast learning for the size similarity judgment module using the generated first training data and performing contrast learning for the shape similarity judgment module using the generated second training data.

[0035] In various embodiments, the plurality of first image attention vectors generated above may be calculated using the following mathematical formula 6.

[0036] <Mathematical Formula 6>

[0037]

[0038] Here, the above is the first image attention vector, the above is the first query matrix, the above is the transpose matrix of the first key matrix, the above is the size of the dimension and the above can be a first value matrix.

[0039] In various embodiments, the plurality of first text attention vectors generated above may be calculated using the following mathematical formula 7.

[0040] <Mathematical Formula 7>

[0041]

[0042] Here, the above is the first text attention vector, above is the second query matrix, the above is the transpose matrix of the second key matrix, the above is the size of the dimension and the above can be a second value matrix.

[0043] In various embodiments, the step of training the demand forecasting model may include: generating a plurality of training data using the plurality of first product data; classifying the generated plurality of training data according to a plurality of key performance indicators (KPIs) related to demand; and generating a demand forecasting model that derives at least one of the plurality of key performance indicators as result data using specific product data as input data by training the demand forecasting model using the classified plurality of training data.

[0044] In various embodiments, the step of predicting demand may include the step of generating a second embedding vector by encoding the second product data, the step of selecting at least one first product data among the plurality of first product data based on the generated second embedding vector, and the step of predicting the demand for the second product using the selected at least one first product data.

[0045] In various embodiments, the second product data includes image data and text data containing information regarding the second product, and the step of generating the second embedding vector may include the step of generating a second image embedding vector by encoding the image data and the step of generating a second text embedding vector by encoding the text data.

[0046] In various embodiments, the step of generating the second embedding vector comprises: a step of generating a second image attention vector by attention pooling a second image embedding vector generated by encoding image data included in the second product data; and a step of generating a second text attention vector by attention pooling a second text embedding vector generated by encoding text data included in the second product data; and the step of selecting at least one first product data comprises: selecting a second size embedding vector among the generated second image attention vector and the generated second text attention vector that reflects characteristics regarding the size of the embedding vector, and calculating the size similarity between the selected second size embedding vector and a plurality of first size embedding vectors—the plurality of first size embedding vectors being generated from the plurality of first product data—through a size similarity judgment module included in the demand forecasting model, selecting a second shape embedding vector among the generated second image attention vector and the generated second text attention vector that reflects characteristics regarding the shape of the embedding vector, and The method may include a step of calculating shape similarity between the selected second shape embedding vector and a plurality of first shape embedding vectors—where the plurality of first size embedding vectors are generated from the plurality of first product data—through a shape similarity determination module included in the demand forecasting model, and a step of selecting at least one first product data among the plurality of first product data based on the calculated size similarity and the calculated shape similarity.

[0047] In various embodiments, the step of predicting the demand for the second product may include the step of predicting the demand for the second product by analyzing the selected at least one first product data through a deep learning-based demand prediction module included in the demand prediction model.

[0048] In various embodiments, the second product data includes meta information corresponding to a new sound source to be released at a specific time, and the plurality of first product data includes meta information corresponding to a plurality of existing sound sources released prior to the specific time and sales information, and the step of predicting demand may include the step of generating a second embedding vector by encoding meta information corresponding to the new sound source, the step of selecting at least one existing sound source among the plurality of existing sound sources by comparing a plurality of first embedding vectors generated from meta information corresponding to the plurality of existing sound sources with the generated second embedding vector, and the step of calculating the expected ranking, expected sales volume, and expected revenue of the new sound source as expected demand using sales information corresponding to the selected at least one existing sound source.

[0049] In various embodiments, the second product data includes meta-information corresponding to a new book to be published at a specific time, and the plurality of first product data includes meta-information corresponding to a plurality of existing books published prior to the specific time and sales information, and the step of predicting demand may include the step of generating a second embedding vector by encoding the meta-information corresponding to the new book, the step of selecting at least one existing book among the plurality of existing books by comparing the plurality of first embedding vectors generated from the meta-information corresponding to the plurality of existing books with the generated second embedding vector, and the step of calculating the expected sales volume of the new book as expected demand using the sales information corresponding to the selected at least one existing book.

[0050] A computing device for performing an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, according to another embodiment of the present disclosure for solving the above-described problem, comprises a processor, a network interface, a memory, and a computer program loaded into the memory and executed by the processor, wherein the computer program may include an instruction for learning a demand forecasting model using a plurality of first product data for first products as learning data, and an instruction for forecasting the demand of a second product by analyzing second product data for a second product using the learned demand forecasting model.

[0051] A computer program according to another embodiment of the present disclosure for solving the above-described problem may be combined with a computing device and stored on a recording medium readable by the computing device to execute an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, comprising the steps of: training a demand forecasting model using a plurality of first product data for first products as learning data; and predicting the demand of a second product by analyzing second product data for a second product using the trained demand forecasting model.

[0052] Other specific details of the present disclosure are included in the detailed description and drawings.

[0053] According to various embodiments of the present disclosure, by training a demand forecasting model using a plurality of existing product data corresponding to existing products as training data and analyzing new product data corresponding to new products using the same, there is an advantage in that demand can be accurately predicted even for new products for which past sales data does not exist.

[0054] The effects of the present disclosure are not limited to those mentioned above, and other unmentioned effects will be clearly understood by a person skilled in the art from the description below.

[0055] The following drawings attached to this specification illustrate preferred embodiments of the present disclosure and serve to further enhance understanding of the technical concept of the present disclosure together with the detailed description of the invention; therefore, the present disclosure should not be interpreted as being limited only to the matters described in such drawings.

[0056] FIG. 1 is a diagram illustrating an artificial intelligence-based demand forecasting model learning and a product demand forecasting system using the same according to one embodiment of the present disclosure.

[0057] FIG. 2 is a diagram illustrating an exemplary artificial intelligence-based demand forecasting model applicable to various embodiments.

[0058] FIG. 3 is a diagram illustrating the hardware configuration of a computing device that performs an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, according to another embodiment of the present disclosure.

[0059] FIG. 4 is a flowchart of an artificial intelligence-based demand forecasting model learning method according to another embodiment of the present disclosure.

[0060] FIG. 5 is a flowchart of a method for predicting the demand for a product using an artificial intelligence-based demand forecasting model according to another embodiment of the present disclosure.

[0061] FIGS. 6 and 7 are diagrams illustrating, in various embodiments, the process of predicting demand using a demand forecasting model.

[0062] The advantages and features of the present disclosure and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present disclosure is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure is complete and to fully inform those skilled in the art of the scope of the present disclosure, and the present disclosure is defined only by the scope of the claims.

[0063] The terms used herein are for describing the embodiments and are not intended to limit the disclosure. In this specification, the singular form includes the plural form unless specifically stated otherwise in the text. As used herein, "comprises" and / or "comprising" do not exclude the presence or addition of one or more other components in addition to the components mentioned.

[0064] Throughout this specification, the same reference numerals refer to the same components, and "and / or" includes each of the mentioned components and all combinations of one or more thereof. Although terms such as "first," "second," etc., are used to describe various components, they are not limited by these terms. These terms are used merely to distinguish one component from another. Accordingly, the first component mentioned below may be the second component within the technical scope of this disclosure.

[0065] As used herein, the terms “part” or “module” refer to hardware components such as software, FPGAs, or ASICs, and the “part” or “module” perform certain roles. However, the meaning of “part” or “module” is not limited to software or hardware. The “part” or “module” may be configured to reside in an addressable storage medium or may be configured to run on one or more processors. Thus, by example, the “part” or “module” includes components such as software components, object-oriented software components, class components, and task components, as well as processes, functions, attributes, procedures, subroutines, segments of program code, drivers, firmware, microcode, circuits, data, databases, data structures, tables, arrays, and variables. The functions provided within the components and “parts” or “modules” may be combined into a smaller number of components and “parts” or “modules,” or further separated into additional components and “parts” or “modules.”

[0066] Spatially relative terms such as "below," "beneath," "lower," "above," and "upper" may be used to facilitate the description of the relationship between one component and other components as illustrated in the drawings. Spatially relative terms should be understood as encompassing different orientations of components during use or operation, in addition to the orientations depicted in the drawings. For example, if a component depicted in a drawing is inverted, a component described as "below" or "beneath" of another component may be placed "above" of that component. Therefore, the exemplary term "below" may encompass both the lower and upper directions. Components may also be oriented in other directions, and accordingly, spatially relative terms may be interpreted according to the orientation.

[0067] Expressions such as "first," "second," or "first," "second" as used in this specification are used to distinguish one object from another when referring to a plurality of objects of the same kind, unless otherwise indicated in the context, and do not limit the order or importance of said objects.

[0068] Expressions used herein such as “A, B, and C,” “A, B, or C,” “A, B, and / or C,” or “at least one of A, B, and C,” “at least one of A, B, or C,” “at least one of A, B, and / or C,” “at least one selected from A, B, and C,” “at least one selected from A, B, or C,” “at least one selected from A, B, and / or C,” etc., may mean each of the listed items or all possible combinations of the listed items. For example, “at least one selected from A and B” may refer to (1) A, (2) at least one of A, (3) B, (4) at least one of B, (5) at least one of A and at least one of B, (6) at least one of A and B, (7) at least one of B and A, and (8) all of A and B.

[0069] As used herein, the expression “based on” is used to describe one or more factors affecting an act or action of a decision or judgment described in the phrase or sentence containing such expression, and such expression does not exclude additional factors affecting said act or action of a decision or judgment.

[0070] As used in this specification, the expression that a certain component (e.g., a first component) is "connected" or "connected" to another component (e.g., a second component) may mean that the said certain component is not only directly connected or connected to the said other component, but is also connected or connected through a new other component (e.g., a third component).

[0071] As used herein, the expression "configured to" may have meanings such as "set to," "capable of," "modified to," "made to," or "capable of." Such expression is not limited to the meaning of "specifically designed in hardware," and, for example, a processor configured to perform a specific operation may mean a generic-purpose processor capable of performing that specific operation by executing software.

[0072] Unless otherwise defined, all terms used herein (including technical and scientific terms) may be used in a meaning commonly understood by those skilled in the art to which this disclosure pertains. Additionally, terms defined in commonly used dictionaries are not to be interpreted ideally or excessively unless explicitly and specifically defined otherwise.

[0073] In this specification, the term "computer" refers to any type of hardware device comprising at least one processor, and may be understood to include software configurations operating on said hardware device according to the embodiments. For example, the term "computer" may be understood to include smartphones, tablet PCs, desktops, laptops, and user clients and applications running on each of these devices, but is not limited thereto.

[0074] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0075] Each step described in this specification is described as being performed by a computer, but the subject of each step is not limited thereto, and depending on the embodiment, at least some of each step may be performed on different devices.

[0076]

[0077] FIG. 1 is a diagram illustrating an artificial intelligence-based demand forecasting model learning and a product demand forecasting system using the same according to one embodiment of the present disclosure.

[0078] Referring to FIG. 1, an artificial intelligence-based demand forecasting model learning and product demand forecasting system using the same according to one embodiment of the present disclosure may include a computing device (100), a user terminal (200), an external server (300), and a network (400).

[0079] Here, the artificial intelligence-based demand forecasting model learning and product demand forecasting system using the same illustrated in FIG. 1 is according to one embodiment, and the components thereof are not limited to the embodiment illustrated in FIG. 1 and may be added, changed, or deleted as needed.

[0080] In one embodiment, the computing device (100) can perform an artificial intelligence-based demand forecasting model learning method for predicting the demand for a product.

[0081] In various embodiments, the computing device (100) can train a demand forecasting model using a plurality of first product data for first products as training data for the purpose of forecasting demand for a second product.

[0082] Here, the first product refers to existing products that have been used to train a demand forecasting model, and the second product refers to a product for which demand is to be predicted that is to be newly launched.

[0083] For example, the first product may be existing sound sources released prior to a specific point in time, and the second product may be new sound sources to be released at a specific point in time and / or after a specific point in time.

[0084] As another example, the first product may be existing books published prior to a specific point in time, and the second product may be new books to be published at and / or after a specific point in time. However, it is not limited thereto.

[0085] Additionally, the product data (first product data and second product data) is data containing information about the product, and may include, for example, image data containing information about the product and / or text data containing information about the product.

[0086] In addition, the product information here may include product meta-information and sales information.

[0087] Meta-information regarding a product is information regarding the characteristics of the product, and may include, for example, internal factor information such as product type, brand, model, price, size, color, material, weight, genre, manufacturer, grade, version, platform, etc., and external factor information such as weather, season, event.

[0088] In addition, sales information regarding the product is information related to product sales, and may include, for example, product sales volume (by time of day, period, season, region, age group, etc.), sales outlets, etc.

[0089] Here, a demand forecasting model (e.g., a neural network) consists of one or more network functions, and one or more network functions may consist of a set of interconnected computational units that can generally be referred to as 'nodes'. These 'nodes' may also be referred to as 'neurons'. One or more network functions are composed of at least one node. The nodes (or neurons) constituting one or more network functions may be interconnected by one or more 'links'.

[0090] In a demand forecasting model, one or more nodes connected via links can form a relative relationship between input and output nodes. The concepts of input and output nodes are relative; any node in an output node relationship with respect to one node may be in an input node relationship with respect to another node, and vice versa. As previously mentioned, the input node versus output node relationship can be generated based on links. One or more output nodes may be connected to a single input node via links, and vice versa.

[0091] In a relationship between input and output nodes connected via a single link, the value of the output node can be determined based on data input into the input node. Here, the link interconnecting the input and output nodes may have a weight. The weight may be variable and may be varied by a user or an algorithm to perform the desired function of the demand forecasting model. For example, if one or more input nodes are interconnected to a single output node by respective links, the output node value may be determined based on the values ​​input into the input nodes connected to the output node and the weight set on the link corresponding to each input node.

[0092] As described above, a demand forecasting model comprises one or more nodes interconnected through one or more links, forming input and output node relationships within the model. The characteristics of a demand forecasting model can be determined by the number of nodes and links within the model, the relationships between the nodes and links, and the weight values ​​assigned to each link. For example, if two demand forecasting models exist with the same number of nodes and links but different weight values ​​between the links, the two demand forecasting models may be recognized as distinct from each other.

[0093] Some of the nodes constituting a demand forecasting model may form a layer based on their distances from the initial input node. For example, a set of nodes with a distance of n from the initial input node may form n layers. The distance from the initial input node can be defined by the minimum number of links that must be traversed to reach the corresponding node from the initial input node. However, this definition of a layer is arbitrary for illustrative purposes, and the degree of a layer within the demand forecasting model may be defined in a way different from that described above. For example, the layer of nodes may be defined by their distance from the final output node.

[0094] The initial input node may refer to one or more nodes within the demand forecasting model to which data is directly input without passing through links in relation to other nodes. Alternatively, within the demand forecasting model network, in terms of relationships between nodes based on links, it may refer to nodes that do not have other input nodes connected by links. Similarly, the final output node may refer to one or more nodes within the demand forecasting model that do not have output nodes in relation to other nodes. Additionally, the hidden node may refer to nodes constituting the demand forecasting model that are neither the initial input node nor the final output node. A demand forecasting model according to one embodiment of the present disclosure may have more nodes in the input layer than nodes in the hidden layer that are close to the output layer, and may be a demand forecasting model in which the number of nodes decreases as one progresses from the input layer to the hidden layer.

[0095] A demand forecasting model may include one or more hidden layers. The hidden nodes of a hidden layer can take the output of the previous layer and the output of neighboring hidden nodes as inputs. The number of hidden nodes for each hidden layer may be the same or different. The number of nodes in the input layer may be determined based on the number of data fields in the input data and may be the same or different from the number of hidden nodes. The input data fed into the input layer can be processed by the hidden nodes of the hidden layer and output by the fully connected layer (FCL), which is the output layer.

[0096] In various embodiments, the demand forecasting model may be a deep learning model (e.g., FIG. 4).

[0097] A deep learning model (e.g., a deep neural network (DNN)) can refer to an artificial intelligence model that includes multiple hidden layers in addition to input and output layers. Using a deep neural network, one can identify the latent structures of data. That is, one can identify the latent structures of photos, text, videos, voice, and music (e.g., what objects are in a photo, what the content and emotions of the text are, what the content and emotions of the voice are, etc.).

[0098] Deep neural networks may include, but are not limited to, convolutional neural networks (CNN), recurrent neural networks (RNN), autoencoders, Generative Adversarial Networks (GAN), restricted Boltzmann machines (RBM), deep belief networks (DBN), Q networks, U networks, and Siamese networks.

[0099] In various embodiments, the network function may include an autoencoder. Here, the autoencoder may be a type of artificial neural network for outputting output data similar to the input data.

[0100] An autoencoder may include at least one hidden layer, and an odd number of hidden layers may be placed between the input and output layers. The number of nodes in each layer may be reduced from the number of nodes in the input layer to an intermediate layer called the bottleneck layer (encoding), and then expanded symmetrically from the bottleneck layer to the output layer (symmetrical to the input layer). The nodes of the dimensionality reduction layer and the dimensionality restoration layer may or may not be symmetrical. Additionally, the autoencoder can perform non-linear dimensionality reduction. The number of nodes in the input and output layers may correspond to the number of sensors remaining after the preprocessing of the input data. In the autoencoder structure, the number of nodes in the hidden layers included in the encoder may have a structure where it decreases as it moves away from the input layer. Since the number of nodes in the bottleneck layer (the layer with the fewest nodes located between the encoder and the decoder) may not transmit a sufficient amount of information if it is too small, it may be maintained at a certain number or higher (e.g., more than half the number of nodes in the input layer).

[0101] In various embodiments, the computing device (100) may perform a demand forecasting method for a product that forecasts the demand for a product using a demand forecasting model. To this end, the demand forecasting model may include, but is not limited to, an encoder, a similarity-based module, and a demand forecasting module as shown in FIG. 2.

[0102] First, the encoder can generate embedding vectors by encoding input data. In this case, the encoder can be implemented as a multimodal model capable of processing different input data simultaneously. For example, the encoder may be a combination of a ResNet-based model that converts image data into embedding vectors and a BERT-based model that converts text data into embedding vectors, but is not limited to this.

[0103] Next, the similarity determination module may be a module capable of selecting products similar to a specific product by generating multiple product groups with mutually similar features and patterns through pre-training by grouping data of products having similar features and patterns, and determining which product group a specific product belongs to based on the embedding vector of a specific product generated through an encoder. For example, the similarity determination module may be a model based on InfoNCE (Information Noise Contrastive Estimation) trained according to contrastive learning, but is not limited thereto.

[0104] In various embodiments, the similarity determination module may include, but is not limited to, a size similarity determination module that compares size similarity between embedding vectors and a shape similarity determination module that compares shape similarity between embedding vectors.

[0105] Next, the demand forecasting module may be a module that predicts the demand for a specific product by analyzing product data of products selected through the similarity judgment module. For example, the demand forecasting module may be a deep learning-based model, but is not limited thereto.

[0106] In various embodiments, the computing device (100) may be connected to a user terminal (200) via a network (400), and may obtain product data regarding a product for which demand is to be predicted from the user terminal (200), predict demand by analyzing the product data of the product through a demand prediction model, and provide information regarding the predicted demand to the user terminal (200).

[0107] Here, the user terminal (200) may refer to any form of entity(s) in a system having a mechanism for communicating with a computing device (100). For example, such a user terminal (200) may include a PC (personal computer), a notebook, a mobile terminal, a smartphone, a tablet PC, and a wearable device, and may include any type of terminal capable of connecting to a wired or wireless network. Additionally, the user terminal (200) may include any computing device implemented by at least one of an agent, an API (Application Programming Interface), and a plug-in. Additionally, the user terminal (200) may include an application source and / or a client application.

[0108] Additionally, the network (400) may refer to a connection structure capable of exchanging information between each node, such as multiple terminals and servers. For example, the network (400) may include a Local Area Network (LAN), a Wide Area Network (WAN), the World Wide Web (WWW), a wired / wireless data network, a telephone network, a wired / wireless television network, a Controller Area Network (CAN), and Ethernet.

[0109] Wireless data communication networks may include, but are not limited to, 3G, 4G, 5G, 3GPP (3rd Generation Partnership Project), 5GPP (5th Generation Partnership Project), LTE (Long Term Evolution), WIMAX (World Interoperability for Microwave Access), Wi-Fi, Internet, LAN (Local Area Network), Wireless LAN (Wireless Local Area Network), WAN (Wide Area Network), PAN (Personal Area Network), RF (Radio Frequency), Bluetooth network, NFC (Near-Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc.

[0110] In one embodiment, an external server (300) may be connected to a computing device (100) via a network (400) and may store and manage various information (e.g., product data for multiple products) necessary for the computing device (100) to learn an artificial intelligence-based demand forecasting model and perform a method for forecasting the demand of products using the same, or may collect, store, and manage various information and data (e.g., demand forecasting results) derived as the computing device (100) learns an artificial intelligence-based demand forecasting model and performs a method for forecasting the demand of products using the same. For example, the external server (300) may be a storage server separately provided outside the computing device (100), but is not limited thereto. Hereinafter, with reference to FIG. 3, the hardware configuration of a computing device (100) that learns an artificial intelligence-based demand forecasting model and performs a method for forecasting the demand of products using the same will be described.

[0111]

[0112] FIG. 3 is a diagram illustrating the hardware configuration of a computing device that performs an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, according to another embodiment of the present disclosure.

[0113] Referring to FIG. 3, a computing device (100) according to another embodiment of the present disclosure may include one or more processors (110), a memory (120) for loading a computer program (151) executed by the processor (110), a bus (130), a communication interface (140), and a storage (150) for storing the computer program (151). Here, FIG. 3 illustrates only the components related to the embodiments of the present disclosure. Accordingly, a person skilled in the art to which the present disclosure pertains will understand that other general-purpose components may be included in addition to the components illustrated in FIG. 3.

[0114] The processor (110) controls the overall operation of each component of the computing device (100). The processor (110) may be configured to include a CPU (Central Processing Unit), an MPU (Micro Processor Unit), an MCU (Micro Controller Unit), a GPU (Graphic Processing Unit), or any form of processor well known in the art of the present disclosure.

[0115] Additionally, the processor (110) may perform operations for at least one application or program for executing the method according to the embodiments of the present disclosure, and the computing device (100) may have one or more processors.

[0116] In various embodiments, the processor (110) may further include Random Access Memory (RAM) (not shown) and Read-Only Memory (ROM) (not shown) for temporarily and / or permanently storing signals (or data) processed within the processor (110). Additionally, the processor (110) may be implemented in the form of a System on Chip (SoC) comprising at least one of a graphics processing unit, RAM, and ROM.

[0117] Memory (120) stores various data, instructions and / or information. Memory (120) may load a computer program (151) from storage (150) to execute a method / operation according to various embodiments of the present disclosure. When the computer program (151) is loaded into memory (120), the processor (110) may perform the method / operation by executing one or more instructions constituting the computer program (151). Memory (120) may be implemented as a volatile memory such as RAM, but the technical scope of the present disclosure is not limited thereto.

[0118] The bus (130) provides communication functions between components of the computing device (100). The bus (130) can be implemented as various types of buses, such as an address bus, a data bus, and a control bus.

[0119] The communication interface (140) supports wired and wireless internet communication of the computing device (100). Additionally, the communication interface (140) may support various communication methods other than internet communication. To this end, the communication interface (140) may be configured to include a communication module well known in the art of the present disclosure. In some embodiments, the communication interface (140) may be omitted.

[0120] Storage (150) can store computer programs (151) non-temporarily. When an AI-based demand forecasting model learning and a product demand forecasting process using the same are performed through a computing device (100), storage (150) can store various information necessary to provide AI-based demand forecasting model learning and a product demand forecasting process using the same.

[0121] The storage (150) may be configured to include non-volatile memory such as ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), flash memory, a hard disk, a removable disk, or any form of computer-readable recording medium well known in the art to which the present disclosure belongs.

[0122] A computer program (151) may include one or more instructions that cause a processor (110) to perform a method / operation according to various embodiments of the present disclosure when loaded into memory (120). That is, the processor (110) may perform the method / operation according to various embodiments of the present disclosure by executing the one or more instructions.

[0123] In one embodiment, the computer program (151) may include one or more instructions for performing an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, the steps of: learning a demand forecasting model using a plurality of first product data for first products as learning data; and predicting the demand for a second product by analyzing second product data for a second product using the learned demand forecasting model.

[0124] The steps of the method or algorithm described in connection with the embodiments of the present disclosure may be implemented directly in hardware, implemented as a software module executed by hardware, or implemented by a combination thereof. The software module may reside in RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically Erasable Programmable ROM), Flash Memory, a hard disk, a removable disk, a CD-ROM, or any form of computer-readable recording medium well known in the art to which the present disclosure belongs.

[0125] The components of the present disclosure may be implemented as a program (or application) and stored on a medium to be executed in combination with a computer, which is hardware. The components of the present disclosure may be executed as software programming or software elements, and similarly, embodiments may be implemented in programming or scripting languages ​​such as C, C++, Java, assembler, etc., including various algorithms implemented as combinations of data structures, processes, routines, or other programming components. Functional aspects may be implemented as algorithms executed on one or more processors. Hereinafter, with reference to FIGS. 4 and 5, an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same according to various embodiments of the present disclosure will be described.

[0126]

[0127] FIG. 4 is a flowchart of an artificial intelligence-based demand forecasting model learning method according to another embodiment of the present disclosure.

[0128] Referring to FIG. 4, in step S110, the computing device (100) can determine a loss function for learning a demand forecasting model.

[0129] Here, the loss function can be expressed as shown in Equation 1 below.

[0130] <Mathematical Formula 1>

[0131]

[0132] Here, is the determined loss function, is the first loss function, and is the second loss function and the third loss function, is the fourth loss function, , and can be a pre-set weight.

[0133] First, the first loss function is a loss function determined in correspondence with the demand forecasting module included in the demand forecasting model, and may be a loss function used for deep learning training for the demand forecasting module.

[0134] The first loss function may be used to fine-tune the demand forecasting model to minimize the difference between the predicted demand value and the actual sales volume based on product data, and can be expressed as Equation 2 below.

[0135] <Mathematical Formula 2>

[0136]

[0137] Here, is the first loss function, is the actual demand value of the first products, and It may be the expected demand value of the first products derived by analyzing multiple first product data through a demand forecasting model.

[0138] In addition, the expected demand value of the first products derived through the demand forecasting model can be calculated using the following mathematical formula 2-1.

[0139] <Mathematical Formula 2-1>

[0140]

[0141] Here, is the weight matrix, and is a size embedding vector generated from multiple first product data, and is a shape embedding vector generated from multiple first product data and can be a bias term added to the linear combination of weights and input vectors.

[0142] In other words, the first loss function is a function that calculates the absolute error between the predicted value and the actual value, and by training the demand forecasting model to minimize this first loss function, the model can be guided to learn in a direction that produces more accurate demand forecasting results.

[0143] Next, the second loss function and the third loss function are loss functions determined in correspondence with the similarity judgment module included in the demand forecasting model, and may be loss functions used for contrastive learning for the similarity judgment module.

[0144] In addition, the second and third loss functions may be intended to induce the similarity judgment module to learn mutually similar product data pairs as close distances and dissimilar product data pairs as far distances.

[0145] Here, the second loss function is intended to perform contrastive learning for a size similarity determination module that determines the size similarity between embedding vectors generated from product data, and can be expressed as Equation 3 below.

[0146] <Mathematical Formula 3>

[0147]

[0148] Here, is the second loss function, and is a size embedding vector generated from multiple first product data, is the actual demand value of the first products included in multiple first product data, is the result of row-wise shuffling the actual demand values ​​of the first products, Is The result of shuffling row by row, is a function that calculates the similarity between vectors and can be a size similarity function that calculates the size similarity between time series data.

[0149] Also, here, a function that calculates the similarity between vectors ( ) can be determined according to the similarity calculation method to be applied (e.g., cosine similarity, inner product, etc.).

[0150] Also, here, a size similarity function that calculates the size similarity between time series data It can be expressed as shown in the following mathematical formula 3-1.

[0151] <Mathematical Formula 3-1>

[0152]

[0153] Here, the third loss function is intended to perform contrastive learning for a shape similarity determination module that determines shape similarity between embedding vectors generated from product data, and can be expressed as Equation 4 below.

[0154] <Mathematical Formula 4>

[0155]

[0156] Here, is the third loss function, and is a shape embedding vector generated from multiple first product data, is the actual demand value of the first products included in multiple first product data, is the result of row-wise shuffling the actual demand values ​​of the first products, Is The result of shuffling row by row, is a function that calculates the similarity between vectors and can be a shape similarity function that calculates shape similarity between time series data.

[0157] Here, a shape similarity function that calculates the shape similarity between time series data It can be expressed as shown in the following mathematical formula 4-1.

[0158] <Mathematical Formula 4-1>

[0159]

[0160] Finally, the fourth loss function may be a loss function determined by considering data corruption in the similarity judgment module included in the demand forecasting model.

[0161] When high-dimensional product data, such as image and text data, is converted into low-dimensional embedding vectors and used to train a demand forecasting model, a problem of data degradation may occur where the embedding vectors gradually lose diversity and converge to similar values ​​during the training process.

[0162] This phenomenon occurs when embedding vectors fail to reflect diverse characteristics and provide differentiated representations among similar data points. If different input data are represented by similar vectors during the embedding process, the demand forecasting model becomes unable to learn the differences between these vectors, resulting in a problem where the forecasting performance of the model deteriorates.

[0163] Considering these points, the function determines a fourth loss function to ensure that embedding vectors maintain diverse representations and prevent them from becoming excessively similar to each other, and trains the demand forecasting model to minimize the fourth loss function, thereby enabling the demand forecasting model to learn while maintaining the differentiated characteristics of the input data. Here, the fourth loss function can be expressed as shown in Equation 5 below.

[0164] <Mathematical Formula 5>

[0165]

[0166] Here, is the fourth loss function, , , and is an embedding vector generated from multiple first product data, , , and Is , , and The result of row-wise shuffling and can be a function that calculates the similarity between vectors.

[0167] In step S120, the computing device (100) can train a demand forecasting model such that the loss function determined through step S110 is minimized.

[0168] In various embodiments, the computing device (100) can perform comparative learning for a similarity judgment module of a demand forecasting model using a plurality of first product data as training data.

[0169] More specifically, first, the computing device (100) can generate a plurality of first embedding vectors using a plurality of first product data. For example, the computing device (100) can generate a plurality of first embedding vectors by encoding a plurality of first product data through an encoder of a demand forecasting model.

[0170] In various embodiments, the computing device (100) can generate a plurality of first image embedding vectors by encoding a plurality of image data included in a plurality of first product data, and generate a plurality of first text embedding vectors by encoding a plurality of text data included in a plurality of first product data.

[0171] Next, the computing device (100) can generate a plurality of training data including a plurality of positive samples including pairs of mutually similar embedding vectors and a plurality of voice samples including pairs of mutually dissimilar embedding vectors using a plurality of first embedding vectors.

[0172] In various embodiments, the computing device (100) may individually generate first training data for a size similarity determination module and second training data for a shape similarity determination module for the purpose of performing individual contrast learning for a size similarity determination module for determining size similarity between embedding vectors and a shape similarity determination module for determining shape similarity between embedding vectors.

[0173] For example, the computing device (100) can generate a plurality of first image attention vectors by attention pooling a plurality of first image embedding vectors and generate a plurality of first text attention vectors by attention pooling a plurality of first text embedding vectors, and can generate first training data and second training data using the plurality of first image attention vectors and the plurality of first text attention vectors.

[0174] For example, the computing device (100) has a first size embedding vector (which reflects characteristics regarding the size of the embedding vector among a plurality of first image attention vectors and a plurality of first text attention vectors). , You can select ) and generate first training data using the first size embedding vector.

[0175] In addition, the computing device (100) has a first shape embedding vector (which reflects characteristics regarding the shape of the embedding vector among a plurality of first image attention vectors and a plurality of first text attention vectors) , You can select ) and generate second training data using the first form embedding vector.

[0176] Here, multiple first image attention vectors can be calculated through the following mathematical formula 6.

[0177] <Mathematical Formula 6>

[0178]

[0179] Here, is the first image attention vector, is the first query matrix, is the transpose of the first key matrix, is the size of the dimension and can be a first value matrix.

[0180] In addition, the first query matrix, the first key matrix, and the first value matrix can be expressed by the following mathematical formula 6-1.

[0181] <Mathematical Formula 6-1>

[0182]

[0183] Here, is the first query matrix, is the first key matrix, is the first value matrix, is a function that generates image embedding vectors from image data, , and is the weight matrix for each of the first query matrix, the first key matrix, and the first value matrix, , and can be the bias term of the first query matrix, the first key matrix, and the first value matrix, respectively.

[0184] In addition, here, a plurality of first text attention vectors can be calculated through the following mathematical formula 7.

[0185] <Mathematical Formula 7>

[0186]

[0187] Here, is the first text attention vector, is the second query matrix, is the transpose of the second key matrix and can be a second-valued matrix.

[0188] In addition, the second query matrix, the second key matrix, and the second value matrix can be expressed by the following mathematical formula 7-1.

[0189] <Mathematical Formula 7-1>

[0190]

[0191] Here, is the second query matrix, is the second key matrix, is the second-valued matrix, is a function that generates text embedding vectors from text data, , and is the weight matrix for each of the second query matrix, second key matrix, and second value matrix, , and can be the bias term of the second query matrix, the second key matrix, and the second value matrix, respectively.

[0192] Next, the computing device (100) can perform contrast learning using multiple training data so that the similarity judgment module learns similarity between embedding vectors. For example, the computing device (100) can perform contrast learning for a size similarity judgment module using multiple first training data and perform contrast learning for a shape similarity judgment module using multiple second training data.

[0193] In various embodiments, the computing device (100) can classify multiple training data generated using multiple first product data according to multiple key performance indicators (KPIs) related to demand, and can generate a demand forecasting model that derives at least one key performance indicator among multiple key performance indicators as result data by training a demand forecasting model using multiple training data classified according to key performance indicators.

[0194] For example, when constructing a demand forecasting model to predict the demand for music, product data of existing music can be classified according to multiple key performance indicators related to the demand for music (e.g., daily / weekly / monthly / yearly rankings, album sales, streaming counts, music rankings, etc.), and a demand forecasting model can be trained using the product data classified according to multiple key performance indicators related to the demand for music as training data. This allows for the construction of a demand forecasting model that takes product data of a specific music as input data and derives at least one of the weekly / monthly / yearly rankings, album sales, streaming counts, and music rankings as result data.

[0195] That is, the computing device (100) can build a demand forecasting model capable of deriving multiple different key performance indicators, rather than building a model that derives each key performance indicator individually by utilizing product data containing information regarding various types of key performance indicators.

[0196]

[0197] FIG. 5 is a flowchart of a method for predicting the demand for a product using an artificial intelligence-based demand forecasting model according to another embodiment of the present disclosure.

[0198] Referring to FIG. 5, in step S210, the computing device (100) can generate a second embedding vector by encoding second product data for the second product. For example, the computing device (100) can generate a second embedding vector by encoding the second product data through an encoder of a demand forecasting model.

[0199] In various embodiments, when the second product data includes image data and text data containing information about the second product, the computing device (100) can generate a second image embedding vector by encoding the image data and generate a second text embedding vector by encoding the text data.

[0200] In step S220, the computing device (100) can select at least one first product data among a plurality of first product data based on the second embedding vector generated through step S210. For example, the computing device (100) can select at least one first product data among a plurality of first product data through a similarity judgment module of a demand forecasting model.

[0201] In various embodiments, the computing device (100) can select at least one first product data similar to the second product data among the plurality of first product data by comparing the size and shape similarity between an embedding vector generated from the second product data and an embedding vector generated from the plurality of first product data.

[0202] More specifically, first, the computing device (100) can generate a second image attention vector by attention pooling the second image embedding vector generated by encoding image data included in the second product data.

[0203] Additionally, the computing device (100) can generate a second text attention vector by attention pooling the second text embedding vector generated by encoding the text data included in the second product data.

[0204] Subsequently, the computing device (100) selects a second size embedding vector that reflects characteristics regarding the size of the embedding vector among the second image attention vector and the second text attention vector, and can calculate the size similarity between the second size embedding vector and a plurality of first size embedding vectors through a size similarity judgment module included in the demand forecasting model.

[0205] Additionally, the computing device (100) can select a second shape embedding vector that reflects characteristics regarding the shape of the embedding vector among the second image attention vector and the second text attention vector, and calculate the shape similarity between the second shape embedding vector and a plurality of first shape embedding vectors through a shape similarity judgment module included in the demand forecasting model.

[0206] Here, a plurality of first-size embedding vectors and a plurality of first-shape embedding vectors are generated from first product data for a plurality of first products, and may be generated in the same manner as the second-size embedding vector and the second-shape embedding vector.

[0207] Subsequently, the computing device (100) may select at least one first product data among a plurality of first product data based on size similarity and shape similarity calculated according to the above method. For example, the computing device (100) may sequentially select N first product data starting from the first product data with a high sum of size similarity and shape similarity, or select first product data among the plurality of first product data where the sum of size similarity and shape similarity is greater than or equal to a reference value, but is not limited thereto.

[0208] In step S230, the computing device (100) can predict the demand for a second product based on at least one first product data selected through step S220. For example, the computing device (100) can predict the demand for a second product by analyzing at least one first product data through a demand prediction module of a demand prediction model.

[0209] For example, referring to FIG. 6, if a second product and second product data include a new sound source to be released at a specific time and corresponding meta information, and a plurality of first products and a plurality of first product data include a plurality of existing sound sources released before a specific time and corresponding meta information and sales information, the computing device (100) can determine the demand for the new sound source by analyzing the meta information of the new sound source.

[0210] First, the computing device (100) can generate a second embedding vector by encoding meta information corresponding to a new sound source through the encoder of the demand forecasting model.

[0211] Subsequently, the computing device (100) can select at least one existing sound source that is determined to be similar to a new sound source among the existing sound sources by comparing a plurality of first embedding vectors and a second embedding vector generated from meta information corresponding to a plurality of existing sound sources through a similarity judgment module of a demand forecasting model.

[0212] Afterwards, the computing device (100) can derive the expected ranking, expected sales volume, and expected revenue of the new sound source as the demand for the new sound source by analyzing sales information corresponding to the existing sound source through the demand forecasting module of the demand forecasting model.

[0213] As another example, referring to FIG. 7, when a second product and second product data include a new book to be published at a specific time and corresponding meta information, and a plurality of first products and a plurality of first product data include a plurality of existing books published before a specific time and corresponding meta information and sales information, the computing device (100) can determine the demand for the new book by analyzing the meta information of the new book.

[0214] First, the computing device (100) can generate a second embedding vector by encoding meta-information corresponding to the new book through the encoder of the demand forecasting model.

[0215] Subsequently, the computing device (100) can select at least one existing book among the existing books that is judged to be similar to the new book by comparing a plurality of first embedding vectors and a second embedding vector generated from meta information corresponding to a plurality of existing books through a similarity judgment module of the demand forecasting model.

[0216] Afterwards, the computing device (100) can derive the expected sales volume of the new book as the demand for the new book by analyzing sales information corresponding to the existing book through the demand forecasting module of the demand forecasting model.

[0217]

[0218] The aforementioned AI-based demand forecasting model training and the method for forecasting product demand using the same have been explained with reference to the flowchart illustrated in the drawings. For the sake of simplicity, the AI-based demand forecasting model training and the method for forecasting product demand using the same have been illustrated and described using a series of blocks; however, the present disclosure is not limited to the order of the blocks, and some blocks may be performed in a different order than that illustrated and described in this specification or simultaneously. Furthermore, new blocks not described in this specification and drawings may be added, or some blocks may be deleted or modified.

[0219]

[0220] Although embodiments of the present disclosure have been described above with reference to the attached drawings, those skilled in the art will understand that the present disclosure may be implemented in other specific forms without altering its technical concept or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive.

Claims

1. In a method performed by a computing device, A step of training a demand forecasting model using multiple first product data for the first products as training data; and A method comprising the step of predicting the demand for the second product by analyzing the second product data for the second product using the above-mentioned learned demand forecasting model, AI-based demand forecasting model training and product demand forecasting method using the same.

2. In Paragraph 1, The step of training the above demand forecasting model is, Step to determine the loss function; and A step comprising training a demand forecasting model such that the above-determined loss function is minimized, AI-based demand forecasting model training and product demand forecasting method using the same.

3. In Paragraph 2, The loss function determined above is, Characterized by being expressed as in the following mathematical formula 1, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 1> Here, the above is the above-determined loss function, above is a first loss function corresponding to a demand forecasting module included in the above demand forecasting model, the above and above is a second loss function and a third loss function corresponding to a similarity judgment module included in the above demand forecasting model, the above is a fourth loss function corresponding to data corruption of the above similarity judgment module, the above , above and above is a pre-set weight 4. In Paragraph 3, The above first loss function is, Characterized by being expressed as in mathematical formula 2 below, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 2> Here, the above is the above first loss function, above is the actual demand value of the first products included in the plurality of first product data and the is the expected demand value of the first products derived by analyzing the plurality of first product data through the above demand forecasting model.

5. In Paragraph 3, The above second loss function is, Characterized by being expressed as in mathematical formula 3 below, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 3> Here, the above is the above second loss function, above and above is a size embedding vector generated from the plurality of first product data, the is the actual demand value of the first products included in the plurality of first product data, the is the result of row-wise shuffling the actual demand values ​​of the above-mentioned first products, the above The above As a result of shuffling row by row, the above is a function that calculates the similarity between vectors and the above is a size similarity function that calculates the size similarity between time series data.

6. In Paragraph 3, The above third loss function is, Characterized by being expressed as in the following mathematical formula 4, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 4> Here, the above is the above third loss function, above and above is a shape embedding vector generated from the plurality of first product data, the is the actual demand value of the first products included in the plurality of first product data, the is the result of row-wise shuffling the actual demand values ​​of the above-mentioned first products, the above The above As a result of shuffling row by row, the above is a function that calculates the similarity between vectors and the above is a shape similarity function that calculates the shape similarity between time series data.

7. In Paragraph 3, The above fourth loss function is, Characterized by being expressed as in mathematical formula 5 below, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 5> Here, the above is the above-mentioned fourth loss function, above , above , above and above is an embedding vector generated from the plurality of first product data above, the , above , above and above The above , above , above and above The result of row-wise shuffling and the above is a function that calculates the similarity between vectors.

8. In Paragraph 2, The step of training a demand forecasting model so that the above-determined loss function is minimized is: A step of generating a plurality of first embedding vectors by encoding the plurality of first product data above; A step of generating a plurality of training data including a plurality of positive samples including mutually similar embedding vector pairs and a plurality of negative samples including mutually dissimilar embedding vector pairs using the plurality of first embedding vectors generated above; and A step comprising performing contrastive learning using the plurality of training data generated above so that a similarity determination module included in the demand forecasting model learns similarity between embedding vectors, AI-based demand forecasting model training and product demand forecasting method using the same.

9. In Paragraph 8, The plurality of first embedding vectors generated above are, It includes a plurality of first image embedding vectors generated by encoding a plurality of image data included in the plurality of first product data and a plurality of first text embedding vectors generated by encoding a plurality of text data included in the plurality of first product data, The above similarity judgment module is, It includes a size similarity determination module for determining size similarity between embedding vectors and a shape similarity determination module for determining shape similarity between embedding vectors, The step of generating the above plurality of training data is, A step of generating a plurality of first image attention vectors by attention pooling the plurality of first image embedding vectors; A step of generating a plurality of first text attention vectors by attention pooling the plurality of first text embedding vectors; A step of selecting a first size embedding vector among the plurality of first image attention vectors generated above and the plurality of first text attention vectors generated above, which reflects a characteristic regarding the size of the embedding vector, and generating first training data using the selected first size embedding vector; and The method includes the step of selecting a first shape embedding vector among the plurality of first image attention vectors and the plurality of first text attention vectors generated above, wherein the first shape embedding vector reflects characteristics regarding the shape of the embedding vector, and generating second training data using the selected first shape embedding vector. The step of performing the above contrast learning is, A method comprising the step of performing contrast learning for the size similarity determination module using the first training data generated above, and performing contrast learning for the shape similarity determination module using the second training data generated above. AI-based demand forecasting model training and product demand forecasting method using the same.

10. In Paragraph 9, The plurality of first image attention vectors generated above are, Characterized by being calculated using the following mathematical formula 6, AI-based demand forecasting model training and product demand forecasting method using the same. It may be calculated using Food Formula 6. <Mathematical Formula 6> Here, the above is the first image attention vector, the above is the first query matrix, the above is the transpose matrix of the first key matrix, the above is the size of the dimension and the above is the first value matrix.

11. In Paragraph 9, The plurality of first text attention vectors generated above are, Characterized by being calculated using the following mathematical formula 7, AI-based demand forecasting model training and product demand forecasting method using the same. <Mathematical Formula 7> Here, the above is the first text attention vector, above is the second query matrix, the above is the transpose matrix of the second key matrix, the above is the size of the dimension and the above is the value matrix.

12. In Paragraph 1, The step of training the above demand forecasting model is, A step of generating a plurality of training data using the above plurality of first product data; A step of classifying the plurality of training data generated above according to a plurality of Key Performance Indicators (KPIs) related to demand; and A method comprising the step of generating a demand forecasting model that derives at least one of the plurality of key performance indicators as result data by training the demand forecasting model using the plurality of classified training data, using specific product data as input data. AI-based demand forecasting model training and product demand forecasting method using the same.

13. In Paragraph 1, The step of forecasting the above demand is, A step of generating a second embedding vector by encoding the second product data; A step of selecting at least one first product data among the plurality of first product data based on the second embedding vector generated above; and A step comprising predicting the demand for the second product using at least one selected first product data, AI-based demand forecasting model training and product demand forecasting method using the same.

14. In Paragraph 13, The above second product data is, It includes image data and text data containing information regarding the second product mentioned above, The step of generating the second embedding vector above is, A step of generating a second image embedding vector by encoding the above image data; and A method comprising the step of generating a second text embedding vector by encoding the above text data, AI-based demand forecasting model training and product demand forecasting method using the same.

15. In Paragraph 13, The step of generating the second embedding vector above is, A step of generating a second image attention vector by attention pooling a second image embedding vector generated by encoding image data included in the second product data; and The method includes the step of generating a second text attention vector by attention pooling a second text embedding vector generated by encoding text data included in the second product data. The step of selecting at least one first product data above is, A step of selecting a second size embedding vector that reflects characteristics regarding the size of the embedding vector among the second image attention vector and the second text attention vector generated above, and calculating the size similarity between the selected second size embedding vector and a plurality of first size embedding vectors—where the plurality of first size embedding vectors are generated from the plurality of first product data—through a size similarity judgment module included in the demand forecasting model; A step of selecting a second shape embedding vector among the generated second image attention vector and the generated second text attention vector that reflects characteristics regarding the shape of the embedding vector, and calculating the shape similarity between the selected second shape embedding vector and a plurality of first shape embedding vectors—where the plurality of first size embedding vectors are generated from the plurality of first product data—through a shape similarity judgment module included in the demand forecasting model; and A method comprising the step of selecting at least one first product data among the plurality of first product data based on the size similarity and shape similarity calculated above. AI-based demand forecasting model training and product demand forecasting method using the same.

16. In Paragraph 13, The step of predicting the demand for the second product mentioned above is, A method comprising the step of predicting the demand for the second product by analyzing at least one selected first product data through a deep learning-based demand prediction module included in the demand prediction model. AI-based demand forecasting model training and product demand forecasting method using the same.

17. In Paragraph 1, The above second product data is, Includes metadata corresponding to new music to be released at a specific time, and The above plurality of first product data are, It includes meta information and sales information corresponding to multiple existing sound sources released prior to the aforementioned specific point in time, and The step of forecasting the above demand is, A step of generating a second embedding vector by encoding meta information corresponding to the above-mentioned new sound source; A step of selecting at least one existing sound source among the plurality of existing sound sources by comparing a plurality of first embedding vectors generated from meta information corresponding to the plurality of existing sound sources with the generated second embedding vector; and The method comprises the step of calculating the expected ranking, expected sales volume, and expected revenue of the new sound source as expected demand using sales information corresponding to at least one existing sound source selected above. AI-based demand forecasting model training and product demand forecasting method using the same.

18. In Paragraph 1, The above second product data is, Includes meta-information corresponding to new books to be published at a specific point in time, and The above plurality of first product data are, It includes meta-information and sales information corresponding to a plurality of existing books published prior to the aforementioned specific point in time, and The step of forecasting the above demand is, A step of generating a second embedding vector by encoding meta-information corresponding to the above-mentioned new book; A step of selecting at least one existing book among the plurality of existing books by comparing a plurality of first embedding vectors generated from meta information corresponding to the plurality of existing books and the generated second embedding vector; and A step comprising calculating the expected sales volume of the new book as expected demand using sales information corresponding to at least one existing book selected above, AI-based demand forecasting model training and product demand forecasting method using the same.

19. Processor; Network interface; Memory; and It includes a computer program that is loaded into the memory and executed by the processor, The above computer program is, Instructions for training a demand forecasting model using multiple first product data for first products as training data; and Instructions for predicting the demand for the second product by analyzing the second product data for the second product using the above-mentioned learned demand forecasting model, A computing device that performs artificial intelligence-based demand forecasting model training and a method for forecasting product demand using the same.

20. Combined with a computing device, A step of training a demand forecasting model using multiple first product data for the first products as training data; and A computer program stored on a recording medium readable by a computing device for executing an artificial intelligence-based demand forecasting model learning and a method for forecasting the demand of a product using the same, comprising the step of forecasting the demand of the second product by analyzing second product data for the second product using the above-mentioned learned demand forecasting model.

Images