Method and apparatus for determining recommended content item based on sequential consumption history of content items

By calculating recommendation weights based on repeated consumption using an attention mechanism in a deep learning model, the method addresses the challenge of accurately recommending content items, enhancing personalization and reducing computational requirements.

KR103044652B1Active Publication Date: 2026-11-06NAVER WEBTOON LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
KR1020230167917
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2023-11-28
Publication Date
2026-11-06
Estimated Expiration
2043-11-28

AI Technical Summary

Technical Problem

Existing methods struggle to provide personalized content recommendations that accurately reflect a user's preferences when their consumption history includes repeated consumption of the same content item, especially when using deep learning-based models.

Method used

A method and apparatus that determine recommended content items by calculating a recommendation weight based on the number of repeated consumptions of each content item, using an attention mechanism within a deep learning model to account for repeated consumption histories.

Benefits of technology

This approach allows for more accurate content recommendations that reflect a user's preferences, reducing computational load while effectively handling repeated consumption, thereby improving personalized content suggestions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure R1020230167917_ABST
    Figure R1020230167917_ABST
Patent Text Reader

Abstract

A method for determining recommended content items among content items provided through a content service is provided. The method for determining recommended content items includes the step of obtaining a user's sequential consumption history of content items including a repeated consumption history for at least one content item, and the step of determining at least one recommended content item among the content items by using the obtained consumption history as input to a pre-built recommendation model.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present disclosure relates to a method and apparatus for determining recommended content items based on a sequential consumption history of content items including a repeated consumption history for at least one content item, and relates to a method and apparatus for determining recommended content items by calculating a recommendation weight that reflects a weight reflecting the repeated consumption of content items. Background Technology

[0002] For online shopping malls selling a wide variety of products or content provision services offering a wide range of content items (e.g., music, movies, webtoons, web novels, etc.), it is important to recommend products or content items that match the user's preferences. Such recommendations are made using the user's history of product consumption or content consumption. Recently, there have been cases where such personalized recommendations are provided to users using artificial intelligence models, such as deep learning.

[0003] However, when a user's product consumption history or content item consumption history includes repeated consumption of the same product or content item, it is difficult to provide personalized recommendations to the user by appropriately reflecting such repeated consumption history.

[0004] When content items are frequently consumed repeatedly, it is difficult to provide recommendations that reflect the user's appropriate long-term preferences when recommendations are provided based on consumption history over a certain period. Furthermore, even when using deep learning-based models, problems may arise where recommendation results fail to generalize or where recommendation results do not properly reflect the user's preferences, depending on the complexity of the model and a lack of information.

[0005] Meanwhile, Korean Patent Publication No. 10-2023-0054946 (published on April 25, 2023) discloses a system for providing content recommendation adjustment services using reverse engineering and machine learning.

[0006] The information described above is for illustrative purposes only and may include content that does not constitute part of the prior art and may not include what the prior art would present to a person skilled in the art. The problem to be solved

[0007] One embodiment may provide a method for determining a content item, comprising the steps of: obtaining a user’s sequential consumption history of content items including a repeated consumption history for at least one content item; and determining at least one recommended content item among the content items by using the consumption history obtained as an input to a pre-built recommendation model.

[0008] One embodiment may provide a method for determining recommended content items by calculating a recommendation weight applied based on the number of repeated consumptions of each content item for each content item of a first number of content items included in a consumption history, and determining recommended content items based on the calculated recommendation weight. means of solving the problem

[0009] In one aspect, a method for determining a recommended content item among content items provided through a content service, performed by a computer system, comprises: a step of obtaining a user’s sequential consumption history for the content items, wherein the consumption history includes a consumption history for a first number of content items among the content items, and includes at least two repeated consumption histories for at least one content item among the first number of content items; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-built recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied with a weight based on the number of repeated consumptions of each content item for each of the first number of content items; and a step of determining the recommended content item among the content items based on the recommendation weight calculated for each content item.

[0010] Each of the above content items includes multiple rounds or multiple episodes, and in the consumption history, consumption of at least two rounds or episodes of each of the above content items can be treated as a repeated consumption history for each of the above content items.

[0011] The above recommendation model includes an attention mechanism, and the recommendation weight calculated for each content item can be calculated based on the value obtained by applying the weight to the attention score for each content item calculated through the attention mechanism.

[0012] The step of calculating the recommendation weight may include: inputting each content item into the attention mechanism to calculate the attention score for each content item; inputting the weight based on the number of repeated consumptions of each content item into the attention mechanism to apply the weight to the calculated attention score; and calculating the recommendation weight based on the attention score to which the weight has been applied.

[0013] The above-mentioned weighted attention score can be calculated by applying the above-mentioned weighted softmax function to the above-mentioned attention score.

[0014] The attention score with the above weights applied may be the attention weight for each of the above content items.

[0015] The sum of the recommendation weights calculated for each of the above-mentioned first number of content items can be 1.

[0016] The above attention mechanism may be configured to calculate the recommendation weight using an activation function for calculating the attention weight for each content item, which reflects the weight based on the number of times each content item is repeatedly consumed.

[0017] The above activation function is expressed by mathematical formula 1, and

[0018] [Mathematical Formula 1]

[0019]

[0020] The above h i is the distribution of attention weights of each content item for determining the recommended content items, j is a value from 1 to the first number, and S j is the attention score for the j-th content item included in the above consumption history, and the above w j is the weight for the j-th content item, and the above v jcan be the content of the j-th content item.

[0021] The above S j is expressed by mathematical formula 2, and

[0022] [Mathematical Formula 2]

[0023]

[0024] q i represents the query vector corresponding to the above recommended content item, and k j represents a key vector corresponding to the j-th content item included in the above consumption history, and c may represent a value obtained by taking the square root of the dimension of the key vector.

[0025] The weight based on the number of times each of the above-mentioned content items is consumed repeatedly may have a value proportional to the number of times each of the above-mentioned content items are consumed repeatedly.

[0026] The method for determining the recommended content items further includes the step of determining the first number, wherein the first number is the number of input tokens that are input to the recommendation model, and the determining step may determine the first number with a larger value than when the user is requested to determine the recommended content items reflecting the user's short-term preferences when the user is requested to determine the recommended content items reflecting the user's long-term preferences.

[0027] The above recommendation model may be a deep learning model including a transformer in which a plurality of the above attention mechanisms are stacked.

[0028] The above recommendation weight may represent the weight of at least one of the genre, content, and attribute of each content item for determining the recommended content item.

[0029] The recommendation weight mentioned above represents the weight of how much information of each content item included in the consumption history needs to be retrieved to determine the recommended content item, and the recommendation model may be pre-trained to determine the recommendation weight by using the recommended content item as a query and each content item as a key.

[0030] The above recommendation model may be configured to determine the same recommended content item even if the consumption order of the content items included in the consumption history is different, when the content items included in the consumption history and the number of repeated consumptions of each content item are the same.

[0031] In another aspect, a computer system for determining recommended content items among content items provided through a content service comprises at least one processor implemented to execute a readable command in the computer system, wherein the at least one processor obtains a user’s sequential consumption history of the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—and determines at least one recommended content item among the content items using the consumption history obtained as input to a pre-built recommendation model, wherein for each content item of the first number of content items, a recommendation weight is applied with a weight based on the number of repeated consumptions of each content item, and the recommended content item among the content items is determined based on the recommendation weight calculated for each content item. Effects of the invention

[0032] By determining recommended content items for a user using the user's sequential consumption history of content items that includes a history of repeated consumption of at least one content item, recommendations that more accurately reflect the user's preferences can be provided to the user.

[0033] For each content item of the first number of content items included in the consumption history, a recommendation weight is calculated by applying a weight based on the number of repeated consumptions of each content item, and by determining the recommended content item based on the calculated recommendation weight, the recommendation result is not distorted even if there is repeated consumption of the content item, and the user's taste can be accurately reflected.

[0034] When using a recommendation model that utilizes an attention mechanism, it is possible to reduce computations based on the consumption history of input content items while calculating an attention weight (or attention distribution) that reflects the number of repeated consumptions for each content item.

[0035] The performance of personalized recommendations for serialized content items that are frequently consumed repeatedly can be improved. Brief explanation of the drawing

[0036] FIG. 1 illustrates a method for determining recommended content items based on a sequential consumption history of content items by a user, according to one embodiment. FIG. 2 is a block diagram showing a computer system for determining recommended content items and a user terminal for receiving recommended content items, according to one embodiment. FIG. 3 is a flowchart illustrating a method for determining recommended content items based on a sequential consumption history of content items by a user, according to one embodiment. Figure 4 is a flowchart illustrating a method for calculating the recommendation weight for each content item used to determine recommended content items according to one example. Figure 5 illustrates a method for determining recommended content items based on a sequential consumption history of content items composed of serialized content items according to one example. Figure 6 shows a recommendation model including an attention mechanism that calculates the attention weight for each content item used to determine recommended content items. Figure 7 shows a recommendation model including an attention mechanism that calculates an attention weight for each content item used to determine recommended content items by considering weights representing repeated consumption of content items according to one example. FIGS. 8 to 10 illustrate a method for calculating attention weight by considering weights representing repeated consumption of content items according to one example. Figure 11 shows a schematic diagram of a personalized recommendation system according to one example. Specific details for implementing the invention

[0037] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0039] FIG. 1 illustrates a method for determining recommended content items based on a sequential consumption history of content items by a user, according to one embodiment.

[0040] Referring to FIG. 1, a method is described in which a computer system (100) acquires a consumption history (60) of a user’s content items among a plurality of content items provided through a content service, and determines a recommended content item (50) to be provided to a user terminal (50) based on the acquired consumption history (60).

[0041] The content service may be a service that provides requested content items to a user terminal (50) in response to a request from the user terminal (50). The content service may be linked with a platform that manages and provides content to provide requested content items to the user terminal (50), and may enable a user account associated with the user terminal (50) to use the content items. The user may use the content items provided by the content service through a user account associated with the user terminal (50) (e.g., an account (ID) of a user registered or subscribed to such content service). In the description to be made later, 'consumption' of content items may refer to the use, viewing, playback, or viewing of such content items.

[0042] Content services may be provided to a user terminal (50) (user) through a server, or may be provided through an illustrated computer system (100). In this case, the computer system (100) may include a content platform that manages and provides content, or may be configured to communicate with such a content platform. The computer system (100) may include at least one server.

[0043] "Content items" provided through a content service are digital goods and may include multimedia content and text content. For example, they may include content or service video content, movie content, music content, webtoon content, broadcast program content, book content (web novel content), etc. "Content items" may represent individual content. For example, "Content items" may represent a webtoon or web novel containing at least one episode. Alternatively, such content items may be serialized and may include multiple episodes or multiple installments.

[0044] Content services may be provided to the user terminal through a dedicated application installed on the user terminal (50). This dedicated application may be an application installed on the user terminal (50) to provide content items to the user account through a content platform that includes a DB storing content items or communicates with said DB. For example, this dedicated application may be a webtoon or web novel providing application.

[0045] In the embodiment, information regarding the content item (20) viewed by the user terminal (50) may be managed as a consumption history (60) for the content item. The consumption history (60) may represent a sequential consumption history for the content items. For example, the consumption history (60) in the illustrated example may indicate that the user terminal (50) consumed the content items in the order of content item A, content item A, content item B, content item C, and content item C. Here, the consumption history (60) may include a repeated consumption history for content item A and content item C. That is to say, the consumption history (60) may include consuming content item A twice, consuming content item B once, and consuming content item C twice. The consumption history (60) may be a selected part of the user's total consumption history for the content items.

[0046] A computer system (100) can determine recommended content items (30) that suit the user's taste based on the user's consumption history (60). At this time, the computer system (100) can determine recommended content items (30) that suit the user's taste by considering the repeated consumption history for each content item included in the consumption history (60). The computer system (100) can determine recommended content items (30) by giving greater weight to content items that have been consumed more repeatedly within the consumption history (60). For example, the recommended content items (30) may be more similar to content items A and C that include two consumptions compared to content item B that includes one consumption.

[0047] Thus, in the embodiment, when determining the recommended content item (30), the user's repeated consumption history for each content item is taken into account, thereby allowing the content item that is more accurate to the user's taste to be determined as the recommended content item (30).

[0048] The computer system (100) can determine the weight (hereinafter, recommendation weight) used in determining the recommended content item (30) for each content item included in the consumption history (60), and can determine the recommended content item (30) based on this recommendation weight. The computer system (100) may use a pre-built recommendation model to determine the recommendation weight and the recommended content item (30). The recommendation model may be pre-trained to take the consumption history (60) as input and to determine (or calculate) the recommendation weight for each content item included in the consumption history (60) in order to determine the recommended content item (30).

[0049] A more specific method for a computer system (100) to determine recommendation weights using a recommendation model and to determine recommendation content items (30) is described in more detail with reference to FIGS. 2 to 11, which will be described later.

[0051] FIG. 2 is a block diagram showing a computer system for determining recommended content items and a user terminal for receiving recommended content items, according to one embodiment.

[0052] The computer system (100) may be a server on which the aforementioned recommendation model is built. The computer system (100) may acquire a certain range of consumption history to be used to determine recommended content items (30) from the consumption history of the user's content items, and may determine recommended content items (30) for the user by using the said consumption history as input to the recommendation model. The computer system (100) may provide the determined recommended content items (30) to the user terminal (50).

[0053] Meanwhile, the computer system (100) may also serve as a server providing the aforementioned content service, and, depending on the embodiment, may include a content platform providing the content service or be configured to communicate with such a content platform.

[0054] The computer system (100) may be configured to include at least one computing device, and such computing device may be implemented as at least one server.

[0055] As illustrated, the computer system (100) may include memory (130), a processor (120), a communication unit (110), and an input / output interface (140).

[0056] Memory (130) is a computer-readable recording medium and may include a non-perishable mass storage device such as RAM (random access memory), ROM (read only memory), and a disk drive. Here, the ROM and the non-perishable mass storage device may be included as separate permanent storage devices separated from the memory (130). Additionally, an operating system and at least one program code may be stored in the memory (130). These software components may be loaded from a computer-readable recording medium separate from the memory (130). This separate computer-readable recording medium may include a computer-readable recording medium such as a floppy drive, disk, tape, DVD / CD-ROM drive, or memory card. In another embodiment, the software components may be loaded into the memory (130) through a communication unit (110) rather than a computer-readable recording medium.

[0057] The processor (120) may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input / output operations. Instructions may be provided to the processor (120) by memory (130) or a communication unit (110). For example, the processor (120) may be configured to execute instructions received according to program code loaded in memory (130).

[0058] The communication unit (110) may be a configuration for the computer system (100) to communicate with another device (such as a user terminal (50) or another server). That is to say, the communication unit (110) may be a hardware module such as an antenna, data bus, network interface card, network interface chip, and networking interface port of the computer system (100) that transmits / receives data and / or information to / from another device, or a software module such as a network device driver or a networking program.

[0059] The input / output interface (140) may be a means for interfacing with an input device such as a keyboard or mouse and an output device such as a display or speaker.

[0060] The processor (120) can manage components of the computer system (100), execute a program or application for providing the aforementioned content service, and process operations necessary for the execution of said program or application and processing of data, etc. The processor (120) may be at least one processor (CPU and / or GPU, etc.) of the computer system (100) or at least one core within the processor.

[0061] Additionally, in the embodiments, the computer system (100) and the processor (120) may include more components than the illustrated components.

[0062] The user terminal (50) is a terminal possessed and used by a user, and may be an electronic device that executes content services provided by a computer system (100), etc., or has a dedicated application for content services installed and executes content services through said dedicated application. For convenience of explanation, the dedicated application may be referred to as an application. The user terminal (50) may be implemented as a computer system. The user terminal (50) may be configured to transmit a request for a content item (e.g., a purchase request or a viewing request) to the computer system (100) through a user account registered for the content service, and to consume the requested content item. Depending on the consumption of the content item, the consumption history indicating the consumption may be transmitted to and managed by the computer system (100) or a content platform providing the content service. The user terminal (50) may receive recommended content items (30) determined by the computer system (100). The recommended content items (30) may be displayed on the user terminal (50) and consumed by the user.

[0063] The user terminal (50) is a computing device and may include, for example, a PC, a laptop, a smartphone, a tablet, an Internet of Things device, or a wearable computer.

[0064] As illustrated, the user terminal (50) may include a memory (135), a processor (125), a communication unit (115), and an input / output interface (145). Since the description of the memory (130), processor (120), communication unit (110), and input / output interface (140) of the aforementioned computer system (100) can be similarly applied to the memory (135), processor (125), communication unit (115), and input / output interface (145), redundant descriptions are omitted.

[0065] The processor (125) can manage the components of the user terminal (50), execute the aforementioned content service, transmit a request for a content item to the computer system (100), and be configured to output the requested content item. Additionally, the processor (125) can execute the aforementioned dedicated application (or program) and process operations necessary for the execution of the dedicated application and the processing of data. The processor (125) may be at least one processor (CPU and / or GPU, etc.) of the user terminal (50) or at least one core within the processor.

[0066] The user terminal (50) may also be referred to as a computer system, but for convenience of explanation, the user terminal (50) and the computer system (100) will be described separately in the detailed description to be described later.

[0067] A more specific method for obtaining a consumption history (60) of content items according to the operation of a computer system (100), determining recommended content items (30) based thereon, and providing recommended content items (30) to a user terminal (50) based on the determined information is described in more detail with reference to FIGS. 3 to 11, which will be described later.

[0068] The description of the technical features described above with reference to FIG. 1 can be applied as is to FIG. 2, so redundant descriptions are omitted.

[0070] In the detailed description to be provided below, the operation performed by the configurations of the computer system (100) or processor (120) or the configurations of the user terminal (50) or processor (125) may be described as an operation performed by the computer system (100) or user terminal (50) for convenience of explanation.

[0072] FIG. 3 is a flowchart illustrating a method for determining recommended content items based on a sequential consumption history of content items by a user, according to one embodiment.

[0073] Referring to FIG. 3, a method is described in which a computer system (100) determines recommended content items for a user among content items provided through a content service using a recommendation model.

[0074] In step (310), the computer system (100) may acquire a user's sequential consumption history (60) for content items provided through a content service. The consumption history (60) may be a selected part of the user's total consumption history for content items. The acquired consumption history (60) may include a consumption history for a first number of content items among the content items provided through the content service (or, among them, content items consumed by the user). At this time, the consumption history (60) may be configured to include at least two repeated consumption histories for at least one content item among the first number of content items. That is to say, as described above with reference to FIG. 1, the consumption history (60) may include a repeated consumption history for at least one content item (e.g., content items A and C in FIG. 1).

[0075] As described above, each content item may include multiple installments or multiple episodes as a serial type. In the consumption history (60), consumption of at least two installments or episodes of each content item may be treated as a repeated consumption history for each content item. That is to say, in the embodiment, if multiple installments (episodes) of the same content item are consumed, it may be considered that the same content item has been consumed multiple times rather than that different content items have been consumed, and this may be included in the consumption history (60) as a repeated consumption history. For example, the two repeated consumption histories of content A shown in FIG. 1 may indicate that different installments of content A have been consumed.

[0076] Meanwhile, as in step (305), the computer system (100) may first determine a first number to determine the range of consumption history (60) to be used for determining content items (30). The first number may be a value set by the computer system (100) or its manager. For example, if the computer system (100) or its manager is requested to determine recommended content items (30) by reflecting the user's long-term preferences, the first number may be determined as a larger value than if the user is requested to determine recommended content items (30) by reflecting the user's short-term preferences.

[0077] The first number determined may represent the number of input tokens that serve as input to the recommendation model. Each input token may correspond to each content item. Accordingly, the range of the consumption history (60) may be determined to include the first number of input tokens (i.e., content items). For example, if the first number is 3, the range of the consumption history (60) may be determined to include content item A, content item B, and content item C as illustrated in FIG. 1.

[0078] In step (320), the computer system (100) can determine at least one recommended content item (30) among the content items provided through the content service by using the consumption history (60) obtained as input to a pre-built recommendation model.

[0079] The computer system (100) can determine recommended content items (30) that match the user's taste by considering the repeated consumption history for each content item included in the consumption history (60). Therefore, in the embodiment, recommendations for content items that more accurately reflect the user's taste can be made.

[0080] To this end, in step (322), the computer system (100) can calculate a recommendation weight applied to each content item of the first number of content items included in the consumption history (60), with a weight based on the number of times each content item is repeatedly consumed. The computer system (100) can calculate a weight associated with each content item based on the number of times the user repeatedly consumes each content item. For example, the weight can be determined by the number of times the content item is repeatedly consumed. The weight determined based on the number of times each content item is repeatedly consumed may have a value proportional to the number of times each content item is repeatedly consumed. That is to say, the weight may be determined to be larger for content items that have been repeatedly consumed more within the consumption history (60). The computer system (100) can calculate a recommendation weight that reflects this determined weight.

[0081] In step (324), the computer system (100) can determine the recommended content item (30) based on the recommendation weight calculated for each content item of the first number of content items.

[0082] The computer system (100) can, based on the calculated recommendation weight, allow content items included in the consumption history (60) that have a higher recommendation weight to be given more consideration in determining the recommended content item (30). That is to say, a content item that is more similar to the content item with a higher recommendation weight can be determined as the recommended content item (30).

[0083] For example, the calculated recommendation weight may represent the weight of at least one of the genre, content, and attribute (or type (e.g., whether it is a novel or a webtoon, etc.)) of each content item for determining the recommended content item. In other words, depending on the recommendation weight, the similarity of genre, similarity of content, and similarity of attribute with the content item with the higher recommendation weight may be further considered when the computer system (100) determines the recommended content item (30).

[0084] The configuration of the recommendation model of the embodiment and a more specific method for determining the recommended content item (30) by calculating the recommendation weight will be explained in more detail with reference to FIGS. 4 to 11, which will be described later.

[0085] The description of the technical features described above with reference to FIGS. 1 and FIGS. 2 can be applied as is to FIGS. 3, so redundant descriptions are omitted.

[0087] Figure 4 is a flowchart illustrating a method for calculating the recommendation weight for each content item used to determine recommended content items according to one example.

[0088] Referring to FIG. 3, the aforementioned recommendation model may be an artificial intelligence (e.g., deep learning)-based recommendation model. The recommendation model may be pre-trained to determine (or calculate) the recommendation weight for each content item included in the consumption history (60) in order to determine the recommended content item (30) using the consumption history (60) as input, and may include such a pre-trained model. For example, the recommendation model may be pre-trained to determine the recommendation weight for each content item included in the consumption history (60) in order to determine the recommended content item (recommended content item (30)) to be consumed by the user next using the user's consumption history (60) as input. Such recommendation weight for each included content item may be an intermediate result for determining the content item (30).

[0089] The recommendation model of the embodiment may be a 'sequential recommendation model' that provides appropriate recommendations using a sequential consumption history (60) of content items.

[0090] The recommendation model of the embodiment may be configured to include an attention mechanism. The recommendation weight calculated for each content item in step (322) of FIG. 3 described above may be calculated based on a value obtained by applying weights to the attention score for each content item calculated through this attention mechanism.

[0091] Below, we first explain the attention mechanism.

[0092] An attention mechanism may include an algorithm that reconstructs the information of each element (or token) in sequential information containing multiple elements into the information of each other element (or each of all elements containing that element). For example, in sequential information such as the sentence [see, that, girl, run], this sequential information may include each word as an element. In this case, the attention mechanism can reconstruct each element, and, for instance, can reconstruct it so that the information of the word 'that' contains the most information of the word 'girl' among the elements. That is, the attention mechanism may refer to an algorithm that determines the word (element) to focus on according to the context in the above sequential input. The attention (or attention score) for each element (token) may indicate the importance of that element.

[0093] This attention can be calculated by the function of Equation 1 below.

[0094] [Mathematical Formula 1]

[0095]

[0096] Here, Key (K), Value (V), and Query (Q) can be three elements for calculating attention. K, V, and Q can each be a vector. Q can be the query vector, K can be the key vector, and V can be the value vector.

[0097] Assuming an example where documents are searched by entering a search term, Q can correspond to the search term, K can be the key value for searching via the term (e.g., the document title), and V can correspond to the content of the document. In other words, Q can be used as the "search term," K as the "title" (used for document searching), and V as the "body" of the document to reconstruct the respective information of the input tokens. d k can represent the dimensionality of the key vector.

[0098] In the above mathematical formula 1 can represent the similarity between a query vector and a key vector. This similarity can be an attention score. The softmax function is used to normalize the attention score, and by applying the softmax function to the attention score, the attention distribution (attention weight) can be calculated. When the attention score to which the softmax function has been applied is multiplied by a value (V), the attention value can be calculated.

[0099] The recommendation model for the content item of the embodiment may be a recommendation model based on an attention mechanism. Such a recommendation model can determine the recommended content item (30) by using information (genre, content, and attributes, etc.) that the content item itself possesses. However, the attention mechanism based on the above mathematical formula 1 may not reflect the repeated consumption of each element (i.e., content item).

[0100] In contrast, the recommendation model for the content item of the embodiment can be implemented using an attention mechanism that reflects weights based on the number of times the content item is repeatedly consumed.

[0101] In the embodiment, the aforementioned recommendation weight can be calculated using an attention mechanism that reflects weights based on the number of times these content items are repeatedly consumed. The recommendation weight can be calculated based on steps (410 to 430).

[0102] In step (410), the computer system (100) can input each content item included in the consumption history (60) into the attention mechanism to calculate an attention score for each content item.

[0103] In step (420), the computer system (100) inputs a weight based on the number of repeated consumptions of each content item included in the consumption history (60) into the attention mechanism, thereby applying the weight to the attention score calculated in step (410). The weight may be the number of repeated consumptions of the corresponding content item. Alternatively, the weight determined based on the number of repeated consumptions of each content item may be a value proportional to the number of repeated consumptions of each content item. The weighted attention score may be calculated by applying a softmax function with the weight to the attention score calculated in step (410). The 'weighted attention score' calculated accordingly may represent the attention weight for each content item. The attention weight for each content item may be the attention weight of each content item.

[0104] In step (430), the computer system (100) can calculate a recommendation weight based on a weighted attention score. The weighted attention score can be a recommendation weight. For example, the 'weighted attention score' calculated by applying the aforementioned softmax function can be a recommendation weight for each content item, as an attention weight (attention distribution) for each content item. Alternatively, the recommendation weight can be a value calculated based on the weighted attention score.

[0105] The recommendation weight (or attention weight) of each content item can be used to determine the content item (recommended content item (30)) that the user will consume next. In other words, the recommended content item (30) can be determined by considering the recommendation weight of each determined content item.

[0106] Meanwhile, depending on the application of the softmax function, the sum of the recommendation weights calculated for each of the first number of content items included in the consumption history (60) can be 1.

[0107] The activation function for calculating the attention weight of the attention mechanism included in the recommendation model of the example can be defined as Equation 2 below.

[0108] [Mathematical Formula 2]

[0109]

[0110] In other words, the attention mechanism of the embodiment may use the activation function of Equation 2 to calculate the attention weight (attention distribution) for each content item, which reflects a weight based on the number of repeated consumptions of each content item included in the consumption history (60), and thereby the recommendation weight for each content item can be calculated.

[0111] In the above mathematical formula 2, h i may be a distribution of the attention weight of each content item included in the consumption history (60) for determining the recommended content item (30). i may be the number of the token corresponding to the recommended content item (60). j may be a value from 1 to the first number (i.e., the number of tokens). i may be j+1. That is, h i may be a distribution of the attention weight of each content item included in the consumption history (60) for determining the recommended content item (30) that the user will consume next after the consumption history (60). S i can be the attention score for the j-th content item included in the consumption history. w j can be the weight for the j-th content item. w j can represent the number of times the j-th content item is consumed repeatedly. The above v j can be the content (value) of the j-th content item.

[0112] Here, Sj is It can be defined as follows. q ican represent a vector of content items corresponding to token i, and k can represent a key vector corresponding to the j-th content item. The key vector can represent the characteristics of the content items. In other words, q i can represent a query vector corresponding to a recommended content item (30) as a query vector, and k j can represent a key vector corresponding to the j-th content item included in the above consumption history (60). c can be the square root of the dimension of the key vector. Sj can represent the similarity between the query vector and the key vector. By applying this to the above mathematical formula 2 and re-expressing the activation function, it can be as shown in mathematical formula 3.

[0113] [Mathematical Formula 3]

[0114]

[0115] Meanwhile, for comparison purposes, the activation function of an attention mechanism that does not reflect weights based on the number of repeated consumptions of each content item included in the consumption history (60) can be expressed as Equation 4 below.

[0116] [Mathematical Formula 4]

[0117]

[0118] As described above, the result of calculating the activation function of the above mathematical formula 3 may represent the distribution of the attention weight of each content item included in the consumption history (60) for determining the recommended content item (30). Alternatively, this may be expressed as the total sum of the attention weights.

[0119] For example, in the case where the i-th token corresponds to the content item (recommended content item (30)) that the user will consume next, h iThis can represent the weight of how much information of each content item included in the consumption history (60) is retrieved when the i-th content item (recommended content item (30)) is used as a query. In other words, the recommendation model of the embodiment may be pre-trained to calculate the attention weight (i.e., recommendation weight) of each content item, which is an intermediate result for determining the i-th content item corresponding to the query, in order to determine the i-th content item. The recommendation weight may represent the weight of how much information of each content item included in the consumption history (60) is retrieved to determine the recommended content item. The recommendation model of the embodiment may be pre-trained to determine the recommendation weight by using the recommended content item (30) as a query (vector) and each content item as a key (vector).

[0120] The recommendation model of the embodiment may be a deep learning model comprising a transformer in which a plurality of attention mechanisms using the activation functions of the aforementioned mathematical formulas 2 and 3 are stacked.

[0121] The description of the technical features described above with reference to FIGS. 1 to 3 can be applied as is to FIG. 4, so redundant descriptions are omitted.

[0123] Figure 5 illustrates a method for determining recommended content items based on a sequential consumption history of content items composed of serialized content items according to one example.

[0124] Figure 5 conceptually illustrates the determination of recommended content items (30) by the recommendation model of the embodiment, although it is not identical to the recommendation model of the embodiment.

[0125] In the consumption history (500) for the illustrated content items, repeated consumption of the same content item was included. That is, as illustrated, content item A was consumed 4 times (i.e., episodes 1 through 4 were consumed), and content items B and C were consumed 2 times (i.e., episodes 1 and 2 were consumed).

[0126] Through the recommendation model of the embodiment, as an intermediate result for determining D, which is a recommended content item (510), the extent to which each of content items A, B, and C is considered in the recommendation (i.e., recommendation weight) can be calculated based on a weight that considers the number of times each of content items A, B, and C is repeatedly consumed. The recommended content item (510) can be determined by considering this calculated recommendation weight. As illustrated, for example, in the case where the genre of content item A is X, the genre of content B is Y, and the genre of content C is Z, content item D having genre W can be determined as content item (510) based on the recommendation weight calculated for each content item.

[0127] The description of the technical features described above with reference to FIGS. 1 to 4 can be applied as is to FIG. 5, so redundant descriptions are omitted.

[0129] Figure 6 shows a recommendation model including an attention mechanism that calculates the attention weight for each content item used to determine recommended content items.

[0130] In FIG. 6, a recommendation model (600) is shown that calculates a recommendation weight (620) based on a consumption history (610) and determines a recommended content item (630) using an attention mechanism that does not reflect a weight based on the number of repeated consumptions of each content item, as described above with reference to mathematical formula 4.

[0131] The consumption history (610) may include three tokens (content items C, C, A). Since the recommendation model (600) only utilizes the consumption history (610) of the three corresponding content items for recommendation, it cannot perform recommendations that take into account the user's long-term preferences. Additionally, since it does not take into account the user's repeated consumption of the same content item, appropriate recommendations that take this into account cannot be made. In the illustrated example, in order for appropriate recommendations to be provided to the user, a consumption history (610) including eight tokens may need to be input into the recommendation model (600), which can significantly increase the computational load of the recommendation model (600).

[0132] Meanwhile, FIG. 7 shows a recommendation model including an attention mechanism that calculates an attention weight for each content item used to determine recommended content items by considering weights representing repeated consumption of content items according to one example.

[0133] In FIG. 7, a recommendation model (700) is illustrated that calculates a recommendation weight (720) based on a consumption history (710) and determines a recommended content item (730) by using an attention mechanism that reflects a weight based on the number of repeated consumptions of each content item as described above with reference to mathematical formulas 2 and 3.

[0134] The consumption history (710) may include three tokens (content items A, B, and C). That is, the recommendation model (700) can receive a weight w corresponding to the number of times each content item is repeatedly consumed, taking into account the number of times each content item is repeatedly consumed, and can treat the same content item with a consumption history as a single token. Therefore, the recommendation model (700) can effectively perform recommendations that consider the user's long-term preferences with less computational power than the recommendation model (300). In other words, unlike the recommendation model (600), the consumption history (710) including consumption of eight content items is received as input with three tokens and the weight of each token, and the recommendation weight (720) is determined and the recommended content item (730) is determined.

[0135] In other words, the recommendation model (700) can realize the determination of recommended content items (730) that reflect the user's long-term preferences and repeated consumption with less computational power than the recommendation model (600) by using an attention mechanism that is weighted by the number of times each content item is repeatedly consumed. Therefore, the recommendation model (700) is a transformer-based recommendation model and can efficiently condense computations for content items that are repeatedly consumed.

[0136] Meanwhile, the recommendation model (700) can be configured to determine the same recommended content item (730) even if the consumption order of the content items included in the consumption history (710) is different (i.e., regardless of the consumption order, independently of the consumption order), when the content items included in the consumption history (710) and the number of repeated consumptions of each content item are the same. In other words, when determining the recommended content item (730), the recommendation model (700) can ignore the difference in the amount of information according to the consumption order of the content items included in the consumption history (710), or the sequential significance of the content items included in the consumption history (710).

[0137] The description of the technical features described above with reference to FIGS. 1 to 5 can be applied as is to FIGS. 6 and 7, so redundant descriptions are omitted.

[0139] FIGS. 8 to 10 illustrate a method for calculating attention weight by considering weights representing repeated consumption of content items according to one example.

[0140] Referring to FIGS. 8 to 10, a method for deriving an attention mechanism having the activation functions of the aforementioned mathematical formulas 2 and 3 is explained.

[0141] In FIGS. 8 to 10, a method for determining what a masked content item (i.e., content item M) is is illustrated in a consumption history including four repeated consumptions of content item A (episodes 1 to 4), two repeated consumptions of content item B (episodes 1 and 2), and consumption of a masked content item. The determination of the masked content item can correspond to the determination of the aforementioned recommended content item (730).

[0142] Figure 10 is a mathematical representation of the operations schematically explained in Figures 8 and 9.

[0143] In FIG. 10, K can represent a key vector corresponding to each content item, and S can represent an attention score. A can represent the value obtained by applying a softmax function to the attention score. As a result of the calculation in FIG. 10, it can be confirmed that an activation function similar in form to the aforementioned mathematical equations 2 and 3 is derived. Z m is h of mathematical formulas 2 and 3 i It can respond to.

[0144] In Fig. 10, A i can represent the attention weight of the i-th content item. S i can represent the attention score of the i-th content item. V ican represent the content or value of the i-th content item. n i can represent the number of repeated consumptions (weights based on repeated consumption) for the i-th content item.

[0145] From the result of FIG. 10, the activation function (Z) of the recommendation model calculated by considering each consumed content item as a token without considering the weight corresponding to the number of times the content item is repeatedly consumed, as illustrated in FIG. 8 and 9. m ) is an activation function (h) of a recommendation model calculated by considering only unique content items as tokens, while taking into account weights corresponding to the number of repeated consumptions of content items as described above with reference to FIG. 6. i It can be confirmed that it becomes the same as ).

[0146] Accordingly, a recommendation model including an attention mechanism that considers the weights of the embodiment described above with reference to FIG. 6 can obtain recommendation results with substantially the same performance as a recommendation model that provides recommendations by inputting content items of the entire consumption history as tokens, even with fewer operations.

[0147] Thus, the recommendation model of the embodiment can generate recommendations that reflect a user's repeated consumption or long-term preferences for content items with a small amount of computation in an environment with limited computational resources. In particular, the recommendation model of the embodiment can generate recommendations that reflect repeated consumption of content items when the content item is a serialized type that is frequently consumed repeatedly.

[0148] The description of the technical features described above with reference to FIGS. 1 to 7 can be applied as is to FIGS. 8 to 10, so redundant descriptions are omitted.

[0150] Figure 11 shows a schematic diagram of a personalized recommendation system according to one example.

[0151] The illustrated personalized recommendation system (1100) may be the aforementioned computer system (100) or a system including a computer system (100) that includes a configuration corresponding to a recommendation model (1130).

[0152] The personalized recommendation system (1100) may include a data collection and processing device (1110), a recommendation model learning device (1120), and a recommendation model (1130).

[0153] The recommendation model (1130) may be a recommendation model that includes an attention mechanism that considers weights based on the number of times each content item is repeatedly consumed, as described above with reference to FIGS. 1 to 10.

[0154] The recommendation model (1130) may be pre-trained to calculate (determine) a recommendation weight representing the weight of how much information of each content item included in the consumption history (1150) should be retrieved, in order to determine the content item that the user will consume next (i.e., the recommended content item (1160) which is the content item corresponding to '?') based on the user's sequential consumption history (1150) of content items. This recommendation weight may be the aforementioned attention weight (distribution) and may be an intermediate result for determining the recommended content item (1160). The recommendation model (1130) may determine an appropriate recommended content item (1160) from the content items provided by the content service by considering the recommendation weight derived as such an intermediate result.

[0155] The recommendation model training device (1120) may be a device for training such a recommendation model (1130), and the data collection and processing device (1110) may be a device for collecting and processing training data required to train the recommendation model (1130). The data collection and processing device (1110) may collect consumption history of content items of multiple users from a content service and may be configured to process the collected consumption history into a form suitable for training the recommendation model (1130). The recommendation model training device (1120) may train the recommendation model (1130) using the training data processed into a form suitable for training the recommendation model (1130).

[0156] The trained recommendation model (1130) can be installed in a computer system (100) and used to determine recommended content items (1160) based on consumption history (1150). The recommendation model (1130) can be updated by using the user's consumption history regarding recommended content items (1160) or the user's subsequent consumption history regarding content items as additional training data.

[0157] The recommendation model (1130) of the embodiment is a deep learning-based recommendation model that can determine recommended content items by appropriately considering the user's repeated consumption of content items and can not generate recommendations that incorrectly determine the number of repeated consumptions (generalized recommendations, recommendations not proportional to the number of repeated consumptions).

[0158] The description of the technical features described above with reference to FIGS. 1 to 10 can be applied as is to FIG. 11, so redundant descriptions are omitted.

[0160] The device described above may be implemented as a hardware component, a software component, and / or a combination of a hardware component and a software component. For example, the device and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing unit may be described as being used as a single unit, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. In addition, other processing configurations, such as parallel processors, are also possible.

[0161] Software may include computer programs, code, instructions, or a combination of one or more of these, and may configure a processing unit to operate as desired or instruct the processing unit independently or collectively. Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by the processing unit or to provide instructions or data to the processing unit. Software may be distributed over networked computer systems and may be stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0162] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. In this case, the medium may continuously store a program executable by a computer, or temporarily store it for execution or download. Additionally, the medium may be various recording or storage means in the form of a single or multiple hardware combined, and may not be limited to a medium directly connected to a computer system but may exist distributed over a network. Examples of media may include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; magneto-optical media such as floptical disks; and media configured to store program instructions, including ROM, RAM, and flash memory. Additionally, other examples of media may include recording or storage media managed by app stores that distribute applications or sites and servers that supply or distribute various other software.

[0163] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results can be achieved even if the described techniques are performed in a different order than described, and / or the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents.

[0164] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims set forth below.

Claims

Claim 1 A method for determining recommended content items among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history regarding the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, the weight based on the number of repeated consumptions of each content item. A method for determining a recommended content item among the content items based on a recommendation weight calculated for each of the content items, wherein the recommendation model includes an attention mechanism, and the recommendation weight calculated for each of the content items is calculated based on a value obtained by applying the weight to an attention score for each of the content items calculated through the attention mechanism. Claim 2 A method for determining recommended content items according to claim 1, wherein each content item includes a plurality of rounds or a plurality of episodes, and in the consumption history, consumption of at least two rounds or episodes of each content item is treated as a repeated consumption history for each content item. Claim 3 delete Claim 4 A method for determining recommended content items according to claim 1, wherein the step of calculating the recommendation weight comprises: inputting each content item into the attention mechanism to calculate the attention score for each content item; inputting the weight based on the number of repeated consumptions of each content item into the attention mechanism to apply the weight to the calculated attention score; and calculating the recommendation weight based on the attention score to which the weight is applied. Claim 5 In paragraph 4, the weighted attention score is a method for determining recommended content items, wherein the weighted softmax function is applied to the attention score. Claim 6 A method for determining recommended content items, wherein the attention score to which the weights are applied is the attention proportion for each content item. Claim 7 A method for determining recommended content items in claim 5, wherein the sum of recommendation weights calculated for each of the first number of content items is 1. Claim 8 A method for determining recommended content items according to claim 1, wherein the attention mechanism is configured to calculate the recommendation weight using an activation function for calculating the attention weight for each content item reflecting the weight based on the number of times each content item is repeatedly consumed. Claim 9 In paragraph 8, the above activation function is expressed by Equation 1, [Equation 1] The above h i is the distribution of attention weights of each content item for determining the recommended content items, j is a value from 1 to the first number, and S j is the attention score for the j-th content item included in the above consumption history, and the above w j is the weight for the j-th content item, and the above v j A method for determining recommended content items, which are the contents of the j-th content item. Claim 10 In Clause 9, the above S j is expressed by mathematical formula 2, [mathematical formula 2] q i represents the query vector corresponding to the above recommended content item, and k j A method for determining recommended content items, wherein represents a key vector corresponding to the j-th content item included in the above consumption history, and c represents a value obtained by taking the square root of the dimension of the key vector. Claim 11 A method for determining recommended content items among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history for the content items, wherein the consumption history includes a consumption history for a first number of content items among the content items, and includes at least two repeated consumption histories for at least one content item among the first number of content items; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, wherein a weight based on the number of repeated consumptions of each content item is applied; and a step of determining the recommended content item among the content items based on the recommendation weight calculated for each content item, wherein the weight based on the number of repeated consumptions of each content item has a value proportional to the number of repeated consumptions of each content item. Claim 12 A method for determining recommended content items among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history regarding the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, the weight based on the number of repeated consumptions of each content item. A method for determining recommended content items, comprising the step of determining the recommended content items among the content items based on the recommendation weight calculated for each of the above content items, and further comprising the step of determining the first number, wherein the first number is the number of input tokens that are input to the recommendation model, and wherein the determining step determines the first number with a larger value than when the user is requested to determine the recommended content items by reflecting the user's consumption history of a second period which is relatively shorter than the first period. Claim 13 A method for determining content items according to claim 1, wherein the recommendation model is a deep learning model comprising a transformer in which a plurality of the attention mechanisms are stacked. Claim 14 A method for determining a recommended content item among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history for the content items, wherein the consumption history includes a consumption history for a first number of content items among the content items, and includes at least two repeated consumption histories for at least one content item among the first number of content items; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, the weight based on the number of repeated consumptions of each content item; and a step of determining the recommended content item among the content items based on the recommendation weight calculated for each content item, wherein the recommendation weight represents the weight of at least one of the genre, content, and attribute of each content item for determining the recommended content item. Claim 15 A method for determining recommended content items among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history regarding the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, the weight based on the number of repeated consumptions of each content item. A method for determining a content item, comprising the step of determining the recommended content item among the content items based on the recommendation weight calculated for each of the content items, wherein the recommendation weight represents the weight of how much information of each of the content items included in the consumption history must be retrieved to determine the recommended content item, and the recommendation model is pre-trained to determine the recommendation weight by using the recommended content item as a query and each of the content items as a key. Claim 16 A method for determining recommended content items among content items provided through a content service, performed by a computer system, comprising: a step of obtaining a user’s sequential consumption history regarding the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—; and a step of determining at least one recommended content item among the content items using the consumption history obtained as input to a pre-established recommendation model, wherein the step of determining the recommended content item comprises: a step of calculating a recommendation weight applied to each content item of the first number of content items, the weight based on the number of repeated consumptions of each content item. A method for determining recommended content items, comprising the step of determining the recommended content item among the content items based on the recommendation weight calculated for each of the above content items, wherein the recommendation model is configured to determine the same recommended content item even if the consumption order of the content items included in the consumption history is different, under the condition that the types of content items included in the consumption history and the number of repeated consumptions of each content item are given equally. Claim 17 A program recorded on a computer-readable recording medium for executing the method of claim 1 on the computer system. Claim 18 A computer-readable recording medium for executing the method of claim 1 on the computer system. Claim 19 A computer system for determining recommended content items among content items provided through a content service comprises at least one processor implemented to execute a readable command in the computer system, wherein the at least one processor obtains a user’s sequential consumption history of the content items—the consumption history includes a consumption history for a first number of content items among the content items, wherein the consumption history includes at least two repeated consumption histories for at least one content item among the first number of content items—and determines at least one recommended content item among the content items using the consumption history obtained as input to a pre-built recommendation model, wherein for each content item of the first number of content items, a recommendation weight is calculated with a weight applied based on the number of repeated consumptions of each content item, and the recommended content item among the content items is determined based on the recommendation weight calculated for each content item, wherein the recommendation model includes an attention mechanism, and the recommendation weight calculated for each content item is based on a value obtained by applying the weight to an attention score for each content item calculated through the attention mechanism A computer system being calculated.

Citation Information

Patent Citations

  • Method and system for recommending products based on collaborative filtering based on natural language programming analogy

    US20230230141A1

  • Content recommending method, and apparatus and system for providing same

    KR1020150020390A